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> |
66114cad | 29 | #include <linux/backing-dev.h> |
9bffad1e TT |
30 | #include <trace/events/ext4.h> |
31 | ||
a0b30c12 TT |
32 | #ifdef CONFIG_EXT4_DEBUG |
33 | ushort ext4_mballoc_debug __read_mostly; | |
34 | ||
35 | module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); | |
36 | MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); | |
37 | #endif | |
38 | ||
c9de560d AT |
39 | /* |
40 | * MUSTDO: | |
41 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
42 | * - search for metadata in few groups | |
43 | * | |
44 | * TODO v4: | |
45 | * - normalization should take into account whether file is still open | |
46 | * - discard preallocations if no free space left (policy?) | |
47 | * - don't normalize tails | |
48 | * - quota | |
49 | * - reservation for superuser | |
50 | * | |
51 | * TODO v3: | |
52 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
53 | * - track min/max extents in each group for better group selection | |
54 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
55 | * - tree of groups sorted by number of free blocks | |
56 | * - error handling | |
57 | */ | |
58 | ||
59 | /* | |
60 | * The allocation request involve request for multiple number of blocks | |
61 | * near to the goal(block) value specified. | |
62 | * | |
b713a5ec TT |
63 | * During initialization phase of the allocator we decide to use the |
64 | * group preallocation or inode preallocation depending on the size of | |
65 | * the file. The size of the file could be the resulting file size we | |
66 | * would have after allocation, or the current file size, which ever | |
67 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
68 | * select to use the group preallocation. The default value of | |
69 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
70 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
71 | * terms of number of blocks. | |
c9de560d AT |
72 | * |
73 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 74 | * ensure that we have small files closer together on the disk. |
c9de560d | 75 | * |
b713a5ec TT |
76 | * First stage the allocator looks at the inode prealloc list, |
77 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
78 | * spaces for this particular inode. The inode prealloc space is | |
79 | * represented as: | |
c9de560d AT |
80 | * |
81 | * pa_lstart -> the logical start block for this prealloc space | |
82 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
83 | * pa_len -> length for this prealloc space (in clusters) |
84 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
85 | * |
86 | * The inode preallocation space is used looking at the _logical_ start | |
87 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
88 | * space we will consume the particular prealloc space. This makes sure that |
89 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
90 | * |
91 | * The important thing to be noted in case of inode prealloc space is that | |
92 | * we don't modify the values associated to inode prealloc space except | |
93 | * pa_free. | |
94 | * | |
95 | * If we are not able to find blocks in the inode prealloc space and if we | |
96 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 97 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
98 | * |
99 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
100 | * | |
101 | * The reason for having a per cpu locality group is to reduce the contention | |
102 | * between CPUs. It is possible to get scheduled at this point. | |
103 | * | |
104 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 105 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
106 | * |
107 | * If we can't allocate blocks via inode prealloc or/and locality group | |
108 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
109 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
110 | * mapped to the buddy and bitmap information regarding different | |
111 | * groups. The buddy information is attached to buddy cache inode so that | |
112 | * we can access them through the page cache. The information regarding | |
113 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
114 | * block bitmap and buddy information. The information are stored in the | |
115 | * inode as: | |
116 | * | |
117 | * { page } | |
c3a326a6 | 118 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
119 | * |
120 | * | |
121 | * one block each for bitmap and buddy information. So for each group we | |
122 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
123 | * blocksize) blocks. So it can have information regarding groups_per_page | |
124 | * which is blocks_per_page/2 | |
125 | * | |
126 | * The buddy cache inode is not stored on disk. The inode is thrown | |
127 | * away when the filesystem is unmounted. | |
128 | * | |
129 | * We look for count number of blocks in the buddy cache. If we were able | |
130 | * to locate that many free blocks we return with additional information | |
131 | * regarding rest of the contiguous physical block available | |
132 | * | |
133 | * Before allocating blocks via buddy cache we normalize the request | |
134 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
135 | * blocks that we get after allocation is added to the respective prealloc | |
136 | * list. In case of inode preallocation we follow a list of heuristics | |
137 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
138 | * we are doing a group prealloc we try to normalize the request to | |
27baebb8 TT |
139 | * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is |
140 | * dependent on the cluster size; for non-bigalloc file systems, it is | |
c9de560d | 141 | * 512 blocks. This can be tuned via |
d7a1fee1 | 142 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
143 | * terms of number of blocks. If we have mounted the file system with -O |
144 | * stripe=<value> option the group prealloc request is normalized to the | |
d7a1fee1 DE |
145 | * the smallest multiple of the stripe value (sbi->s_stripe) which is |
146 | * greater than the default mb_group_prealloc. | |
c9de560d | 147 | * |
d7a1fee1 | 148 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 149 | * |
b713a5ec TT |
150 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
151 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
152 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 153 | * |
b713a5ec | 154 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
155 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
156 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 157 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 158 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
159 | * stripe size. This should result in better allocation on RAID setups. If |
160 | * not, we search in the specific group using bitmap for best extents. The | |
161 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 162 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 163 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
164 | * best extent in the found extents. Searching for the blocks starts with |
165 | * the group specified as the goal value in allocation context via | |
166 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 167 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
168 | * checked. |
169 | * | |
170 | * Both the prealloc space are getting populated as above. So for the first | |
171 | * request we will hit the buddy cache which will result in this prealloc | |
172 | * space getting filled. The prealloc space is then later used for the | |
173 | * subsequent request. | |
174 | */ | |
175 | ||
176 | /* | |
177 | * mballoc operates on the following data: | |
178 | * - on-disk bitmap | |
179 | * - in-core buddy (actually includes buddy and bitmap) | |
180 | * - preallocation descriptors (PAs) | |
181 | * | |
182 | * there are two types of preallocations: | |
183 | * - inode | |
184 | * assiged to specific inode and can be used for this inode only. | |
185 | * it describes part of inode's space preallocated to specific | |
186 | * physical blocks. any block from that preallocated can be used | |
187 | * independent. the descriptor just tracks number of blocks left | |
188 | * unused. so, before taking some block from descriptor, one must | |
189 | * make sure corresponded logical block isn't allocated yet. this | |
190 | * also means that freeing any block within descriptor's range | |
191 | * must discard all preallocated blocks. | |
192 | * - locality group | |
193 | * assigned to specific locality group which does not translate to | |
194 | * permanent set of inodes: inode can join and leave group. space | |
195 | * from this type of preallocation can be used for any inode. thus | |
196 | * it's consumed from the beginning to the end. | |
197 | * | |
198 | * relation between them can be expressed as: | |
199 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
200 | * | |
201 | * this mean blocks mballoc considers used are: | |
202 | * - allocated blocks (persistent) | |
203 | * - preallocated blocks (non-persistent) | |
204 | * | |
205 | * consistency in mballoc world means that at any time a block is either | |
206 | * free or used in ALL structures. notice: "any time" should not be read | |
207 | * literally -- time is discrete and delimited by locks. | |
208 | * | |
209 | * to keep it simple, we don't use block numbers, instead we count number of | |
210 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
211 | * | |
212 | * all operations can be expressed as: | |
213 | * - init buddy: buddy = on-disk + PAs | |
214 | * - new PA: buddy += N; PA = N | |
215 | * - use inode PA: on-disk += N; PA -= N | |
216 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
217 | * - use locality group PA on-disk += N; PA -= N | |
218 | * - discard locality group PA buddy -= PA; PA = 0 | |
219 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
220 | * is used in real operation because we can't know actual used | |
221 | * bits from PA, only from on-disk bitmap | |
222 | * | |
223 | * if we follow this strict logic, then all operations above should be atomic. | |
224 | * given some of them can block, we'd have to use something like semaphores | |
225 | * killing performance on high-end SMP hardware. let's try to relax it using | |
226 | * the following knowledge: | |
227 | * 1) if buddy is referenced, it's already initialized | |
228 | * 2) while block is used in buddy and the buddy is referenced, | |
229 | * nobody can re-allocate that block | |
230 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
231 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
232 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
233 | * block | |
234 | * | |
235 | * so, now we're building a concurrency table: | |
236 | * - init buddy vs. | |
237 | * - new PA | |
238 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
239 | * until PA is linked to allocation group to avoid concurrent buddy init | |
240 | * - use inode PA | |
241 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
242 | * given (3) we care that PA-=N operation doesn't interfere with init | |
243 | * - discard inode PA | |
244 | * the simplest way would be to have buddy initialized by the discard | |
245 | * - use locality group PA | |
246 | * again PA-=N must be serialized with init | |
247 | * - discard locality group PA | |
248 | * the simplest way would be to have buddy initialized by the discard | |
249 | * - new PA vs. | |
250 | * - use inode PA | |
251 | * i_data_sem serializes them | |
252 | * - discard inode PA | |
253 | * discard process must wait until PA isn't used by another process | |
254 | * - use locality group PA | |
255 | * some mutex should serialize them | |
256 | * - discard locality group PA | |
257 | * discard process must wait until PA isn't used by another process | |
258 | * - use inode PA | |
259 | * - use inode PA | |
260 | * i_data_sem or another mutex should serializes them | |
261 | * - discard inode PA | |
262 | * discard process must wait until PA isn't used by another process | |
263 | * - use locality group PA | |
264 | * nothing wrong here -- they're different PAs covering different blocks | |
265 | * - discard locality group PA | |
266 | * discard process must wait until PA isn't used by another process | |
267 | * | |
268 | * now we're ready to make few consequences: | |
269 | * - PA is referenced and while it is no discard is possible | |
270 | * - PA is referenced until block isn't marked in on-disk bitmap | |
271 | * - PA changes only after on-disk bitmap | |
272 | * - discard must not compete with init. either init is done before | |
273 | * any discard or they're serialized somehow | |
274 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
275 | * | |
276 | * a special case when we've used PA to emptiness. no need to modify buddy | |
277 | * in this case, but we should care about concurrent init | |
278 | * | |
279 | */ | |
280 | ||
281 | /* | |
282 | * Logic in few words: | |
283 | * | |
284 | * - allocation: | |
285 | * load group | |
286 | * find blocks | |
287 | * mark bits in on-disk bitmap | |
288 | * release group | |
289 | * | |
290 | * - use preallocation: | |
291 | * find proper PA (per-inode or group) | |
292 | * load group | |
293 | * mark bits in on-disk bitmap | |
294 | * release group | |
295 | * release PA | |
296 | * | |
297 | * - free: | |
298 | * load group | |
299 | * mark bits in on-disk bitmap | |
300 | * release group | |
301 | * | |
302 | * - discard preallocations in group: | |
303 | * mark PAs deleted | |
304 | * move them onto local list | |
305 | * load on-disk bitmap | |
306 | * load group | |
307 | * remove PA from object (inode or locality group) | |
308 | * mark free blocks in-core | |
309 | * | |
310 | * - discard inode's preallocations: | |
311 | */ | |
312 | ||
313 | /* | |
314 | * Locking rules | |
315 | * | |
316 | * Locks: | |
317 | * - bitlock on a group (group) | |
318 | * - object (inode/locality) (object) | |
319 | * - per-pa lock (pa) | |
320 | * | |
321 | * Paths: | |
322 | * - new pa | |
323 | * object | |
324 | * group | |
325 | * | |
326 | * - find and use pa: | |
327 | * pa | |
328 | * | |
329 | * - release consumed pa: | |
330 | * pa | |
331 | * group | |
332 | * object | |
333 | * | |
334 | * - generate in-core bitmap: | |
335 | * group | |
336 | * pa | |
337 | * | |
338 | * - discard all for given object (inode, locality group): | |
339 | * object | |
340 | * pa | |
341 | * group | |
342 | * | |
343 | * - discard all for given group: | |
344 | * group | |
345 | * pa | |
346 | * group | |
347 | * object | |
348 | * | |
349 | */ | |
c3a326a6 AK |
350 | static struct kmem_cache *ext4_pspace_cachep; |
351 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 352 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
353 | |
354 | /* We create slab caches for groupinfo data structures based on the | |
355 | * superblock block size. There will be one per mounted filesystem for | |
356 | * each unique s_blocksize_bits */ | |
2892c15d | 357 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
358 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
359 | ||
2892c15d ES |
360 | static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
361 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", | |
362 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
363 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
364 | }; | |
365 | ||
c3a326a6 AK |
366 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
367 | ext4_group_t group); | |
7a2fcbf7 AK |
368 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
369 | ext4_group_t group); | |
c3a326a6 | 370 | |
ffad0a44 AK |
371 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
372 | { | |
c9de560d | 373 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
374 | *bit += ((unsigned long) addr & 7UL) << 3; |
375 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 376 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
377 | *bit += ((unsigned long) addr & 3UL) << 3; |
378 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
379 | #else |
380 | #error "how many bits you are?!" | |
381 | #endif | |
ffad0a44 AK |
382 | return addr; |
383 | } | |
c9de560d AT |
384 | |
385 | static inline int mb_test_bit(int bit, void *addr) | |
386 | { | |
387 | /* | |
388 | * ext4_test_bit on architecture like powerpc | |
389 | * needs unsigned long aligned address | |
390 | */ | |
ffad0a44 | 391 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
392 | return ext4_test_bit(bit, addr); |
393 | } | |
394 | ||
395 | static inline void mb_set_bit(int bit, void *addr) | |
396 | { | |
ffad0a44 | 397 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
398 | ext4_set_bit(bit, addr); |
399 | } | |
400 | ||
c9de560d AT |
401 | static inline void mb_clear_bit(int bit, void *addr) |
402 | { | |
ffad0a44 | 403 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
404 | ext4_clear_bit(bit, addr); |
405 | } | |
406 | ||
eabe0444 AS |
407 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
408 | { | |
409 | addr = mb_correct_addr_and_bit(&bit, addr); | |
410 | return ext4_test_and_clear_bit(bit, addr); | |
411 | } | |
412 | ||
ffad0a44 AK |
413 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
414 | { | |
e7dfb246 | 415 | int fix = 0, ret, tmpmax; |
ffad0a44 | 416 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 417 | tmpmax = max + fix; |
ffad0a44 AK |
418 | start += fix; |
419 | ||
e7dfb246 AK |
420 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
421 | if (ret > max) | |
422 | return max; | |
423 | return ret; | |
ffad0a44 AK |
424 | } |
425 | ||
426 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
427 | { | |
e7dfb246 | 428 | int fix = 0, ret, tmpmax; |
ffad0a44 | 429 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 430 | tmpmax = max + fix; |
ffad0a44 AK |
431 | start += fix; |
432 | ||
e7dfb246 AK |
433 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
434 | if (ret > max) | |
435 | return max; | |
436 | return ret; | |
ffad0a44 AK |
437 | } |
438 | ||
c9de560d AT |
439 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
440 | { | |
441 | char *bb; | |
442 | ||
c5e8f3f3 | 443 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
444 | BUG_ON(max == NULL); |
445 | ||
446 | if (order > e4b->bd_blkbits + 1) { | |
447 | *max = 0; | |
448 | return NULL; | |
449 | } | |
450 | ||
451 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
452 | if (order == 0) { |
453 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 454 | return e4b->bd_bitmap; |
84b775a3 | 455 | } |
c9de560d | 456 | |
c5e8f3f3 | 457 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
458 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
459 | ||
460 | return bb; | |
461 | } | |
462 | ||
463 | #ifdef DOUBLE_CHECK | |
464 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
465 | int first, int count) | |
466 | { | |
467 | int i; | |
468 | struct super_block *sb = e4b->bd_sb; | |
469 | ||
470 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
471 | return; | |
bc8e6740 | 472 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
473 | for (i = 0; i < count; i++) { |
474 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
475 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
476 | |
477 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 478 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 479 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
480 | inode ? inode->i_ino : 0, |
481 | blocknr, | |
482 | "freeing block already freed " | |
483 | "(bit %u)", | |
484 | first + i); | |
c9de560d AT |
485 | } |
486 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
487 | } | |
488 | } | |
489 | ||
490 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
491 | { | |
492 | int i; | |
493 | ||
494 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
495 | return; | |
bc8e6740 | 496 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
497 | for (i = 0; i < count; i++) { |
498 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
499 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
500 | } | |
501 | } | |
502 | ||
503 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
504 | { | |
505 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
506 | unsigned char *b1, *b2; | |
507 | int i; | |
508 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
509 | b2 = (unsigned char *) bitmap; | |
510 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
511 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
512 | ext4_msg(e4b->bd_sb, KERN_ERR, |
513 | "corruption in group %u " | |
514 | "at byte %u(%u): %x in copy != %x " | |
515 | "on disk/prealloc", | |
516 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
517 | BUG(); |
518 | } | |
519 | } | |
520 | } | |
521 | } | |
522 | ||
523 | #else | |
524 | static inline void mb_free_blocks_double(struct inode *inode, | |
525 | struct ext4_buddy *e4b, int first, int count) | |
526 | { | |
527 | return; | |
528 | } | |
529 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
530 | int first, int count) | |
531 | { | |
532 | return; | |
533 | } | |
534 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
535 | { | |
536 | return; | |
537 | } | |
538 | #endif | |
539 | ||
540 | #ifdef AGGRESSIVE_CHECK | |
541 | ||
542 | #define MB_CHECK_ASSERT(assert) \ | |
543 | do { \ | |
544 | if (!(assert)) { \ | |
545 | printk(KERN_EMERG \ | |
546 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
547 | function, file, line, # assert); \ | |
548 | BUG(); \ | |
549 | } \ | |
550 | } while (0) | |
551 | ||
552 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
553 | const char *function, int line) | |
554 | { | |
555 | struct super_block *sb = e4b->bd_sb; | |
556 | int order = e4b->bd_blkbits + 1; | |
557 | int max; | |
558 | int max2; | |
559 | int i; | |
560 | int j; | |
561 | int k; | |
562 | int count; | |
563 | struct ext4_group_info *grp; | |
564 | int fragments = 0; | |
565 | int fstart; | |
566 | struct list_head *cur; | |
567 | void *buddy; | |
568 | void *buddy2; | |
569 | ||
c9de560d AT |
570 | { |
571 | static int mb_check_counter; | |
572 | if (mb_check_counter++ % 100 != 0) | |
573 | return 0; | |
574 | } | |
575 | ||
576 | while (order > 1) { | |
577 | buddy = mb_find_buddy(e4b, order, &max); | |
578 | MB_CHECK_ASSERT(buddy); | |
579 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
580 | MB_CHECK_ASSERT(buddy2); | |
581 | MB_CHECK_ASSERT(buddy != buddy2); | |
582 | MB_CHECK_ASSERT(max * 2 == max2); | |
583 | ||
584 | count = 0; | |
585 | for (i = 0; i < max; i++) { | |
586 | ||
587 | if (mb_test_bit(i, buddy)) { | |
588 | /* only single bit in buddy2 may be 1 */ | |
589 | if (!mb_test_bit(i << 1, buddy2)) { | |
590 | MB_CHECK_ASSERT( | |
591 | mb_test_bit((i<<1)+1, buddy2)); | |
592 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
593 | MB_CHECK_ASSERT( | |
594 | mb_test_bit(i << 1, buddy2)); | |
595 | } | |
596 | continue; | |
597 | } | |
598 | ||
0a10da73 | 599 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
600 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
601 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
602 | ||
603 | for (j = 0; j < (1 << order); j++) { | |
604 | k = (i * (1 << order)) + j; | |
605 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 606 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
607 | } |
608 | count++; | |
609 | } | |
610 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
611 | order--; | |
612 | } | |
613 | ||
614 | fstart = -1; | |
615 | buddy = mb_find_buddy(e4b, 0, &max); | |
616 | for (i = 0; i < max; i++) { | |
617 | if (!mb_test_bit(i, buddy)) { | |
618 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
619 | if (fstart == -1) { | |
620 | fragments++; | |
621 | fstart = i; | |
622 | } | |
623 | continue; | |
624 | } | |
625 | fstart = -1; | |
626 | /* check used bits only */ | |
627 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
628 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
629 | k = i >> j; | |
630 | MB_CHECK_ASSERT(k < max2); | |
631 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
632 | } | |
633 | } | |
634 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
635 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
636 | ||
637 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
638 | list_for_each(cur, &grp->bb_prealloc_list) { |
639 | ext4_group_t groupnr; | |
640 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
641 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
642 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 643 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 644 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
645 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
646 | } | |
647 | return 0; | |
648 | } | |
649 | #undef MB_CHECK_ASSERT | |
650 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 651 | __FILE__, __func__, __LINE__) |
c9de560d AT |
652 | #else |
653 | #define mb_check_buddy(e4b) | |
654 | #endif | |
655 | ||
7c786059 CL |
656 | /* |
657 | * Divide blocks started from @first with length @len into | |
658 | * smaller chunks with power of 2 blocks. | |
659 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
660 | * then increase bb_counters[] for corresponded chunk size. | |
661 | */ | |
c9de560d | 662 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 663 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
664 | struct ext4_group_info *grp) |
665 | { | |
666 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
667 | ext4_grpblk_t min; |
668 | ext4_grpblk_t max; | |
669 | ext4_grpblk_t chunk; | |
26492d8a | 670 | unsigned int border; |
c9de560d | 671 | |
7137d7a4 | 672 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
673 | |
674 | border = 2 << sb->s_blocksize_bits; | |
675 | ||
676 | while (len > 0) { | |
677 | /* find how many blocks can be covered since this position */ | |
678 | max = ffs(first | border) - 1; | |
679 | ||
680 | /* find how many blocks of power 2 we need to mark */ | |
681 | min = fls(len) - 1; | |
682 | ||
683 | if (max < min) | |
684 | min = max; | |
685 | chunk = 1 << min; | |
686 | ||
687 | /* mark multiblock chunks only */ | |
688 | grp->bb_counters[min]++; | |
689 | if (min > 0) | |
690 | mb_clear_bit(first >> min, | |
691 | buddy + sbi->s_mb_offsets[min]); | |
692 | ||
693 | len -= chunk; | |
694 | first += chunk; | |
695 | } | |
696 | } | |
697 | ||
8a57d9d6 CW |
698 | /* |
699 | * Cache the order of the largest free extent we have available in this block | |
700 | * group. | |
701 | */ | |
702 | static void | |
703 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
704 | { | |
705 | int i; | |
706 | int bits; | |
707 | ||
708 | grp->bb_largest_free_order = -1; /* uninit */ | |
709 | ||
710 | bits = sb->s_blocksize_bits + 1; | |
711 | for (i = bits; i >= 0; i--) { | |
712 | if (grp->bb_counters[i] > 0) { | |
713 | grp->bb_largest_free_order = i; | |
714 | break; | |
715 | } | |
716 | } | |
717 | } | |
718 | ||
089ceecc ES |
719 | static noinline_for_stack |
720 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
721 | void *buddy, void *bitmap, ext4_group_t group) |
722 | { | |
723 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 724 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 725 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
726 | ext4_grpblk_t i = 0; |
727 | ext4_grpblk_t first; | |
728 | ext4_grpblk_t len; | |
c9de560d AT |
729 | unsigned free = 0; |
730 | unsigned fragments = 0; | |
731 | unsigned long long period = get_cycles(); | |
732 | ||
733 | /* initialize buddy from bitmap which is aggregation | |
734 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 735 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
736 | grp->bb_first_free = i; |
737 | while (i < max) { | |
738 | fragments++; | |
739 | first = i; | |
ffad0a44 | 740 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
741 | len = i - first; |
742 | free += len; | |
743 | if (len > 1) | |
744 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
745 | else | |
746 | grp->bb_counters[0]++; | |
747 | if (i < max) | |
ffad0a44 | 748 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
749 | } |
750 | grp->bb_fragments = fragments; | |
751 | ||
752 | if (free != grp->bb_free) { | |
1cac41cb MB |
753 | /* for more specific debugging, sangwoo2.lee */ |
754 | struct ext4_group_desc *desc; | |
755 | ext4_fsblk_t bitmap_blk; | |
756 | ||
757 | desc = ext4_get_group_desc(sb, group, NULL); | |
758 | bitmap_blk = ext4_block_bitmap(sb, desc); | |
759 | ||
760 | print_block_data(sb, bitmap_blk, bitmap, 0, EXT4_BLOCK_SIZE(sb)); | |
761 | /* for more specific debugging */ | |
762 | ||
e29136f8 | 763 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
764 | "block bitmap and bg descriptor " |
765 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 766 | free, grp->bb_free); |
e56eb659 | 767 | /* |
163a203d | 768 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
769 | * corrupt and update bb_free using bitmap value |
770 | */ | |
c9de560d | 771 | grp->bb_free = free; |
e43bb4e6 NJ |
772 | if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp)) |
773 | percpu_counter_sub(&sbi->s_freeclusters_counter, | |
774 | grp->bb_free); | |
163a203d | 775 | set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state); |
c9de560d | 776 | } |
8a57d9d6 | 777 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
778 | |
779 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
780 | ||
781 | period = get_cycles() - period; | |
782 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
783 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
784 | EXT4_SB(sb)->s_mb_generation_time += period; | |
785 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
786 | } | |
787 | ||
eabe0444 AS |
788 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
789 | { | |
790 | int count; | |
791 | int order = 1; | |
792 | void *buddy; | |
793 | ||
794 | while ((buddy = mb_find_buddy(e4b, order++, &count))) { | |
795 | ext4_set_bits(buddy, 0, count); | |
796 | } | |
797 | e4b->bd_info->bb_fragments = 0; | |
798 | memset(e4b->bd_info->bb_counters, 0, | |
799 | sizeof(*e4b->bd_info->bb_counters) * | |
800 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
801 | ||
802 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
803 | e4b->bd_bitmap, e4b->bd_group); | |
804 | } | |
805 | ||
c9de560d AT |
806 | /* The buddy information is attached the buddy cache inode |
807 | * for convenience. The information regarding each group | |
808 | * is loaded via ext4_mb_load_buddy. The information involve | |
809 | * block bitmap and buddy information. The information are | |
810 | * stored in the inode as | |
811 | * | |
812 | * { page } | |
c3a326a6 | 813 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
814 | * |
815 | * | |
816 | * one block each for bitmap and buddy information. | |
817 | * So for each group we take up 2 blocks. A page can | |
818 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
819 | * So it can have information regarding groups_per_page which | |
820 | * is blocks_per_page/2 | |
8a57d9d6 CW |
821 | * |
822 | * Locking note: This routine takes the block group lock of all groups | |
823 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
824 | */ |
825 | ||
d8aafd0c | 826 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 827 | { |
8df9675f | 828 | ext4_group_t ngroups; |
c9de560d AT |
829 | int blocksize; |
830 | int blocks_per_page; | |
831 | int groups_per_page; | |
832 | int err = 0; | |
833 | int i; | |
813e5727 | 834 | ext4_group_t first_group, group; |
c9de560d AT |
835 | int first_block; |
836 | struct super_block *sb; | |
837 | struct buffer_head *bhs; | |
fa77dcfa | 838 | struct buffer_head **bh = NULL; |
c9de560d AT |
839 | struct inode *inode; |
840 | char *data; | |
841 | char *bitmap; | |
9b8b7d35 | 842 | struct ext4_group_info *grinfo; |
c9de560d | 843 | |
6ba495e9 | 844 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
845 | |
846 | inode = page->mapping->host; | |
847 | sb = inode->i_sb; | |
8df9675f | 848 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
849 | blocksize = 1 << inode->i_blkbits; |
850 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
851 | ||
852 | groups_per_page = blocks_per_page >> 1; | |
853 | if (groups_per_page == 0) | |
854 | groups_per_page = 1; | |
855 | ||
856 | /* allocate buffer_heads to read bitmaps */ | |
857 | if (groups_per_page > 1) { | |
c9de560d | 858 | i = sizeof(struct buffer_head *) * groups_per_page; |
d8aafd0c | 859 | bh = kzalloc(i, gfp); |
813e5727 TT |
860 | if (bh == NULL) { |
861 | err = -ENOMEM; | |
c9de560d | 862 | goto out; |
813e5727 | 863 | } |
c9de560d AT |
864 | } else |
865 | bh = &bhs; | |
866 | ||
867 | first_group = page->index * blocks_per_page / 2; | |
868 | ||
869 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
870 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
871 | if (group >= ngroups) | |
c9de560d AT |
872 | break; |
873 | ||
813e5727 | 874 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
875 | /* |
876 | * If page is uptodate then we came here after online resize | |
877 | * which added some new uninitialized group info structs, so | |
878 | * we must skip all initialized uptodate buddies on the page, | |
879 | * which may be currently in use by an allocating task. | |
880 | */ | |
881 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
882 | bh[i] = NULL; | |
883 | continue; | |
884 | } | |
9008a58e DW |
885 | bh[i] = ext4_read_block_bitmap_nowait(sb, group); |
886 | if (IS_ERR(bh[i])) { | |
887 | err = PTR_ERR(bh[i]); | |
888 | bh[i] = NULL; | |
c9de560d | 889 | goto out; |
2ccb5fb9 | 890 | } |
813e5727 | 891 | mb_debug(1, "read bitmap for group %u\n", group); |
c9de560d AT |
892 | } |
893 | ||
894 | /* wait for I/O completion */ | |
813e5727 | 895 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
896 | int err2; |
897 | ||
898 | if (!bh[i]) | |
899 | continue; | |
900 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
901 | if (!err) | |
902 | err = err2; | |
813e5727 | 903 | } |
c9de560d AT |
904 | |
905 | first_block = page->index * blocks_per_page; | |
906 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 907 | group = (first_block + i) >> 1; |
8df9675f | 908 | if (group >= ngroups) |
c9de560d AT |
909 | break; |
910 | ||
9b8b7d35 AG |
911 | if (!bh[group - first_group]) |
912 | /* skip initialized uptodate buddy */ | |
913 | continue; | |
914 | ||
bbdc322f LC |
915 | if (!buffer_verified(bh[group - first_group])) |
916 | /* Skip faulty bitmaps */ | |
917 | continue; | |
918 | err = 0; | |
919 | ||
c9de560d AT |
920 | /* |
921 | * data carry information regarding this | |
922 | * particular group in the format specified | |
923 | * above | |
924 | * | |
925 | */ | |
926 | data = page_address(page) + (i * blocksize); | |
927 | bitmap = bh[group - first_group]->b_data; | |
928 | ||
929 | /* | |
930 | * We place the buddy block and bitmap block | |
931 | * close together | |
932 | */ | |
933 | if ((first_block + i) & 1) { | |
934 | /* this is block of buddy */ | |
935 | BUG_ON(incore == NULL); | |
6ba495e9 | 936 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 937 | group, page->index, i * blocksize); |
f307333e | 938 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
939 | grinfo = ext4_get_group_info(sb, group); |
940 | grinfo->bb_fragments = 0; | |
941 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
942 | sizeof(*grinfo->bb_counters) * |
943 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
944 | /* |
945 | * incore got set to the group block bitmap below | |
946 | */ | |
7a2fcbf7 | 947 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
948 | /* init the buddy */ |
949 | memset(data, 0xff, blocksize); | |
c9de560d | 950 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 951 | ext4_unlock_group(sb, group); |
c9de560d AT |
952 | incore = NULL; |
953 | } else { | |
954 | /* this is block of bitmap */ | |
955 | BUG_ON(incore != NULL); | |
6ba495e9 | 956 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 957 | group, page->index, i * blocksize); |
f307333e | 958 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
959 | |
960 | /* see comments in ext4_mb_put_pa() */ | |
961 | ext4_lock_group(sb, group); | |
962 | memcpy(data, bitmap, blocksize); | |
963 | ||
964 | /* mark all preallocated blks used in in-core bitmap */ | |
965 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 966 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
967 | ext4_unlock_group(sb, group); |
968 | ||
969 | /* set incore so that the buddy information can be | |
970 | * generated using this | |
971 | */ | |
972 | incore = data; | |
973 | } | |
974 | } | |
975 | SetPageUptodate(page); | |
976 | ||
977 | out: | |
978 | if (bh) { | |
9b8b7d35 | 979 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
980 | brelse(bh[i]); |
981 | if (bh != &bhs) | |
982 | kfree(bh); | |
983 | } | |
984 | return err; | |
985 | } | |
986 | ||
eee4adc7 | 987 | /* |
2de8807b AG |
988 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
989 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
990 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
991 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 992 | */ |
2de8807b | 993 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
d8aafd0c | 994 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 995 | { |
2de8807b AG |
996 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
997 | int block, pnum, poff; | |
eee4adc7 | 998 | int blocks_per_page; |
2de8807b AG |
999 | struct page *page; |
1000 | ||
1001 | e4b->bd_buddy_page = NULL; | |
1002 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 ES |
1003 | |
1004 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
1005 | /* | |
1006 | * the buddy cache inode stores the block bitmap | |
1007 | * and buddy information in consecutive blocks. | |
1008 | * So for each group we need two blocks. | |
1009 | */ | |
1010 | block = group * 2; | |
1011 | pnum = block / blocks_per_page; | |
2de8807b | 1012 | poff = block % blocks_per_page; |
d8aafd0c | 1013 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1014 | if (!page) |
c57ab39b | 1015 | return -ENOMEM; |
2de8807b AG |
1016 | BUG_ON(page->mapping != inode->i_mapping); |
1017 | e4b->bd_bitmap_page = page; | |
1018 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1019 | ||
1020 | if (blocks_per_page >= 2) { | |
1021 | /* buddy and bitmap are on the same page */ | |
1022 | return 0; | |
eee4adc7 | 1023 | } |
2de8807b AG |
1024 | |
1025 | block++; | |
1026 | pnum = block / blocks_per_page; | |
d8aafd0c | 1027 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1028 | if (!page) |
c57ab39b | 1029 | return -ENOMEM; |
2de8807b AG |
1030 | BUG_ON(page->mapping != inode->i_mapping); |
1031 | e4b->bd_buddy_page = page; | |
1032 | return 0; | |
eee4adc7 ES |
1033 | } |
1034 | ||
2de8807b | 1035 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1036 | { |
2de8807b AG |
1037 | if (e4b->bd_bitmap_page) { |
1038 | unlock_page(e4b->bd_bitmap_page); | |
1039 | page_cache_release(e4b->bd_bitmap_page); | |
1040 | } | |
1041 | if (e4b->bd_buddy_page) { | |
1042 | unlock_page(e4b->bd_buddy_page); | |
1043 | page_cache_release(e4b->bd_buddy_page); | |
eee4adc7 | 1044 | } |
eee4adc7 ES |
1045 | } |
1046 | ||
8a57d9d6 CW |
1047 | /* |
1048 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1049 | * block group lock of all groups for this page; do not hold the BG lock when | |
1050 | * calling this routine! | |
1051 | */ | |
b6a758ec | 1052 | static noinline_for_stack |
d8aafd0c | 1053 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1054 | { |
1055 | ||
b6a758ec | 1056 | struct ext4_group_info *this_grp; |
2de8807b AG |
1057 | struct ext4_buddy e4b; |
1058 | struct page *page; | |
1059 | int ret = 0; | |
b6a758ec | 1060 | |
b10a44c3 | 1061 | might_sleep(); |
b6a758ec | 1062 | mb_debug(1, "init group %u\n", group); |
b6a758ec AK |
1063 | this_grp = ext4_get_group_info(sb, group); |
1064 | /* | |
08c3a813 AK |
1065 | * This ensures that we don't reinit the buddy cache |
1066 | * page which map to the group from which we are already | |
1067 | * allocating. If we are looking at the buddy cache we would | |
1068 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1069 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1070 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1071 | * page accessed. | |
b6a758ec | 1072 | */ |
d8aafd0c | 1073 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1074 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1075 | /* |
1076 | * somebody initialized the group | |
1077 | * return without doing anything | |
1078 | */ | |
b6a758ec AK |
1079 | goto err; |
1080 | } | |
2de8807b AG |
1081 | |
1082 | page = e4b.bd_bitmap_page; | |
d8aafd0c | 1083 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1084 | if (ret) |
1085 | goto err; | |
1086 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1087 | ret = -EIO; |
1088 | goto err; | |
1089 | } | |
b6a758ec | 1090 | |
2de8807b | 1091 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1092 | /* |
1093 | * If both the bitmap and buddy are in | |
1094 | * the same page we don't need to force | |
1095 | * init the buddy | |
1096 | */ | |
2de8807b AG |
1097 | ret = 0; |
1098 | goto err; | |
b6a758ec | 1099 | } |
2de8807b AG |
1100 | /* init buddy cache */ |
1101 | page = e4b.bd_buddy_page; | |
d8aafd0c | 1102 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1103 | if (ret) |
1104 | goto err; | |
1105 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1106 | ret = -EIO; |
1107 | goto err; | |
1108 | } | |
b6a758ec | 1109 | err: |
2de8807b | 1110 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1111 | return ret; |
1112 | } | |
1113 | ||
8a57d9d6 CW |
1114 | /* |
1115 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1116 | * block group lock of all groups for this page; do not hold the BG lock when | |
1117 | * calling this routine! | |
1118 | */ | |
4ddfef7b | 1119 | static noinline_for_stack int |
d8aafd0c KK |
1120 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1121 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1122 | { |
c9de560d AT |
1123 | int blocks_per_page; |
1124 | int block; | |
1125 | int pnum; | |
1126 | int poff; | |
1127 | struct page *page; | |
fdf6c7a7 | 1128 | int ret; |
920313a7 AK |
1129 | struct ext4_group_info *grp; |
1130 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1131 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1132 | |
b10a44c3 | 1133 | might_sleep(); |
6ba495e9 | 1134 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
1135 | |
1136 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 1137 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1138 | |
1139 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1140 | e4b->bd_info = grp; |
c9de560d AT |
1141 | e4b->bd_sb = sb; |
1142 | e4b->bd_group = group; | |
1143 | e4b->bd_buddy_page = NULL; | |
1144 | e4b->bd_bitmap_page = NULL; | |
1145 | ||
f41c0750 | 1146 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1147 | /* |
1148 | * we need full data about the group | |
1149 | * to make a good selection | |
1150 | */ | |
d8aafd0c | 1151 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1152 | if (ret) |
1153 | return ret; | |
f41c0750 AK |
1154 | } |
1155 | ||
c9de560d AT |
1156 | /* |
1157 | * the buddy cache inode stores the block bitmap | |
1158 | * and buddy information in consecutive blocks. | |
1159 | * So for each group we need two blocks. | |
1160 | */ | |
1161 | block = group * 2; | |
1162 | pnum = block / blocks_per_page; | |
1163 | poff = block % blocks_per_page; | |
1164 | ||
1165 | /* we could use find_or_create_page(), but it locks page | |
1166 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1167 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1168 | if (page == NULL || !PageUptodate(page)) { |
1169 | if (page) | |
920313a7 AK |
1170 | /* |
1171 | * drop the page reference and try | |
1172 | * to get the page with lock. If we | |
1173 | * are not uptodate that implies | |
1174 | * somebody just created the page but | |
1175 | * is yet to initialize the same. So | |
1176 | * wait for it to initialize. | |
1177 | */ | |
c9de560d | 1178 | page_cache_release(page); |
d8aafd0c | 1179 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1180 | if (page) { |
1181 | BUG_ON(page->mapping != inode->i_mapping); | |
1182 | if (!PageUptodate(page)) { | |
d8aafd0c | 1183 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1184 | if (ret) { |
1185 | unlock_page(page); | |
1186 | goto err; | |
1187 | } | |
c9de560d AT |
1188 | mb_cmp_bitmaps(e4b, page_address(page) + |
1189 | (poff * sb->s_blocksize)); | |
1190 | } | |
1191 | unlock_page(page); | |
1192 | } | |
1193 | } | |
c57ab39b YL |
1194 | if (page == NULL) { |
1195 | ret = -ENOMEM; | |
1196 | goto err; | |
1197 | } | |
1198 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1199 | ret = -EIO; |
c9de560d | 1200 | goto err; |
fdf6c7a7 | 1201 | } |
2457aec6 MG |
1202 | |
1203 | /* Pages marked accessed already */ | |
c9de560d AT |
1204 | e4b->bd_bitmap_page = page; |
1205 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1206 | |
1207 | block++; | |
1208 | pnum = block / blocks_per_page; | |
1209 | poff = block % blocks_per_page; | |
1210 | ||
2457aec6 | 1211 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1212 | if (page == NULL || !PageUptodate(page)) { |
1213 | if (page) | |
1214 | page_cache_release(page); | |
d8aafd0c | 1215 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1216 | if (page) { |
1217 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1218 | if (!PageUptodate(page)) { |
d8aafd0c KK |
1219 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1220 | gfp); | |
fdf6c7a7 SF |
1221 | if (ret) { |
1222 | unlock_page(page); | |
1223 | goto err; | |
1224 | } | |
1225 | } | |
c9de560d AT |
1226 | unlock_page(page); |
1227 | } | |
1228 | } | |
c57ab39b YL |
1229 | if (page == NULL) { |
1230 | ret = -ENOMEM; | |
1231 | goto err; | |
1232 | } | |
1233 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1234 | ret = -EIO; |
c9de560d | 1235 | goto err; |
fdf6c7a7 | 1236 | } |
2457aec6 MG |
1237 | |
1238 | /* Pages marked accessed already */ | |
c9de560d AT |
1239 | e4b->bd_buddy_page = page; |
1240 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1241 | |
1242 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1243 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1244 | ||
1245 | return 0; | |
1246 | ||
1247 | err: | |
26626f11 YR |
1248 | if (page) |
1249 | page_cache_release(page); | |
c9de560d AT |
1250 | if (e4b->bd_bitmap_page) |
1251 | page_cache_release(e4b->bd_bitmap_page); | |
1252 | if (e4b->bd_buddy_page) | |
1253 | page_cache_release(e4b->bd_buddy_page); | |
1254 | e4b->bd_buddy = NULL; | |
1255 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1256 | return ret; |
c9de560d AT |
1257 | } |
1258 | ||
d8aafd0c KK |
1259 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1260 | struct ext4_buddy *e4b) | |
1261 | { | |
1262 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1263 | } | |
1264 | ||
e39e07fd | 1265 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1266 | { |
1267 | if (e4b->bd_bitmap_page) | |
1268 | page_cache_release(e4b->bd_bitmap_page); | |
1269 | if (e4b->bd_buddy_page) | |
1270 | page_cache_release(e4b->bd_buddy_page); | |
1271 | } | |
1272 | ||
1273 | ||
1274 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1275 | { | |
1276 | int order = 1; | |
12aa7d95 | 1277 | int bb_incr = 1 << (e4b->bd_blkbits - 1); |
c9de560d AT |
1278 | void *bb; |
1279 | ||
c5e8f3f3 | 1280 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1281 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1282 | ||
c5e8f3f3 | 1283 | bb = e4b->bd_buddy; |
c9de560d AT |
1284 | while (order <= e4b->bd_blkbits + 1) { |
1285 | block = block >> 1; | |
1286 | if (!mb_test_bit(block, bb)) { | |
1287 | /* this block is part of buddy of order 'order' */ | |
1288 | return order; | |
1289 | } | |
12aa7d95 NS |
1290 | bb += bb_incr; |
1291 | bb_incr >>= 1; | |
c9de560d AT |
1292 | order++; |
1293 | } | |
1294 | return 0; | |
1295 | } | |
1296 | ||
955ce5f5 | 1297 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1298 | { |
1299 | __u32 *addr; | |
1300 | ||
1301 | len = cur + len; | |
1302 | while (cur < len) { | |
1303 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1304 | /* fast path: clear whole word at once */ | |
1305 | addr = bm + (cur >> 3); | |
1306 | *addr = 0; | |
1307 | cur += 32; | |
1308 | continue; | |
1309 | } | |
955ce5f5 | 1310 | mb_clear_bit(cur, bm); |
c9de560d AT |
1311 | cur++; |
1312 | } | |
1313 | } | |
1314 | ||
eabe0444 AS |
1315 | /* clear bits in given range |
1316 | * will return first found zero bit if any, -1 otherwise | |
1317 | */ | |
1318 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1319 | { | |
1320 | __u32 *addr; | |
1321 | int zero_bit = -1; | |
1322 | ||
1323 | len = cur + len; | |
1324 | while (cur < len) { | |
1325 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1326 | /* fast path: clear whole word at once */ | |
1327 | addr = bm + (cur >> 3); | |
1328 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1329 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1330 | *addr = 0; | |
1331 | cur += 32; | |
1332 | continue; | |
1333 | } | |
1334 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1335 | zero_bit = cur; | |
1336 | cur++; | |
1337 | } | |
1338 | ||
1339 | return zero_bit; | |
1340 | } | |
1341 | ||
c3e94d1d | 1342 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1343 | { |
1344 | __u32 *addr; | |
1345 | ||
1346 | len = cur + len; | |
1347 | while (cur < len) { | |
1348 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1349 | /* fast path: set whole word at once */ | |
1350 | addr = bm + (cur >> 3); | |
1351 | *addr = 0xffffffff; | |
1352 | cur += 32; | |
1353 | continue; | |
1354 | } | |
955ce5f5 | 1355 | mb_set_bit(cur, bm); |
c9de560d AT |
1356 | cur++; |
1357 | } | |
1358 | } | |
1359 | ||
eabe0444 AS |
1360 | /* |
1361 | * _________________________________________________________________ */ | |
1362 | ||
1363 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) | |
1364 | { | |
1365 | if (mb_test_bit(*bit + side, bitmap)) { | |
1366 | mb_clear_bit(*bit, bitmap); | |
1367 | (*bit) -= side; | |
1368 | return 1; | |
1369 | } | |
1370 | else { | |
1371 | (*bit) += side; | |
1372 | mb_set_bit(*bit, bitmap); | |
1373 | return -1; | |
1374 | } | |
1375 | } | |
1376 | ||
1377 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1378 | { | |
1379 | int max; | |
1380 | int order = 1; | |
1381 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1382 | ||
1383 | while (buddy) { | |
1384 | void *buddy2; | |
1385 | ||
1386 | /* Bits in range [first; last] are known to be set since | |
1387 | * corresponding blocks were allocated. Bits in range | |
1388 | * (first; last) will stay set because they form buddies on | |
1389 | * upper layer. We just deal with borders if they don't | |
1390 | * align with upper layer and then go up. | |
1391 | * Releasing entire group is all about clearing | |
1392 | * single bit of highest order buddy. | |
1393 | */ | |
1394 | ||
1395 | /* Example: | |
1396 | * --------------------------------- | |
1397 | * | 1 | 1 | 1 | 1 | | |
1398 | * --------------------------------- | |
1399 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1400 | * --------------------------------- | |
1401 | * 0 1 2 3 4 5 6 7 | |
1402 | * \_____________________/ | |
1403 | * | |
1404 | * Neither [1] nor [6] is aligned to above layer. | |
1405 | * Left neighbour [0] is free, so mark it busy, | |
1406 | * decrease bb_counters and extend range to | |
1407 | * [0; 6] | |
1408 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1409 | * mark [6] free, increase bb_counters and shrink range to | |
1410 | * [0; 5]. | |
1411 | * Then shift range to [0; 2], go up and do the same. | |
1412 | */ | |
1413 | ||
1414 | ||
1415 | if (first & 1) | |
1416 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1417 | if (!(last & 1)) | |
1418 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1419 | if (first > last) | |
1420 | break; | |
1421 | order++; | |
1422 | ||
1423 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1424 | mb_clear_bits(buddy, first, last - first + 1); | |
1425 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1426 | break; | |
1427 | } | |
1428 | first >>= 1; | |
1429 | last >>= 1; | |
1430 | buddy = buddy2; | |
1431 | } | |
1432 | } | |
1433 | ||
7e5a8cdd | 1434 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1435 | int first, int count) |
c9de560d | 1436 | { |
eabe0444 AS |
1437 | int left_is_free = 0; |
1438 | int right_is_free = 0; | |
1439 | int block; | |
1440 | int last = first + count - 1; | |
c9de560d AT |
1441 | struct super_block *sb = e4b->bd_sb; |
1442 | ||
c99d1e6e TT |
1443 | if (WARN_ON(count == 0)) |
1444 | return; | |
eabe0444 | 1445 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1446 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1447 | /* Don't bother if the block group is corrupt. */ |
1448 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1449 | return; | |
1450 | ||
c9de560d AT |
1451 | mb_check_buddy(e4b); |
1452 | mb_free_blocks_double(inode, e4b, first, count); | |
1453 | ||
1454 | e4b->bd_info->bb_free += count; | |
1455 | if (first < e4b->bd_info->bb_first_free) | |
1456 | e4b->bd_info->bb_first_free = first; | |
1457 | ||
eabe0444 AS |
1458 | /* access memory sequentially: check left neighbour, |
1459 | * clear range and then check right neighbour | |
1460 | */ | |
c9de560d | 1461 | if (first != 0) |
eabe0444 AS |
1462 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1463 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1464 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1465 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1466 | ||
1467 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1468 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
1cac41cb MB |
1469 | /* for debugging, sangwoo2.lee */ |
1470 | struct ext4_group_desc *desc; | |
1471 | ext4_fsblk_t blocknr, bitmap_blk; | |
1472 | ||
1473 | desc = ext4_get_group_desc(sb, e4b->bd_group, NULL); | |
1474 | bitmap_blk = ext4_block_bitmap(sb, desc); | |
eabe0444 AS |
1475 | |
1476 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
1477 | blocknr += EXT4_C2B(EXT4_SB(sb), block); | |
1cac41cb MB |
1478 | |
1479 | print_block_data(sb, bitmap_blk, e4b->bd_bitmap, 0 | |
1480 | , EXT4_BLOCK_SIZE(sb)); | |
1481 | /* for debugging */ | |
eabe0444 AS |
1482 | ext4_grp_locked_error(sb, e4b->bd_group, |
1483 | inode ? inode->i_ino : 0, | |
1484 | blocknr, | |
1485 | "freeing already freed block " | |
163a203d DW |
1486 | "(bit %u); block bitmap corrupt.", |
1487 | block); | |
e43bb4e6 NJ |
1488 | if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)) |
1489 | percpu_counter_sub(&sbi->s_freeclusters_counter, | |
1490 | e4b->bd_info->bb_free); | |
163a203d DW |
1491 | /* Mark the block group as corrupt. */ |
1492 | set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, | |
1493 | &e4b->bd_info->bb_state); | |
eabe0444 AS |
1494 | mb_regenerate_buddy(e4b); |
1495 | goto done; | |
1496 | } | |
1497 | ||
1498 | /* let's maintain fragments counter */ | |
1499 | if (left_is_free && right_is_free) | |
c9de560d | 1500 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1501 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1502 | e4b->bd_info->bb_fragments++; |
1503 | ||
eabe0444 AS |
1504 | /* buddy[0] == bd_bitmap is a special case, so handle |
1505 | * it right away and let mb_buddy_mark_free stay free of | |
1506 | * zero order checks. | |
1507 | * Check if neighbours are to be coaleasced, | |
1508 | * adjust bitmap bb_counters and borders appropriately. | |
1509 | */ | |
1510 | if (first & 1) { | |
1511 | first += !left_is_free; | |
1512 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1513 | } | |
1514 | if (!(last & 1)) { | |
1515 | last -= !right_is_free; | |
1516 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1517 | } | |
c9de560d | 1518 | |
eabe0444 AS |
1519 | if (first <= last) |
1520 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1521 | |
eabe0444 | 1522 | done: |
8a57d9d6 | 1523 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1524 | mb_check_buddy(e4b); |
c9de560d AT |
1525 | } |
1526 | ||
15c006a2 | 1527 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1528 | int needed, struct ext4_free_extent *ex) |
1529 | { | |
1530 | int next = block; | |
15c006a2 | 1531 | int max, order; |
c9de560d AT |
1532 | void *buddy; |
1533 | ||
bc8e6740 | 1534 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1535 | BUG_ON(ex == NULL); |
1536 | ||
15c006a2 | 1537 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1538 | BUG_ON(buddy == NULL); |
1539 | BUG_ON(block >= max); | |
1540 | if (mb_test_bit(block, buddy)) { | |
1541 | ex->fe_len = 0; | |
1542 | ex->fe_start = 0; | |
1543 | ex->fe_group = 0; | |
1544 | return 0; | |
1545 | } | |
1546 | ||
15c006a2 RD |
1547 | /* find actual order */ |
1548 | order = mb_find_order_for_block(e4b, block); | |
1549 | block = block >> order; | |
c9de560d AT |
1550 | |
1551 | ex->fe_len = 1 << order; | |
1552 | ex->fe_start = block << order; | |
1553 | ex->fe_group = e4b->bd_group; | |
1554 | ||
1555 | /* calc difference from given start */ | |
1556 | next = next - ex->fe_start; | |
1557 | ex->fe_len -= next; | |
1558 | ex->fe_start += next; | |
1559 | ||
1560 | while (needed > ex->fe_len && | |
d8ec0c39 | 1561 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1562 | |
1563 | if (block + 1 >= max) | |
1564 | break; | |
1565 | ||
1566 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1567 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1568 | break; |
1569 | ||
b051d8dc | 1570 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1571 | |
c9de560d AT |
1572 | block = next >> order; |
1573 | ex->fe_len += 1 << order; | |
1574 | } | |
1575 | ||
1576 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1577 | return ex->fe_len; | |
1578 | } | |
1579 | ||
1580 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1581 | { | |
1582 | int ord; | |
1583 | int mlen = 0; | |
1584 | int max = 0; | |
1585 | int cur; | |
1586 | int start = ex->fe_start; | |
1587 | int len = ex->fe_len; | |
1588 | unsigned ret = 0; | |
1589 | int len0 = len; | |
1590 | void *buddy; | |
1591 | ||
1592 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1593 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1594 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1595 | mb_check_buddy(e4b); |
1596 | mb_mark_used_double(e4b, start, len); | |
1597 | ||
1598 | e4b->bd_info->bb_free -= len; | |
1599 | if (e4b->bd_info->bb_first_free == start) | |
1600 | e4b->bd_info->bb_first_free += len; | |
1601 | ||
1602 | /* let's maintain fragments counter */ | |
1603 | if (start != 0) | |
c5e8f3f3 | 1604 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1605 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1606 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1607 | if (mlen && max) |
1608 | e4b->bd_info->bb_fragments++; | |
1609 | else if (!mlen && !max) | |
1610 | e4b->bd_info->bb_fragments--; | |
1611 | ||
1612 | /* let's maintain buddy itself */ | |
1613 | while (len) { | |
1614 | ord = mb_find_order_for_block(e4b, start); | |
1615 | ||
1616 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1617 | /* the whole chunk may be allocated at once! */ | |
1618 | mlen = 1 << ord; | |
1619 | buddy = mb_find_buddy(e4b, ord, &max); | |
1620 | BUG_ON((start >> ord) >= max); | |
1621 | mb_set_bit(start >> ord, buddy); | |
1622 | e4b->bd_info->bb_counters[ord]--; | |
1623 | start += mlen; | |
1624 | len -= mlen; | |
1625 | BUG_ON(len < 0); | |
1626 | continue; | |
1627 | } | |
1628 | ||
1629 | /* store for history */ | |
1630 | if (ret == 0) | |
1631 | ret = len | (ord << 16); | |
1632 | ||
1633 | /* we have to split large buddy */ | |
1634 | BUG_ON(ord <= 0); | |
1635 | buddy = mb_find_buddy(e4b, ord, &max); | |
1636 | mb_set_bit(start >> ord, buddy); | |
1637 | e4b->bd_info->bb_counters[ord]--; | |
1638 | ||
1639 | ord--; | |
1640 | cur = (start >> ord) & ~1U; | |
1641 | buddy = mb_find_buddy(e4b, ord, &max); | |
1642 | mb_clear_bit(cur, buddy); | |
1643 | mb_clear_bit(cur + 1, buddy); | |
1644 | e4b->bd_info->bb_counters[ord]++; | |
1645 | e4b->bd_info->bb_counters[ord]++; | |
1646 | } | |
8a57d9d6 | 1647 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1648 | |
c5e8f3f3 | 1649 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1650 | mb_check_buddy(e4b); |
1651 | ||
1652 | return ret; | |
1653 | } | |
1654 | ||
1655 | /* | |
1656 | * Must be called under group lock! | |
1657 | */ | |
1658 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1659 | struct ext4_buddy *e4b) | |
1660 | { | |
1661 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1662 | int ret; | |
1663 | ||
1664 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1665 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1666 | ||
1667 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1668 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1669 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1670 | ||
1671 | /* preallocation can change ac_b_ex, thus we store actually | |
1672 | * allocated blocks for history */ | |
1673 | ac->ac_f_ex = ac->ac_b_ex; | |
1674 | ||
1675 | ac->ac_status = AC_STATUS_FOUND; | |
1676 | ac->ac_tail = ret & 0xffff; | |
1677 | ac->ac_buddy = ret >> 16; | |
1678 | ||
c3a326a6 AK |
1679 | /* |
1680 | * take the page reference. We want the page to be pinned | |
1681 | * so that we don't get a ext4_mb_init_cache_call for this | |
1682 | * group until we update the bitmap. That would mean we | |
1683 | * double allocate blocks. The reference is dropped | |
1684 | * in ext4_mb_release_context | |
1685 | */ | |
c9de560d AT |
1686 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1687 | get_page(ac->ac_bitmap_page); | |
1688 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1689 | get_page(ac->ac_buddy_page); | |
c9de560d | 1690 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1691 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1692 | spin_lock(&sbi->s_md_lock); |
1693 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1694 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1695 | spin_unlock(&sbi->s_md_lock); | |
1696 | } | |
1697 | } | |
1698 | ||
1699 | /* | |
1700 | * regular allocator, for general purposes allocation | |
1701 | */ | |
1702 | ||
1703 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1704 | struct ext4_buddy *e4b, | |
1705 | int finish_group) | |
1706 | { | |
1707 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1708 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1709 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1710 | struct ext4_free_extent ex; | |
1711 | int max; | |
1712 | ||
032115fc AK |
1713 | if (ac->ac_status == AC_STATUS_FOUND) |
1714 | return; | |
c9de560d AT |
1715 | /* |
1716 | * We don't want to scan for a whole year | |
1717 | */ | |
1718 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1719 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1720 | ac->ac_status = AC_STATUS_BREAK; | |
1721 | return; | |
1722 | } | |
1723 | ||
1724 | /* | |
1725 | * Haven't found good chunk so far, let's continue | |
1726 | */ | |
1727 | if (bex->fe_len < gex->fe_len) | |
1728 | return; | |
1729 | ||
1730 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1731 | && bex->fe_group == e4b->bd_group) { | |
1732 | /* recheck chunk's availability - we don't know | |
1733 | * when it was found (within this lock-unlock | |
1734 | * period or not) */ | |
15c006a2 | 1735 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1736 | if (max >= gex->fe_len) { |
1737 | ext4_mb_use_best_found(ac, e4b); | |
1738 | return; | |
1739 | } | |
1740 | } | |
1741 | } | |
1742 | ||
1743 | /* | |
1744 | * The routine checks whether found extent is good enough. If it is, | |
1745 | * then the extent gets marked used and flag is set to the context | |
1746 | * to stop scanning. Otherwise, the extent is compared with the | |
1747 | * previous found extent and if new one is better, then it's stored | |
1748 | * in the context. Later, the best found extent will be used, if | |
1749 | * mballoc can't find good enough extent. | |
1750 | * | |
1751 | * FIXME: real allocation policy is to be designed yet! | |
1752 | */ | |
1753 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1754 | struct ext4_free_extent *ex, | |
1755 | struct ext4_buddy *e4b) | |
1756 | { | |
1757 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1758 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1759 | ||
1760 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1761 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1762 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1763 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1764 | ||
1765 | ac->ac_found++; | |
1766 | ||
1767 | /* | |
1768 | * The special case - take what you catch first | |
1769 | */ | |
1770 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1771 | *bex = *ex; | |
1772 | ext4_mb_use_best_found(ac, e4b); | |
1773 | return; | |
1774 | } | |
1775 | ||
1776 | /* | |
1777 | * Let's check whether the chuck is good enough | |
1778 | */ | |
1779 | if (ex->fe_len == gex->fe_len) { | |
1780 | *bex = *ex; | |
1781 | ext4_mb_use_best_found(ac, e4b); | |
1782 | return; | |
1783 | } | |
1784 | ||
1785 | /* | |
1786 | * If this is first found extent, just store it in the context | |
1787 | */ | |
1788 | if (bex->fe_len == 0) { | |
1789 | *bex = *ex; | |
1790 | return; | |
1791 | } | |
1792 | ||
1793 | /* | |
1794 | * If new found extent is better, store it in the context | |
1795 | */ | |
1796 | if (bex->fe_len < gex->fe_len) { | |
1797 | /* if the request isn't satisfied, any found extent | |
1798 | * larger than previous best one is better */ | |
1799 | if (ex->fe_len > bex->fe_len) | |
1800 | *bex = *ex; | |
1801 | } else if (ex->fe_len > gex->fe_len) { | |
1802 | /* if the request is satisfied, then we try to find | |
1803 | * an extent that still satisfy the request, but is | |
1804 | * smaller than previous one */ | |
1805 | if (ex->fe_len < bex->fe_len) | |
1806 | *bex = *ex; | |
1807 | } | |
1808 | ||
1809 | ext4_mb_check_limits(ac, e4b, 0); | |
1810 | } | |
1811 | ||
089ceecc ES |
1812 | static noinline_for_stack |
1813 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1814 | struct ext4_buddy *e4b) |
1815 | { | |
1816 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1817 | ext4_group_t group = ex.fe_group; | |
1818 | int max; | |
1819 | int err; | |
1820 | ||
1821 | BUG_ON(ex.fe_len <= 0); | |
1822 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1823 | if (err) | |
1824 | return err; | |
1825 | ||
1826 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1827 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1828 | |
1829 | if (max > 0) { | |
1830 | ac->ac_b_ex = ex; | |
1831 | ext4_mb_use_best_found(ac, e4b); | |
1832 | } | |
1833 | ||
1834 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1835 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1836 | |
1837 | return 0; | |
1838 | } | |
1839 | ||
089ceecc ES |
1840 | static noinline_for_stack |
1841 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1842 | struct ext4_buddy *e4b) |
1843 | { | |
1844 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1845 | int max; | |
1846 | int err; | |
1847 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1848 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1849 | struct ext4_free_extent ex; |
1850 | ||
1851 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1852 | return 0; | |
838cd0cf YY |
1853 | if (grp->bb_free == 0) |
1854 | return 0; | |
c9de560d AT |
1855 | |
1856 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1857 | if (err) | |
1858 | return err; | |
1859 | ||
163a203d DW |
1860 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
1861 | ext4_mb_unload_buddy(e4b); | |
1862 | return 0; | |
1863 | } | |
1864 | ||
c9de560d | 1865 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 1866 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 1867 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 1868 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
1869 | |
1870 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1871 | ext4_fsblk_t start; | |
1872 | ||
5661bd68 AM |
1873 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1874 | ex.fe_start; | |
c9de560d AT |
1875 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1876 | if (do_div(start, sbi->s_stripe) == 0) { | |
1877 | ac->ac_found++; | |
1878 | ac->ac_b_ex = ex; | |
1879 | ext4_mb_use_best_found(ac, e4b); | |
1880 | } | |
1881 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1882 | BUG_ON(ex.fe_len <= 0); | |
1883 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1884 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1885 | ac->ac_found++; | |
1886 | ac->ac_b_ex = ex; | |
1887 | ext4_mb_use_best_found(ac, e4b); | |
1888 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1889 | /* Sometimes, caller may want to merge even small | |
1890 | * number of blocks to an existing extent */ | |
1891 | BUG_ON(ex.fe_len <= 0); | |
1892 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1893 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1894 | ac->ac_found++; | |
1895 | ac->ac_b_ex = ex; | |
1896 | ext4_mb_use_best_found(ac, e4b); | |
1897 | } | |
1898 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1899 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1900 | |
1901 | return 0; | |
1902 | } | |
1903 | ||
1904 | /* | |
1905 | * The routine scans buddy structures (not bitmap!) from given order | |
1906 | * to max order and tries to find big enough chunk to satisfy the req | |
1907 | */ | |
089ceecc ES |
1908 | static noinline_for_stack |
1909 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1910 | struct ext4_buddy *e4b) |
1911 | { | |
1912 | struct super_block *sb = ac->ac_sb; | |
1913 | struct ext4_group_info *grp = e4b->bd_info; | |
1914 | void *buddy; | |
1915 | int i; | |
1916 | int k; | |
1917 | int max; | |
1918 | ||
1919 | BUG_ON(ac->ac_2order <= 0); | |
1920 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1921 | if (grp->bb_counters[i] == 0) | |
1922 | continue; | |
1923 | ||
1924 | buddy = mb_find_buddy(e4b, i, &max); | |
1925 | BUG_ON(buddy == NULL); | |
1926 | ||
ffad0a44 | 1927 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1928 | BUG_ON(k >= max); |
1929 | ||
1930 | ac->ac_found++; | |
1931 | ||
1932 | ac->ac_b_ex.fe_len = 1 << i; | |
1933 | ac->ac_b_ex.fe_start = k << i; | |
1934 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1935 | ||
1936 | ext4_mb_use_best_found(ac, e4b); | |
1937 | ||
1938 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1939 | ||
1940 | if (EXT4_SB(sb)->s_mb_stats) | |
1941 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1942 | ||
1943 | break; | |
1944 | } | |
1945 | } | |
1946 | ||
1947 | /* | |
1948 | * The routine scans the group and measures all found extents. | |
1949 | * In order to optimize scanning, caller must pass number of | |
1950 | * free blocks in the group, so the routine can know upper limit. | |
1951 | */ | |
089ceecc ES |
1952 | static noinline_for_stack |
1953 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1954 | struct ext4_buddy *e4b) |
1955 | { | |
1956 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 1957 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1958 | struct ext4_free_extent ex; |
1959 | int i; | |
1960 | int free; | |
1961 | ||
1962 | free = e4b->bd_info->bb_free; | |
1963 | BUG_ON(free <= 0); | |
1964 | ||
1965 | i = e4b->bd_info->bb_first_free; | |
1966 | ||
1967 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1968 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1969 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1970 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1971 | /* |
e56eb659 | 1972 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1973 | * free blocks even though group info says we |
1974 | * we have free blocks | |
1975 | */ | |
e29136f8 | 1976 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1977 | "%d free clusters as per " |
fde4d95a | 1978 | "group info. But bitmap says 0", |
26346ff6 | 1979 | free); |
c9de560d AT |
1980 | break; |
1981 | } | |
1982 | ||
15c006a2 | 1983 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
c9de560d | 1984 | BUG_ON(ex.fe_len <= 0); |
26346ff6 | 1985 | if (free < ex.fe_len) { |
e29136f8 | 1986 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1987 | "%d free clusters as per " |
fde4d95a | 1988 | "group info. But got %d blocks", |
26346ff6 | 1989 | free, ex.fe_len); |
e56eb659 AK |
1990 | /* |
1991 | * The number of free blocks differs. This mostly | |
1992 | * indicate that the bitmap is corrupt. So exit | |
1993 | * without claiming the space. | |
1994 | */ | |
1995 | break; | |
26346ff6 | 1996 | } |
ab0c00fc | 1997 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
1998 | ext4_mb_measure_extent(ac, &ex, e4b); |
1999 | ||
2000 | i += ex.fe_len; | |
2001 | free -= ex.fe_len; | |
2002 | } | |
2003 | ||
2004 | ext4_mb_check_limits(ac, e4b, 1); | |
2005 | } | |
2006 | ||
2007 | /* | |
2008 | * This is a special case for storages like raid5 | |
506bf2d8 | 2009 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2010 | */ |
089ceecc ES |
2011 | static noinline_for_stack |
2012 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2013 | struct ext4_buddy *e4b) |
2014 | { | |
2015 | struct super_block *sb = ac->ac_sb; | |
2016 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2017 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2018 | struct ext4_free_extent ex; |
2019 | ext4_fsblk_t first_group_block; | |
2020 | ext4_fsblk_t a; | |
2021 | ext4_grpblk_t i; | |
2022 | int max; | |
2023 | ||
2024 | BUG_ON(sbi->s_stripe == 0); | |
2025 | ||
2026 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2027 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2028 | ||
c9de560d AT |
2029 | a = first_group_block + sbi->s_stripe - 1; |
2030 | do_div(a, sbi->s_stripe); | |
2031 | i = (a * sbi->s_stripe) - first_group_block; | |
2032 | ||
7137d7a4 | 2033 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2034 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2035 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2036 | if (max >= sbi->s_stripe) { |
2037 | ac->ac_found++; | |
ab0c00fc | 2038 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2039 | ac->ac_b_ex = ex; |
2040 | ext4_mb_use_best_found(ac, e4b); | |
2041 | break; | |
2042 | } | |
2043 | } | |
2044 | i += sbi->s_stripe; | |
2045 | } | |
2046 | } | |
2047 | ||
42ac1848 LC |
2048 | /* |
2049 | * This is now called BEFORE we load the buddy bitmap. | |
2050 | * Returns either 1 or 0 indicating that the group is either suitable | |
2051 | * for the allocation or not. In addition it can also return negative | |
2052 | * error code when something goes wrong. | |
2053 | */ | |
c9de560d AT |
2054 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
2055 | ext4_group_t group, int cr) | |
2056 | { | |
2057 | unsigned free, fragments; | |
a4912123 | 2058 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2059 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2060 | ||
2061 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2062 | |
01fc48e8 TT |
2063 | free = grp->bb_free; |
2064 | if (free == 0) | |
2065 | return 0; | |
2066 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2067 | return 0; | |
2068 | ||
163a203d DW |
2069 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
2070 | return 0; | |
2071 | ||
8a57d9d6 CW |
2072 | /* We only do this if the grp has never been initialized */ |
2073 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
d8aafd0c | 2074 | int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS); |
8a57d9d6 | 2075 | if (ret) |
42ac1848 | 2076 | return ret; |
8a57d9d6 | 2077 | } |
c9de560d | 2078 | |
c9de560d | 2079 | fragments = grp->bb_fragments; |
c9de560d AT |
2080 | if (fragments == 0) |
2081 | return 0; | |
2082 | ||
2083 | switch (cr) { | |
2084 | case 0: | |
2085 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2086 | |
a4912123 TT |
2087 | /* Avoid using the first bg of a flexgroup for data files */ |
2088 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2089 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2090 | ((group % flex_size) == 0)) | |
2091 | return 0; | |
2092 | ||
40ae3487 TT |
2093 | if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || |
2094 | (free / fragments) >= ac->ac_g_ex.fe_len) | |
2095 | return 1; | |
2096 | ||
2097 | if (grp->bb_largest_free_order < ac->ac_2order) | |
2098 | return 0; | |
2099 | ||
8a57d9d6 | 2100 | return 1; |
c9de560d AT |
2101 | case 1: |
2102 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
2103 | return 1; | |
2104 | break; | |
2105 | case 2: | |
2106 | if (free >= ac->ac_g_ex.fe_len) | |
2107 | return 1; | |
2108 | break; | |
2109 | case 3: | |
2110 | return 1; | |
2111 | default: | |
2112 | BUG(); | |
2113 | } | |
2114 | ||
2115 | return 0; | |
2116 | } | |
2117 | ||
4ddfef7b ES |
2118 | static noinline_for_stack int |
2119 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2120 | { |
8df9675f | 2121 | ext4_group_t ngroups, group, i; |
c9de560d | 2122 | int cr; |
42ac1848 | 2123 | int err = 0, first_err = 0; |
c9de560d AT |
2124 | struct ext4_sb_info *sbi; |
2125 | struct super_block *sb; | |
2126 | struct ext4_buddy e4b; | |
c9de560d AT |
2127 | |
2128 | sb = ac->ac_sb; | |
2129 | sbi = EXT4_SB(sb); | |
8df9675f | 2130 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2131 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2132 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2133 | ngroups = sbi->s_blockfile_groups; |
2134 | ||
c9de560d AT |
2135 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2136 | ||
2137 | /* first, try the goal */ | |
2138 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2139 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2140 | goto out; | |
2141 | ||
2142 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2143 | goto out; | |
2144 | ||
2145 | /* | |
2146 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
2147 | * if ac2_order is set we also set criteria to 0 so that we | |
2148 | * try exact allocation using buddy. | |
2149 | */ | |
2150 | i = fls(ac->ac_g_ex.fe_len); | |
2151 | ac->ac_2order = 0; | |
2152 | /* | |
2153 | * We search using buddy data only if the order of the request | |
2154 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2155 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
5624ea16 JK |
2156 | * We also support searching for power-of-two requests only for |
2157 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2158 | */ |
5624ea16 | 2159 | if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) { |
c9de560d AT |
2160 | /* |
2161 | * This should tell if fe_len is exactly power of 2 | |
2162 | */ | |
2163 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
2164 | ac->ac_2order = i - 1; | |
2165 | } | |
2166 | ||
4ba74d00 TT |
2167 | /* if stream allocation is enabled, use global goal */ |
2168 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2169 | /* TBD: may be hot point */ |
2170 | spin_lock(&sbi->s_md_lock); | |
2171 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2172 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2173 | spin_unlock(&sbi->s_md_lock); | |
2174 | } | |
4ba74d00 | 2175 | |
c9de560d AT |
2176 | /* Let's just scan groups to find more-less suitable blocks */ |
2177 | cr = ac->ac_2order ? 0 : 1; | |
2178 | /* | |
2179 | * cr == 0 try to get exact allocation, | |
2180 | * cr == 3 try to get anything | |
2181 | */ | |
2182 | repeat: | |
2183 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2184 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2185 | /* |
2186 | * searching for the right group start | |
2187 | * from the goal value specified | |
2188 | */ | |
2189 | group = ac->ac_g_ex.fe_group; | |
2190 | ||
8df9675f | 2191 | for (i = 0; i < ngroups; group++, i++) { |
42ac1848 | 2192 | int ret = 0; |
2ed5724d | 2193 | cond_resched(); |
e6155736 LM |
2194 | /* |
2195 | * Artificially restricted ngroups for non-extent | |
2196 | * files makes group > ngroups possible on first loop. | |
2197 | */ | |
2198 | if (group >= ngroups) | |
c9de560d AT |
2199 | group = 0; |
2200 | ||
8a57d9d6 | 2201 | /* This now checks without needing the buddy page */ |
42ac1848 LC |
2202 | ret = ext4_mb_good_group(ac, group, cr); |
2203 | if (ret <= 0) { | |
2204 | if (!first_err) | |
2205 | first_err = ret; | |
c9de560d | 2206 | continue; |
42ac1848 | 2207 | } |
c9de560d | 2208 | |
c9de560d AT |
2209 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2210 | if (err) | |
2211 | goto out; | |
2212 | ||
2213 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2214 | |
2215 | /* | |
2216 | * We need to check again after locking the | |
2217 | * block group | |
2218 | */ | |
42ac1848 LC |
2219 | ret = ext4_mb_good_group(ac, group, cr); |
2220 | if (ret <= 0) { | |
c9de560d | 2221 | ext4_unlock_group(sb, group); |
e39e07fd | 2222 | ext4_mb_unload_buddy(&e4b); |
42ac1848 LC |
2223 | if (!first_err) |
2224 | first_err = ret; | |
c9de560d AT |
2225 | continue; |
2226 | } | |
2227 | ||
2228 | ac->ac_groups_scanned++; | |
5624ea16 | 2229 | if (cr == 0) |
c9de560d | 2230 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2231 | else if (cr == 1 && sbi->s_stripe && |
2232 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2233 | ext4_mb_scan_aligned(ac, &e4b); |
2234 | else | |
2235 | ext4_mb_complex_scan_group(ac, &e4b); | |
2236 | ||
2237 | ext4_unlock_group(sb, group); | |
e39e07fd | 2238 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2239 | |
2240 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2241 | break; | |
2242 | } | |
2243 | } | |
2244 | ||
2245 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2246 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2247 | /* | |
2248 | * We've been searching too long. Let's try to allocate | |
2249 | * the best chunk we've found so far | |
2250 | */ | |
2251 | ||
2252 | ext4_mb_try_best_found(ac, &e4b); | |
2253 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2254 | /* | |
2255 | * Someone more lucky has already allocated it. | |
2256 | * The only thing we can do is just take first | |
2257 | * found block(s) | |
2258 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2259 | */ | |
2260 | ac->ac_b_ex.fe_group = 0; | |
2261 | ac->ac_b_ex.fe_start = 0; | |
2262 | ac->ac_b_ex.fe_len = 0; | |
2263 | ac->ac_status = AC_STATUS_CONTINUE; | |
2264 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2265 | cr = 3; | |
2266 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2267 | goto repeat; | |
2268 | } | |
2269 | } | |
2270 | out: | |
42ac1848 LC |
2271 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2272 | err = first_err; | |
c9de560d AT |
2273 | return err; |
2274 | } | |
2275 | ||
c9de560d AT |
2276 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2277 | { | |
2278 | struct super_block *sb = seq->private; | |
c9de560d AT |
2279 | ext4_group_t group; |
2280 | ||
8df9675f | 2281 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2282 | return NULL; |
c9de560d | 2283 | group = *pos + 1; |
a9df9a49 | 2284 | return (void *) ((unsigned long) group); |
c9de560d AT |
2285 | } |
2286 | ||
2287 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2288 | { | |
2289 | struct super_block *sb = seq->private; | |
c9de560d AT |
2290 | ext4_group_t group; |
2291 | ||
2292 | ++*pos; | |
8df9675f | 2293 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2294 | return NULL; |
2295 | group = *pos + 1; | |
a9df9a49 | 2296 | return (void *) ((unsigned long) group); |
c9de560d AT |
2297 | } |
2298 | ||
2299 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2300 | { | |
2301 | struct super_block *sb = seq->private; | |
a9df9a49 | 2302 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2303 | int i; |
1c8457ca | 2304 | int err, buddy_loaded = 0; |
c9de560d | 2305 | struct ext4_buddy e4b; |
1c8457ca | 2306 | struct ext4_group_info *grinfo; |
c9de560d AT |
2307 | struct sg { |
2308 | struct ext4_group_info info; | |
7b74c351 | 2309 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2310 | } sg; |
2311 | ||
2312 | group--; | |
2313 | if (group == 0) | |
97b4af2f RV |
2314 | seq_puts(seq, "#group: free frags first [" |
2315 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
2316 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]"); | |
c9de560d AT |
2317 | |
2318 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2319 | sizeof(struct ext4_group_info); | |
1c8457ca AK |
2320 | grinfo = ext4_get_group_info(sb, group); |
2321 | /* Load the group info in memory only if not already loaded. */ | |
2322 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2323 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2324 | if (err) { | |
2325 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2326 | return 0; | |
2327 | } | |
2328 | buddy_loaded = 1; | |
c9de560d | 2329 | } |
1c8457ca | 2330 | |
c9de560d | 2331 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2332 | |
2333 | if (buddy_loaded) | |
2334 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2335 | |
a9df9a49 | 2336 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2337 | sg.info.bb_fragments, sg.info.bb_first_free); |
2338 | for (i = 0; i <= 13; i++) | |
2339 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2340 | sg.info.bb_counters[i] : 0); | |
2341 | seq_printf(seq, " ]\n"); | |
2342 | ||
2343 | return 0; | |
2344 | } | |
2345 | ||
2346 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2347 | { | |
2348 | } | |
2349 | ||
7f1346a9 | 2350 | static const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2351 | .start = ext4_mb_seq_groups_start, |
2352 | .next = ext4_mb_seq_groups_next, | |
2353 | .stop = ext4_mb_seq_groups_stop, | |
2354 | .show = ext4_mb_seq_groups_show, | |
2355 | }; | |
2356 | ||
2357 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2358 | { | |
d9dda78b | 2359 | struct super_block *sb = PDE_DATA(inode); |
c9de560d AT |
2360 | int rc; |
2361 | ||
2362 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2363 | if (rc == 0) { | |
a271fe85 | 2364 | struct seq_file *m = file->private_data; |
c9de560d AT |
2365 | m->private = sb; |
2366 | } | |
2367 | return rc; | |
2368 | ||
2369 | } | |
2370 | ||
ebd173be | 2371 | const struct file_operations ext4_seq_mb_groups_fops = { |
c9de560d AT |
2372 | .owner = THIS_MODULE, |
2373 | .open = ext4_mb_seq_groups_open, | |
2374 | .read = seq_read, | |
2375 | .llseek = seq_lseek, | |
2376 | .release = seq_release, | |
2377 | }; | |
2378 | ||
1cac41cb MB |
2379 | ssize_t ext4_mb_freefrag_show(struct ext4_sb_info *sbi, char *buf) |
2380 | { | |
2381 | #define EXT4_FREEFRAG_COLUMN 14 /* sb->s_blocksize_bits + 2 */ | |
2382 | ext4_group_t group = 0; | |
2383 | int i; | |
2384 | ext4_fsblk_t freeblock[EXT4_FREEFRAG_COLUMN] = {0,}; | |
2385 | char *size[EXT4_FREEFRAG_COLUMN] = {"4K","8K","16K","32K","64K", | |
2386 | "128K","256K","512K","1M","2M","4M","8M","16M","32M"}; | |
2387 | ||
2388 | for (group = 0; group < sbi->s_groups_count; group++) { | |
2389 | struct super_block *sb = sbi->s_sb; | |
2390 | int err, buddy_loaded = 0; | |
2391 | struct ext4_buddy e4b; | |
2392 | struct ext4_group_info *grinfo; | |
2393 | struct sg { | |
2394 | struct ext4_group_info info; | |
2395 | ext4_grpblk_t counters[EXT4_FREEFRAG_COLUMN+2]; | |
2396 | } sg; | |
2397 | ||
2398 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2399 | sizeof(struct ext4_group_info); | |
2400 | grinfo = ext4_get_group_info(sb, group); | |
2401 | /* Load the group info in memory only if not already loaded. */ | |
2402 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2403 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2404 | if (err) { | |
2405 | freeblock[0] = ULLONG_MAX; | |
2406 | goto out; | |
2407 | } | |
2408 | buddy_loaded = 1; | |
2409 | } | |
2410 | ||
2411 | memcpy(&sg, ext4_get_group_info(sb, group), i); | |
2412 | ||
2413 | if (buddy_loaded) | |
2414 | ext4_mb_unload_buddy(&e4b); | |
2415 | for (i = 0; i < EXT4_FREEFRAG_COLUMN; i++) | |
2416 | freeblock[i] += (i <= sb->s_blocksize_bits + 1) ? | |
2417 | sg.info.bb_counters[i]: 0; | |
2418 | } | |
2419 | out: | |
2420 | for (i=0; i < EXT4_FREEFRAG_COLUMN; i++) | |
2421 | snprintf(buf, PAGE_SIZE, "%s\"%s\":\"%llu\",", buf, size[i], | |
2422 | (unsigned long long)freeblock[i]); | |
2423 | buf[strlen(buf)-1] = '\n'; | |
2424 | ||
2425 | return strlen(buf); | |
2426 | } | |
2427 | ||
fb1813f4 CW |
2428 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2429 | { | |
2430 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2431 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2432 | ||
2433 | BUG_ON(!cachep); | |
2434 | return cachep; | |
2435 | } | |
5f21b0e6 | 2436 | |
28623c2f TT |
2437 | /* |
2438 | * Allocate the top-level s_group_info array for the specified number | |
2439 | * of groups | |
2440 | */ | |
2441 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2442 | { | |
2443 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2444 | unsigned size; | |
2445 | struct ext4_group_info ***new_groupinfo; | |
2446 | ||
2447 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2448 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2449 | if (size <= sbi->s_group_info_size) | |
2450 | return 0; | |
2451 | ||
2452 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
2453 | new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL); | |
2454 | if (!new_groupinfo) { | |
2455 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2456 | return -ENOMEM; | |
2457 | } | |
2458 | if (sbi->s_group_info) { | |
2459 | memcpy(new_groupinfo, sbi->s_group_info, | |
2460 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); | |
b93b41d4 | 2461 | kvfree(sbi->s_group_info); |
28623c2f TT |
2462 | } |
2463 | sbi->s_group_info = new_groupinfo; | |
2464 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); | |
2465 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", | |
2466 | sbi->s_group_info_size); | |
2467 | return 0; | |
2468 | } | |
2469 | ||
5f21b0e6 | 2470 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2471 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2472 | struct ext4_group_desc *desc) |
2473 | { | |
fb1813f4 | 2474 | int i; |
5f21b0e6 FB |
2475 | int metalen = 0; |
2476 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2477 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2478 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2479 | |
2480 | /* | |
2481 | * First check if this group is the first of a reserved block. | |
2482 | * If it's true, we have to allocate a new table of pointers | |
2483 | * to ext4_group_info structures | |
2484 | */ | |
2485 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2486 | metalen = sizeof(*meta_group_info) << | |
2487 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 2488 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 2489 | if (meta_group_info == NULL) { |
7f6a11e7 | 2490 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2491 | "for a buddy group"); |
5f21b0e6 FB |
2492 | goto exit_meta_group_info; |
2493 | } | |
2494 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2495 | meta_group_info; | |
2496 | } | |
2497 | ||
5f21b0e6 FB |
2498 | meta_group_info = |
2499 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2500 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2501 | ||
4fdb5543 | 2502 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 2503 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2504 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2505 | goto exit_group_info; |
2506 | } | |
2507 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2508 | &(meta_group_info[i]->bb_state)); | |
2509 | ||
2510 | /* | |
2511 | * initialize bb_free to be able to skip | |
2512 | * empty groups without initialization | |
2513 | */ | |
1cac41cb MB |
2514 | if (ext4_has_group_desc_csum(sb) && |
2515 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 2516 | meta_group_info[i]->bb_free = |
cff1dfd7 | 2517 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2518 | } else { |
2519 | meta_group_info[i]->bb_free = | |
021b65bb | 2520 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2521 | } |
2522 | ||
2523 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2524 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2525 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2526 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 FB |
2527 | |
2528 | #ifdef DOUBLE_CHECK | |
2529 | { | |
2530 | struct buffer_head *bh; | |
2531 | meta_group_info[i]->bb_bitmap = | |
4fdb5543 | 2532 | kmalloc(sb->s_blocksize, GFP_NOFS); |
5f21b0e6 FB |
2533 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); |
2534 | bh = ext4_read_block_bitmap(sb, group); | |
9008a58e | 2535 | BUG_ON(IS_ERR_OR_NULL(bh)); |
5f21b0e6 FB |
2536 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, |
2537 | sb->s_blocksize); | |
2538 | put_bh(bh); | |
2539 | } | |
2540 | #endif | |
2541 | ||
2542 | return 0; | |
2543 | ||
2544 | exit_group_info: | |
2545 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2546 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
5f21b0e6 | 2547 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); |
caaf7a29 TM |
2548 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; |
2549 | } | |
5f21b0e6 FB |
2550 | exit_meta_group_info: |
2551 | return -ENOMEM; | |
2552 | } /* ext4_mb_add_groupinfo */ | |
2553 | ||
c9de560d AT |
2554 | static int ext4_mb_init_backend(struct super_block *sb) |
2555 | { | |
8df9675f | 2556 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2557 | ext4_group_t i; |
c9de560d | 2558 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2559 | int err; |
5f21b0e6 | 2560 | struct ext4_group_desc *desc; |
fb1813f4 | 2561 | struct kmem_cache *cachep; |
5f21b0e6 | 2562 | |
28623c2f TT |
2563 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2564 | if (err) | |
2565 | return err; | |
c9de560d | 2566 | |
c9de560d AT |
2567 | sbi->s_buddy_cache = new_inode(sb); |
2568 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2569 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2570 | goto err_freesgi; |
2571 | } | |
48e6061b YJ |
2572 | /* To avoid potentially colliding with an valid on-disk inode number, |
2573 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2574 | * not in the inode hash, so it should never be found by iget(), but | |
2575 | * this will avoid confusion if it ever shows up during debugging. */ | |
2576 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2577 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2578 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2579 | desc = ext4_get_group_desc(sb, i, NULL); |
2580 | if (desc == NULL) { | |
9d8b9ec4 | 2581 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2582 | goto err_freebuddy; |
2583 | } | |
5f21b0e6 FB |
2584 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2585 | goto err_freebuddy; | |
c9de560d AT |
2586 | } |
2587 | ||
2588 | return 0; | |
2589 | ||
2590 | err_freebuddy: | |
fb1813f4 | 2591 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2592 | while (i-- > 0) |
fb1813f4 | 2593 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2594 | i = sbi->s_group_info_size; |
f1fa3342 | 2595 | while (i-- > 0) |
c9de560d AT |
2596 | kfree(sbi->s_group_info[i]); |
2597 | iput(sbi->s_buddy_cache); | |
2598 | err_freesgi: | |
b93b41d4 | 2599 | kvfree(sbi->s_group_info); |
c9de560d AT |
2600 | return -ENOMEM; |
2601 | } | |
2602 | ||
2892c15d ES |
2603 | static void ext4_groupinfo_destroy_slabs(void) |
2604 | { | |
2605 | int i; | |
2606 | ||
2607 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
2608 | if (ext4_groupinfo_caches[i]) | |
2609 | kmem_cache_destroy(ext4_groupinfo_caches[i]); | |
2610 | ext4_groupinfo_caches[i] = NULL; | |
2611 | } | |
2612 | } | |
2613 | ||
2614 | static int ext4_groupinfo_create_slab(size_t size) | |
2615 | { | |
2616 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2617 | int slab_size; | |
2618 | int blocksize_bits = order_base_2(size); | |
2619 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2620 | struct kmem_cache *cachep; | |
2621 | ||
2622 | if (cache_index >= NR_GRPINFO_CACHES) | |
2623 | return -EINVAL; | |
2624 | ||
2625 | if (unlikely(cache_index < 0)) | |
2626 | cache_index = 0; | |
2627 | ||
2628 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2629 | if (ext4_groupinfo_caches[cache_index]) { | |
2630 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2631 | return 0; /* Already created */ | |
2632 | } | |
2633 | ||
2634 | slab_size = offsetof(struct ext4_group_info, | |
2635 | bb_counters[blocksize_bits + 2]); | |
2636 | ||
2637 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2638 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2639 | NULL); | |
2640 | ||
823ba01f TM |
2641 | ext4_groupinfo_caches[cache_index] = cachep; |
2642 | ||
2892c15d ES |
2643 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2644 | if (!cachep) { | |
9d8b9ec4 TT |
2645 | printk(KERN_EMERG |
2646 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2647 | return -ENOMEM; |
2648 | } | |
2649 | ||
2892c15d ES |
2650 | return 0; |
2651 | } | |
2652 | ||
9d99012f | 2653 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2654 | { |
2655 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2656 | unsigned i, j; |
8b8de1c9 | 2657 | unsigned offset, offset_incr; |
c9de560d | 2658 | unsigned max; |
74767c5a | 2659 | int ret; |
c9de560d | 2660 | |
1927805e | 2661 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2662 | |
2663 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2664 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2665 | ret = -ENOMEM; |
2666 | goto out; | |
c9de560d | 2667 | } |
ff7ef329 | 2668 | |
1927805e | 2669 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2670 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2671 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2672 | ret = -ENOMEM; |
2673 | goto out; | |
2674 | } | |
2675 | ||
2892c15d ES |
2676 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2677 | if (ret < 0) | |
2678 | goto out; | |
c9de560d AT |
2679 | |
2680 | /* order 0 is regular bitmap */ | |
2681 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2682 | sbi->s_mb_offsets[0] = 0; | |
2683 | ||
2684 | i = 1; | |
2685 | offset = 0; | |
8b8de1c9 | 2686 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
2687 | max = sb->s_blocksize << 2; |
2688 | do { | |
2689 | sbi->s_mb_offsets[i] = offset; | |
2690 | sbi->s_mb_maxs[i] = max; | |
8b8de1c9 NS |
2691 | offset += offset_incr; |
2692 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
2693 | max = max >> 1; |
2694 | i++; | |
2695 | } while (i <= sb->s_blocksize_bits + 1); | |
2696 | ||
c9de560d | 2697 | spin_lock_init(&sbi->s_md_lock); |
c9de560d | 2698 | spin_lock_init(&sbi->s_bal_lock); |
1cac41cb | 2699 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
c9de560d AT |
2700 | |
2701 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2702 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2703 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2704 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2705 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
2706 | /* |
2707 | * The default group preallocation is 512, which for 4k block | |
2708 | * sizes translates to 2 megabytes. However for bigalloc file | |
2709 | * systems, this is probably too big (i.e, if the cluster size | |
2710 | * is 1 megabyte, then group preallocation size becomes half a | |
2711 | * gigabyte!). As a default, we will keep a two megabyte | |
2712 | * group pralloc size for cluster sizes up to 64k, and after | |
2713 | * that, we will force a minimum group preallocation size of | |
2714 | * 32 clusters. This translates to 8 megs when the cluster | |
2715 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2716 | * which seems reasonable as a default. | |
2717 | */ | |
2718 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2719 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2720 | /* |
2721 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2722 | * to the lowest multiple of s_stripe which is bigger than | |
2723 | * the s_mb_group_prealloc as determined above. We want | |
2724 | * the preallocation size to be an exact multiple of the | |
2725 | * RAID stripe size so that preallocations don't fragment | |
2726 | * the stripes. | |
2727 | */ | |
2728 | if (sbi->s_stripe > 1) { | |
2729 | sbi->s_mb_group_prealloc = roundup( | |
2730 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2731 | } | |
c9de560d | 2732 | |
730c213c | 2733 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2734 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2735 | ret = -ENOMEM; |
029b10c5 | 2736 | goto out; |
c9de560d | 2737 | } |
730c213c | 2738 | for_each_possible_cpu(i) { |
c9de560d | 2739 | struct ext4_locality_group *lg; |
730c213c | 2740 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2741 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2742 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2743 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2744 | spin_lock_init(&lg->lg_prealloc_lock); |
2745 | } | |
2746 | ||
79a77c5a YJ |
2747 | /* init file for buddy data */ |
2748 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2749 | if (ret != 0) |
2750 | goto out_free_locality_groups; | |
79a77c5a | 2751 | |
7aa0baea TM |
2752 | return 0; |
2753 | ||
2754 | out_free_locality_groups: | |
2755 | free_percpu(sbi->s_locality_groups); | |
2756 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 2757 | out: |
7aa0baea TM |
2758 | kfree(sbi->s_mb_offsets); |
2759 | sbi->s_mb_offsets = NULL; | |
2760 | kfree(sbi->s_mb_maxs); | |
2761 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2762 | return ret; |
c9de560d AT |
2763 | } |
2764 | ||
955ce5f5 | 2765 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2766 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2767 | { | |
2768 | struct ext4_prealloc_space *pa; | |
2769 | struct list_head *cur, *tmp; | |
2770 | int count = 0; | |
2771 | ||
2772 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2773 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2774 | list_del(&pa->pa_group_list); | |
2775 | count++; | |
688f05a0 | 2776 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2777 | } |
2778 | if (count) | |
6ba495e9 | 2779 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2780 | |
2781 | } | |
2782 | ||
2783 | int ext4_mb_release(struct super_block *sb) | |
2784 | { | |
8df9675f | 2785 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2786 | ext4_group_t i; |
2787 | int num_meta_group_infos; | |
2788 | struct ext4_group_info *grinfo; | |
2789 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
fb1813f4 | 2790 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2791 | |
c9de560d | 2792 | if (sbi->s_group_info) { |
8df9675f | 2793 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2794 | grinfo = ext4_get_group_info(sb, i); |
2795 | #ifdef DOUBLE_CHECK | |
2796 | kfree(grinfo->bb_bitmap); | |
2797 | #endif | |
2798 | ext4_lock_group(sb, i); | |
2799 | ext4_mb_cleanup_pa(grinfo); | |
2800 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2801 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2802 | } |
8df9675f | 2803 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2804 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2805 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2806 | for (i = 0; i < num_meta_group_infos; i++) | |
2807 | kfree(sbi->s_group_info[i]); | |
b93b41d4 | 2808 | kvfree(sbi->s_group_info); |
c9de560d AT |
2809 | } |
2810 | kfree(sbi->s_mb_offsets); | |
2811 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 2812 | iput(sbi->s_buddy_cache); |
c9de560d | 2813 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
2814 | ext4_msg(sb, KERN_INFO, |
2815 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2816 | atomic_read(&sbi->s_bal_allocated), |
2817 | atomic_read(&sbi->s_bal_reqs), | |
2818 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2819 | ext4_msg(sb, KERN_INFO, |
2820 | "mballoc: %u extents scanned, %u goal hits, " | |
2821 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2822 | atomic_read(&sbi->s_bal_ex_scanned), |
2823 | atomic_read(&sbi->s_bal_goals), | |
2824 | atomic_read(&sbi->s_bal_2orders), | |
2825 | atomic_read(&sbi->s_bal_breaks), | |
2826 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2827 | ext4_msg(sb, KERN_INFO, |
2828 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2829 | sbi->s_mb_buddies_generated, |
c9de560d | 2830 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2831 | ext4_msg(sb, KERN_INFO, |
2832 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2833 | atomic_read(&sbi->s_mb_preallocated), |
2834 | atomic_read(&sbi->s_mb_discarded)); | |
2835 | } | |
2836 | ||
730c213c | 2837 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2838 | |
2839 | return 0; | |
2840 | } | |
2841 | ||
77ca6cdf | 2842 | static inline int ext4_issue_discard(struct super_block *sb, |
9a878d1d | 2843 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
1cac41cb | 2844 | unsigned long flags, struct bio **biop) |
5c521830 | 2845 | { |
5c521830 JZ |
2846 | ext4_fsblk_t discard_block; |
2847 | ||
84130193 TT |
2848 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
2849 | ext4_group_first_block_no(sb, block_group)); | |
2850 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
2851 | trace_ext4_discard_blocks(sb, |
2852 | (unsigned long long) discard_block, count); | |
1cac41cb MB |
2853 | if (biop) { |
2854 | return __blkdev_issue_discard(sb->s_bdev, | |
2855 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
2856 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
2857 | GFP_NOFS, flags, biop); | |
2858 | } else | |
2859 | return sb_issue_discard(sb, discard_block, count, | |
2860 | GFP_NOFS, flags); | |
5c521830 JZ |
2861 | } |
2862 | ||
1cac41cb MB |
2863 | static void ext4_free_data_in_buddy(struct super_block *sb, |
2864 | struct ext4_free_data *entry) | |
c9de560d | 2865 | { |
c9de560d | 2866 | struct ext4_buddy e4b; |
c894058d | 2867 | struct ext4_group_info *db; |
d9f34504 | 2868 | int err, count = 0, count2 = 0; |
c9de560d | 2869 | |
18aadd47 BJ |
2870 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
2871 | entry->efd_count, entry->efd_group, entry); | |
c9de560d | 2872 | |
18aadd47 BJ |
2873 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
2874 | /* we expect to find existing buddy because it's pinned */ | |
2875 | BUG_ON(err != 0); | |
b90f6870 | 2876 | |
18aadd47 BJ |
2877 | db = e4b.bd_info; |
2878 | /* there are blocks to put in buddy to make them really free */ | |
2879 | count += entry->efd_count; | |
2880 | count2++; | |
2881 | ext4_lock_group(sb, entry->efd_group); | |
2882 | /* Take it out of per group rb tree */ | |
2883 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
2884 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 2885 | |
18aadd47 BJ |
2886 | /* |
2887 | * Clear the trimmed flag for the group so that the next | |
2888 | * ext4_trim_fs can trim it. | |
2889 | * If the volume is mounted with -o discard, online discard | |
2890 | * is supported and the free blocks will be trimmed online. | |
2891 | */ | |
2892 | if (!test_opt(sb, DISCARD)) | |
2893 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 2894 | |
18aadd47 BJ |
2895 | if (!db->bb_free_root.rb_node) { |
2896 | /* No more items in the per group rb tree | |
2897 | * balance refcounts from ext4_mb_free_metadata() | |
2898 | */ | |
2899 | page_cache_release(e4b.bd_buddy_page); | |
2900 | page_cache_release(e4b.bd_bitmap_page); | |
3e624fc7 | 2901 | } |
18aadd47 BJ |
2902 | ext4_unlock_group(sb, entry->efd_group); |
2903 | kmem_cache_free(ext4_free_data_cachep, entry); | |
2904 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2905 | |
6ba495e9 | 2906 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2907 | } |
2908 | ||
1cac41cb MB |
2909 | /* |
2910 | * This function is called by the jbd2 layer once the commit has finished, | |
2911 | * so we know we can free the blocks that were released with that commit. | |
2912 | */ | |
2913 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
2914 | { | |
2915 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2916 | struct ext4_free_data *entry, *tmp; | |
2917 | struct bio *discard_bio = NULL; | |
2918 | struct list_head freed_data_list; | |
2919 | struct list_head *cut_pos = NULL; | |
2920 | int type = REQ_WRITE | REQ_DISCARD | REQ_PRIO; | |
2921 | int err; | |
2922 | ||
2923 | INIT_LIST_HEAD(&freed_data_list); | |
2924 | ||
2925 | spin_lock(&sbi->s_md_lock); | |
2926 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
2927 | if (entry->efd_tid != commit_tid) | |
2928 | break; | |
2929 | cut_pos = &entry->efd_list; | |
2930 | } | |
2931 | if (cut_pos) | |
2932 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
2933 | cut_pos); | |
2934 | spin_unlock(&sbi->s_md_lock); | |
2935 | ||
2936 | if (test_opt(sb, DISCARD)) { | |
2937 | list_for_each_entry(entry, &freed_data_list, efd_list) { | |
2938 | err = ext4_issue_discard(sb, entry->efd_group, | |
2939 | entry->efd_start_cluster, | |
2940 | entry->efd_count, | |
2941 | BLKDEV_DISCARD_SYNC, | |
2942 | &discard_bio); | |
2943 | if (err && err != -EOPNOTSUPP) { | |
2944 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
2945 | " group:%d block:%d count:%d failed" | |
2946 | " with %d", entry->efd_group, | |
2947 | entry->efd_start_cluster, | |
2948 | entry->efd_count, err); | |
2949 | } else if (err == -EOPNOTSUPP) | |
2950 | break; | |
2951 | } | |
2952 | ||
2953 | if (discard_bio) { | |
2954 | submit_bio_wait(type, discard_bio); | |
2955 | bio_put(discard_bio); | |
2956 | } | |
2957 | } | |
2958 | ||
2959 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
2960 | ext4_free_data_in_buddy(sb, entry); | |
2961 | } | |
2962 | ||
5dabfc78 | 2963 | int __init ext4_init_mballoc(void) |
c9de560d | 2964 | { |
16828088 TT |
2965 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2966 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d AT |
2967 | if (ext4_pspace_cachep == NULL) |
2968 | return -ENOMEM; | |
2969 | ||
16828088 TT |
2970 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2971 | SLAB_RECLAIM_ACCOUNT); | |
256bdb49 ES |
2972 | if (ext4_ac_cachep == NULL) { |
2973 | kmem_cache_destroy(ext4_pspace_cachep); | |
2974 | return -ENOMEM; | |
2975 | } | |
c894058d | 2976 | |
18aadd47 BJ |
2977 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
2978 | SLAB_RECLAIM_ACCOUNT); | |
2979 | if (ext4_free_data_cachep == NULL) { | |
c894058d AK |
2980 | kmem_cache_destroy(ext4_pspace_cachep); |
2981 | kmem_cache_destroy(ext4_ac_cachep); | |
2982 | return -ENOMEM; | |
2983 | } | |
c9de560d AT |
2984 | return 0; |
2985 | } | |
2986 | ||
5dabfc78 | 2987 | void ext4_exit_mballoc(void) |
c9de560d | 2988 | { |
60e6679e | 2989 | /* |
3e03f9ca JDB |
2990 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2991 | * before destroying the slab cache. | |
2992 | */ | |
2993 | rcu_barrier(); | |
c9de560d | 2994 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2995 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 2996 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 2997 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
2998 | } |
2999 | ||
3000 | ||
3001 | /* | |
73b2c716 | 3002 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
3003 | * Returns 0 if success or error code |
3004 | */ | |
4ddfef7b ES |
3005 | static noinline_for_stack int |
3006 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 3007 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
3008 | { |
3009 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
3010 | struct ext4_group_desc *gdp; |
3011 | struct buffer_head *gdp_bh; | |
3012 | struct ext4_sb_info *sbi; | |
3013 | struct super_block *sb; | |
3014 | ext4_fsblk_t block; | |
519deca0 | 3015 | int err, len; |
c9de560d AT |
3016 | |
3017 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3018 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
3019 | ||
3020 | sb = ac->ac_sb; | |
3021 | sbi = EXT4_SB(sb); | |
c9de560d | 3022 | |
574ca174 | 3023 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
3024 | if (IS_ERR(bitmap_bh)) { |
3025 | err = PTR_ERR(bitmap_bh); | |
3026 | bitmap_bh = NULL; | |
c9de560d | 3027 | goto out_err; |
9008a58e | 3028 | } |
c9de560d | 3029 | |
5d601255 | 3030 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
c9de560d AT |
3031 | err = ext4_journal_get_write_access(handle, bitmap_bh); |
3032 | if (err) | |
3033 | goto out_err; | |
3034 | ||
3035 | err = -EIO; | |
3036 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3037 | if (!gdp) | |
3038 | goto out_err; | |
3039 | ||
a9df9a49 | 3040 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3041 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3042 | |
5d601255 | 3043 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
c9de560d AT |
3044 | err = ext4_journal_get_write_access(handle, gdp_bh); |
3045 | if (err) | |
3046 | goto out_err; | |
3047 | ||
bda00de7 | 3048 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3049 | |
53accfa9 | 3050 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 3051 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 3052 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3053 | "fs metadata", block, block+len); |
519deca0 | 3054 | /* File system mounted not to panic on error |
3a22cf0c | 3055 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3056 | * We leak some of the blocks here. |
3057 | */ | |
955ce5f5 | 3058 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
3059 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3060 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 3061 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3062 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3063 | if (!err) |
3a22cf0c | 3064 | err = -EFSCORRUPTED; |
519deca0 | 3065 | goto out_err; |
c9de560d | 3066 | } |
955ce5f5 AK |
3067 | |
3068 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3069 | #ifdef AGGRESSIVE_CHECK |
3070 | { | |
3071 | int i; | |
3072 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3073 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3074 | bitmap_bh->b_data)); | |
3075 | } | |
3076 | } | |
3077 | #endif | |
c3e94d1d YY |
3078 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3079 | ac->ac_b_ex.fe_len); | |
1cac41cb MB |
3080 | if (ext4_has_group_desc_csum(sb) && |
3081 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3082 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3083 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3084 | ext4_free_clusters_after_init(sb, |
021b65bb | 3085 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3086 | } |
021b65bb TT |
3087 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3088 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3089 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3090 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3091 | |
3092 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3093 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3094 | /* |
6bc6e63f | 3095 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3096 | */ |
6bc6e63f AK |
3097 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3098 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3099 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3100 | reserv_clstrs); | |
c9de560d | 3101 | |
772cb7c8 JS |
3102 | if (sbi->s_log_groups_per_flex) { |
3103 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3104 | ac->ac_b_ex.fe_group); | |
90ba983f TT |
3105 | atomic64_sub(ac->ac_b_ex.fe_len, |
3106 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
3107 | } |
3108 | ||
0390131b | 3109 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3110 | if (err) |
3111 | goto out_err; | |
0390131b | 3112 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3113 | |
3114 | out_err: | |
42a10add | 3115 | brelse(bitmap_bh); |
c9de560d AT |
3116 | return err; |
3117 | } | |
3118 | ||
3119 | /* | |
3120 | * here we normalize request for locality group | |
d7a1fee1 DE |
3121 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3122 | * s_strip if we set the same via mount option. | |
3123 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3124 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3125 | * |
3126 | * XXX: should we try to preallocate more than the group has now? | |
3127 | */ | |
3128 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3129 | { | |
3130 | struct super_block *sb = ac->ac_sb; | |
3131 | struct ext4_locality_group *lg = ac->ac_lg; | |
3132 | ||
3133 | BUG_ON(lg == NULL); | |
d7a1fee1 | 3134 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 3135 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
3136 | current->pid, ac->ac_g_ex.fe_len); |
3137 | } | |
3138 | ||
3139 | /* | |
3140 | * Normalization means making request better in terms of | |
3141 | * size and alignment | |
3142 | */ | |
4ddfef7b ES |
3143 | static noinline_for_stack void |
3144 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3145 | struct ext4_allocation_request *ar) |
3146 | { | |
53accfa9 | 3147 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3148 | int bsbits, max; |
3149 | ext4_lblk_t end; | |
1592d2c5 CW |
3150 | loff_t size, start_off; |
3151 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3152 | ext4_lblk_t start; |
c9de560d | 3153 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 3154 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3155 | |
3156 | /* do normalize only data requests, metadata requests | |
3157 | do not need preallocation */ | |
3158 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3159 | return; | |
3160 | ||
3161 | /* sometime caller may want exact blocks */ | |
3162 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3163 | return; | |
3164 | ||
3165 | /* caller may indicate that preallocation isn't | |
3166 | * required (it's a tail, for example) */ | |
3167 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3168 | return; | |
3169 | ||
3170 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3171 | ext4_mb_normalize_group_request(ac); | |
3172 | return ; | |
3173 | } | |
3174 | ||
3175 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3176 | ||
3177 | /* first, let's learn actual file size | |
3178 | * given current request is allocated */ | |
53accfa9 | 3179 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
3180 | size = size << bsbits; |
3181 | if (size < i_size_read(ac->ac_inode)) | |
3182 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 3183 | orig_size = size; |
c9de560d | 3184 | |
1930479c VC |
3185 | /* max size of free chunks */ |
3186 | max = 2 << bsbits; | |
c9de560d | 3187 | |
1930479c VC |
3188 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3189 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3190 | |
3191 | /* first, try to predict filesize */ | |
3192 | /* XXX: should this table be tunable? */ | |
3193 | start_off = 0; | |
3194 | if (size <= 16 * 1024) { | |
3195 | size = 16 * 1024; | |
3196 | } else if (size <= 32 * 1024) { | |
3197 | size = 32 * 1024; | |
3198 | } else if (size <= 64 * 1024) { | |
3199 | size = 64 * 1024; | |
3200 | } else if (size <= 128 * 1024) { | |
3201 | size = 128 * 1024; | |
3202 | } else if (size <= 256 * 1024) { | |
3203 | size = 256 * 1024; | |
3204 | } else if (size <= 512 * 1024) { | |
3205 | size = 512 * 1024; | |
3206 | } else if (size <= 1024 * 1024) { | |
3207 | size = 1024 * 1024; | |
1930479c | 3208 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3209 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3210 | (21 - bsbits)) << 21; |
3211 | size = 2 * 1024 * 1024; | |
3212 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3213 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3214 | (22 - bsbits)) << 22; | |
3215 | size = 4 * 1024 * 1024; | |
3216 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3217 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3218 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3219 | (23 - bsbits)) << 23; | |
3220 | size = 8 * 1024 * 1024; | |
3221 | } else { | |
b27b1535 XW |
3222 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
3223 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
3224 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 3225 | } |
5a0790c2 AK |
3226 | size = size >> bsbits; |
3227 | start = start_off >> bsbits; | |
c9de560d AT |
3228 | |
3229 | /* don't cover already allocated blocks in selected range */ | |
3230 | if (ar->pleft && start <= ar->lleft) { | |
3231 | size -= ar->lleft + 1 - start; | |
3232 | start = ar->lleft + 1; | |
3233 | } | |
3234 | if (ar->pright && start + size - 1 >= ar->lright) | |
3235 | size -= start + size - ar->lright; | |
3236 | ||
8774c73c JK |
3237 | /* |
3238 | * Trim allocation request for filesystems with artificially small | |
3239 | * groups. | |
3240 | */ | |
3241 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
3242 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
3243 | ||
c9de560d AT |
3244 | end = start + size; |
3245 | ||
3246 | /* check we don't cross already preallocated blocks */ | |
3247 | rcu_read_lock(); | |
9a0762c5 | 3248 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3249 | ext4_lblk_t pa_end; |
c9de560d | 3250 | |
c9de560d AT |
3251 | if (pa->pa_deleted) |
3252 | continue; | |
3253 | spin_lock(&pa->pa_lock); | |
3254 | if (pa->pa_deleted) { | |
3255 | spin_unlock(&pa->pa_lock); | |
3256 | continue; | |
3257 | } | |
3258 | ||
53accfa9 TT |
3259 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3260 | pa->pa_len); | |
c9de560d AT |
3261 | |
3262 | /* PA must not overlap original request */ | |
3263 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3264 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3265 | ||
38877f4e ES |
3266 | /* skip PAs this normalized request doesn't overlap with */ |
3267 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3268 | spin_unlock(&pa->pa_lock); |
3269 | continue; | |
3270 | } | |
3271 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3272 | ||
38877f4e | 3273 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3274 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3275 | BUG_ON(pa_end < start); | |
3276 | start = pa_end; | |
38877f4e | 3277 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3278 | BUG_ON(pa->pa_lstart > end); |
3279 | end = pa->pa_lstart; | |
3280 | } | |
3281 | spin_unlock(&pa->pa_lock); | |
3282 | } | |
3283 | rcu_read_unlock(); | |
3284 | size = end - start; | |
3285 | ||
3286 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3287 | rcu_read_lock(); | |
9a0762c5 | 3288 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3289 | ext4_lblk_t pa_end; |
53accfa9 | 3290 | |
c9de560d AT |
3291 | spin_lock(&pa->pa_lock); |
3292 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3293 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3294 | pa->pa_len); | |
c9de560d AT |
3295 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3296 | } | |
3297 | spin_unlock(&pa->pa_lock); | |
3298 | } | |
3299 | rcu_read_unlock(); | |
3300 | ||
3301 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3302 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3303 | ext4_msg(ac->ac_sb, KERN_ERR, |
3304 | "start %lu, size %lu, fe_logical %lu", | |
3305 | (unsigned long) start, (unsigned long) size, | |
3306 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 3307 | BUG(); |
c9de560d | 3308 | } |
b5b60778 | 3309 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3310 | |
3311 | /* now prepare goal request */ | |
3312 | ||
3313 | /* XXX: is it better to align blocks WRT to logical | |
3314 | * placement or satisfy big request as is */ | |
3315 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3316 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3317 | |
3318 | /* define goal start in order to merge */ | |
3319 | if (ar->pright && (ar->lright == (start + size))) { | |
3320 | /* merge to the right */ | |
3321 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3322 | &ac->ac_f_ex.fe_group, | |
3323 | &ac->ac_f_ex.fe_start); | |
3324 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3325 | } | |
3326 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3327 | /* merge to the left */ | |
3328 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3329 | &ac->ac_f_ex.fe_group, | |
3330 | &ac->ac_f_ex.fe_start); | |
3331 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3332 | } | |
3333 | ||
6ba495e9 | 3334 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3335 | (unsigned) orig_size, (unsigned) start); |
3336 | } | |
3337 | ||
3338 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3339 | { | |
3340 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3341 | ||
3342 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3343 | atomic_inc(&sbi->s_bal_reqs); | |
3344 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3345 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3346 | atomic_inc(&sbi->s_bal_success); |
3347 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3348 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3349 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3350 | atomic_inc(&sbi->s_bal_goals); | |
3351 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3352 | atomic_inc(&sbi->s_bal_breaks); | |
3353 | } | |
3354 | ||
296c355c TT |
3355 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3356 | trace_ext4_mballoc_alloc(ac); | |
3357 | else | |
3358 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3359 | } |
3360 | ||
b844167e CW |
3361 | /* |
3362 | * Called on failure; free up any blocks from the inode PA for this | |
3363 | * context. We don't need this for MB_GROUP_PA because we only change | |
3364 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3365 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3366 | */ | |
3367 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3368 | { | |
3369 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
3370 | struct ext4_buddy e4b; |
3371 | int err; | |
b844167e | 3372 | |
86f0afd4 | 3373 | if (pa == NULL) { |
c99d1e6e TT |
3374 | if (ac->ac_f_ex.fe_len == 0) |
3375 | return; | |
86f0afd4 TT |
3376 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
3377 | if (err) { | |
3378 | /* | |
3379 | * This should never happen since we pin the | |
3380 | * pages in the ext4_allocation_context so | |
3381 | * ext4_mb_load_buddy() should never fail. | |
3382 | */ | |
3383 | WARN(1, "mb_load_buddy failed (%d)", err); | |
3384 | return; | |
3385 | } | |
3386 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
3387 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
3388 | ac->ac_f_ex.fe_len); | |
3389 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 3390 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
3391 | return; |
3392 | } | |
3393 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 3394 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
3395 | } |
3396 | ||
c9de560d AT |
3397 | /* |
3398 | * use blocks preallocated to inode | |
3399 | */ | |
3400 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3401 | struct ext4_prealloc_space *pa) | |
3402 | { | |
53accfa9 | 3403 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3404 | ext4_fsblk_t start; |
3405 | ext4_fsblk_t end; | |
3406 | int len; | |
3407 | ||
3408 | /* found preallocated blocks, use them */ | |
3409 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3410 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3411 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3412 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3413 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3414 | &ac->ac_b_ex.fe_start); | |
3415 | ac->ac_b_ex.fe_len = len; | |
3416 | ac->ac_status = AC_STATUS_FOUND; | |
3417 | ac->ac_pa = pa; | |
3418 | ||
3419 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3420 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3421 | BUG_ON(pa->pa_free < len); |
3422 | pa->pa_free -= len; | |
3423 | ||
6ba495e9 | 3424 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3425 | } |
3426 | ||
3427 | /* | |
3428 | * use blocks preallocated to locality group | |
3429 | */ | |
3430 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3431 | struct ext4_prealloc_space *pa) | |
3432 | { | |
03cddb80 | 3433 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3434 | |
c9de560d AT |
3435 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3436 | &ac->ac_b_ex.fe_group, | |
3437 | &ac->ac_b_ex.fe_start); | |
3438 | ac->ac_b_ex.fe_len = len; | |
3439 | ac->ac_status = AC_STATUS_FOUND; | |
3440 | ac->ac_pa = pa; | |
3441 | ||
3442 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3443 | * possible race when the group is being loaded concurrently |
c9de560d | 3444 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3445 | * in on-disk bitmap -- see ext4_mb_release_context() |
3446 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3447 | */ |
6ba495e9 | 3448 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3449 | } |
3450 | ||
5e745b04 AK |
3451 | /* |
3452 | * Return the prealloc space that have minimal distance | |
3453 | * from the goal block. @cpa is the prealloc | |
3454 | * space that is having currently known minimal distance | |
3455 | * from the goal block. | |
3456 | */ | |
3457 | static struct ext4_prealloc_space * | |
3458 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3459 | struct ext4_prealloc_space *pa, | |
3460 | struct ext4_prealloc_space *cpa) | |
3461 | { | |
3462 | ext4_fsblk_t cur_distance, new_distance; | |
3463 | ||
3464 | if (cpa == NULL) { | |
3465 | atomic_inc(&pa->pa_count); | |
3466 | return pa; | |
3467 | } | |
79211c8e AM |
3468 | cur_distance = abs(goal_block - cpa->pa_pstart); |
3469 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 3470 | |
5a54b2f1 | 3471 | if (cur_distance <= new_distance) |
5e745b04 AK |
3472 | return cpa; |
3473 | ||
3474 | /* drop the previous reference */ | |
3475 | atomic_dec(&cpa->pa_count); | |
3476 | atomic_inc(&pa->pa_count); | |
3477 | return pa; | |
3478 | } | |
3479 | ||
c9de560d AT |
3480 | /* |
3481 | * search goal blocks in preallocated space | |
3482 | */ | |
4ddfef7b ES |
3483 | static noinline_for_stack int |
3484 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3485 | { |
53accfa9 | 3486 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3487 | int order, i; |
c9de560d AT |
3488 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3489 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3490 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3491 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3492 | |
3493 | /* only data can be preallocated */ | |
3494 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3495 | return 0; | |
3496 | ||
3497 | /* first, try per-file preallocation */ | |
3498 | rcu_read_lock(); | |
9a0762c5 | 3499 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3500 | |
3501 | /* all fields in this condition don't change, | |
3502 | * so we can skip locking for them */ | |
3503 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3504 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3505 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3506 | continue; |
3507 | ||
fb0a387d | 3508 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3509 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3510 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3511 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3512 | continue; |
3513 | ||
c9de560d AT |
3514 | /* found preallocated blocks, use them */ |
3515 | spin_lock(&pa->pa_lock); | |
3516 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3517 | atomic_inc(&pa->pa_count); | |
3518 | ext4_mb_use_inode_pa(ac, pa); | |
3519 | spin_unlock(&pa->pa_lock); | |
3520 | ac->ac_criteria = 10; | |
3521 | rcu_read_unlock(); | |
3522 | return 1; | |
3523 | } | |
3524 | spin_unlock(&pa->pa_lock); | |
3525 | } | |
3526 | rcu_read_unlock(); | |
3527 | ||
3528 | /* can we use group allocation? */ | |
3529 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3530 | return 0; | |
3531 | ||
3532 | /* inode may have no locality group for some reason */ | |
3533 | lg = ac->ac_lg; | |
3534 | if (lg == NULL) | |
3535 | return 0; | |
6be2ded1 AK |
3536 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3537 | if (order > PREALLOC_TB_SIZE - 1) | |
3538 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3539 | order = PREALLOC_TB_SIZE - 1; | |
3540 | ||
bda00de7 | 3541 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3542 | /* |
3543 | * search for the prealloc space that is having | |
3544 | * minimal distance from the goal block. | |
3545 | */ | |
6be2ded1 AK |
3546 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3547 | rcu_read_lock(); | |
3548 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3549 | pa_inode_list) { | |
3550 | spin_lock(&pa->pa_lock); | |
3551 | if (pa->pa_deleted == 0 && | |
3552 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3553 | |
3554 | cpa = ext4_mb_check_group_pa(goal_block, | |
3555 | pa, cpa); | |
6be2ded1 | 3556 | } |
c9de560d | 3557 | spin_unlock(&pa->pa_lock); |
c9de560d | 3558 | } |
6be2ded1 | 3559 | rcu_read_unlock(); |
c9de560d | 3560 | } |
5e745b04 AK |
3561 | if (cpa) { |
3562 | ext4_mb_use_group_pa(ac, cpa); | |
3563 | ac->ac_criteria = 20; | |
3564 | return 1; | |
3565 | } | |
c9de560d AT |
3566 | return 0; |
3567 | } | |
3568 | ||
7a2fcbf7 AK |
3569 | /* |
3570 | * the function goes through all block freed in the group | |
3571 | * but not yet committed and marks them used in in-core bitmap. | |
3572 | * buddy must be generated from this bitmap | |
955ce5f5 | 3573 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3574 | */ |
3575 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3576 | ext4_group_t group) | |
3577 | { | |
3578 | struct rb_node *n; | |
3579 | struct ext4_group_info *grp; | |
3580 | struct ext4_free_data *entry; | |
3581 | ||
3582 | grp = ext4_get_group_info(sb, group); | |
3583 | n = rb_first(&(grp->bb_free_root)); | |
3584 | ||
3585 | while (n) { | |
18aadd47 BJ |
3586 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3587 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3588 | n = rb_next(n); |
3589 | } | |
3590 | return; | |
3591 | } | |
3592 | ||
c9de560d AT |
3593 | /* |
3594 | * the function goes through all preallocation in this group and marks them | |
3595 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3596 | * Need to be called with ext4 group lock held |
c9de560d | 3597 | */ |
089ceecc ES |
3598 | static noinline_for_stack |
3599 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3600 | ext4_group_t group) |
3601 | { | |
3602 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3603 | struct ext4_prealloc_space *pa; | |
3604 | struct list_head *cur; | |
3605 | ext4_group_t groupnr; | |
3606 | ext4_grpblk_t start; | |
3607 | int preallocated = 0; | |
c9de560d AT |
3608 | int len; |
3609 | ||
3610 | /* all form of preallocation discards first load group, | |
3611 | * so the only competing code is preallocation use. | |
3612 | * we don't need any locking here | |
3613 | * notice we do NOT ignore preallocations with pa_deleted | |
3614 | * otherwise we could leave used blocks available for | |
3615 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3616 | * is dropping preallocation | |
3617 | */ | |
3618 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3619 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3620 | spin_lock(&pa->pa_lock); | |
3621 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3622 | &groupnr, &start); | |
3623 | len = pa->pa_len; | |
3624 | spin_unlock(&pa->pa_lock); | |
3625 | if (unlikely(len == 0)) | |
3626 | continue; | |
3627 | BUG_ON(groupnr != group); | |
c3e94d1d | 3628 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3629 | preallocated += len; |
c9de560d | 3630 | } |
6ba495e9 | 3631 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3632 | } |
3633 | ||
3634 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3635 | { | |
3636 | struct ext4_prealloc_space *pa; | |
3637 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
3638 | |
3639 | BUG_ON(atomic_read(&pa->pa_count)); | |
3640 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
3641 | kmem_cache_free(ext4_pspace_cachep, pa); |
3642 | } | |
3643 | ||
3644 | /* | |
3645 | * drops a reference to preallocated space descriptor | |
3646 | * if this was the last reference and the space is consumed | |
3647 | */ | |
3648 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3649 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3650 | { | |
a9df9a49 | 3651 | ext4_group_t grp; |
d33a1976 | 3652 | ext4_fsblk_t grp_blk; |
c9de560d | 3653 | |
c9de560d AT |
3654 | /* in this short window concurrent discard can set pa_deleted */ |
3655 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
3656 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
3657 | spin_unlock(&pa->pa_lock); | |
3658 | return; | |
3659 | } | |
3660 | ||
c9de560d AT |
3661 | if (pa->pa_deleted == 1) { |
3662 | spin_unlock(&pa->pa_lock); | |
3663 | return; | |
3664 | } | |
3665 | ||
3666 | pa->pa_deleted = 1; | |
3667 | spin_unlock(&pa->pa_lock); | |
3668 | ||
d33a1976 | 3669 | grp_blk = pa->pa_pstart; |
60e6679e | 3670 | /* |
cc0fb9ad AK |
3671 | * If doing group-based preallocation, pa_pstart may be in the |
3672 | * next group when pa is used up | |
3673 | */ | |
3674 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3675 | grp_blk--; |
3676 | ||
bd86298e | 3677 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
3678 | |
3679 | /* | |
3680 | * possible race: | |
3681 | * | |
3682 | * P1 (buddy init) P2 (regular allocation) | |
3683 | * find block B in PA | |
3684 | * copy on-disk bitmap to buddy | |
3685 | * mark B in on-disk bitmap | |
3686 | * drop PA from group | |
3687 | * mark all PAs in buddy | |
3688 | * | |
3689 | * thus, P1 initializes buddy with B available. to prevent this | |
3690 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3691 | * against that pair | |
3692 | */ | |
3693 | ext4_lock_group(sb, grp); | |
3694 | list_del(&pa->pa_group_list); | |
3695 | ext4_unlock_group(sb, grp); | |
3696 | ||
3697 | spin_lock(pa->pa_obj_lock); | |
3698 | list_del_rcu(&pa->pa_inode_list); | |
3699 | spin_unlock(pa->pa_obj_lock); | |
3700 | ||
3701 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3702 | } | |
3703 | ||
3704 | /* | |
3705 | * creates new preallocated space for given inode | |
3706 | */ | |
4ddfef7b ES |
3707 | static noinline_for_stack int |
3708 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3709 | { |
3710 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3711 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3712 | struct ext4_prealloc_space *pa; |
3713 | struct ext4_group_info *grp; | |
3714 | struct ext4_inode_info *ei; | |
3715 | ||
3716 | /* preallocate only when found space is larger then requested */ | |
3717 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3718 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3719 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3720 | ||
3721 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3722 | if (pa == NULL) | |
3723 | return -ENOMEM; | |
3724 | ||
3725 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3726 | int winl; | |
3727 | int wins; | |
3728 | int win; | |
3729 | int offs; | |
3730 | ||
3731 | /* we can't allocate as much as normalizer wants. | |
3732 | * so, found space must get proper lstart | |
3733 | * to cover original request */ | |
3734 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3735 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3736 | ||
3737 | /* we're limited by original request in that | |
3738 | * logical block must be covered any way | |
3739 | * winl is window we can move our chunk within */ | |
3740 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3741 | ||
3742 | /* also, we should cover whole original request */ | |
53accfa9 | 3743 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3744 | |
3745 | /* the smallest one defines real window */ | |
3746 | win = min(winl, wins); | |
3747 | ||
53accfa9 TT |
3748 | offs = ac->ac_o_ex.fe_logical % |
3749 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3750 | if (offs && offs < win) |
3751 | win = offs; | |
3752 | ||
53accfa9 | 3753 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3754 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3755 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3756 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3757 | } | |
3758 | ||
3759 | /* preallocation can change ac_b_ex, thus we store actually | |
3760 | * allocated blocks for history */ | |
3761 | ac->ac_f_ex = ac->ac_b_ex; | |
3762 | ||
3763 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3764 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3765 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3766 | pa->pa_free = pa->pa_len; | |
3767 | atomic_set(&pa->pa_count, 1); | |
3768 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3769 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3770 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3771 | pa->pa_deleted = 0; |
cc0fb9ad | 3772 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3773 | |
6ba495e9 | 3774 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3775 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3776 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3777 | |
3778 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3779 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3780 | |
3781 | ei = EXT4_I(ac->ac_inode); | |
3782 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3783 | ||
3784 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3785 | pa->pa_inode = ac->ac_inode; | |
3786 | ||
3787 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3788 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3789 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3790 | ||
3791 | spin_lock(pa->pa_obj_lock); | |
3792 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3793 | spin_unlock(pa->pa_obj_lock); | |
3794 | ||
3795 | return 0; | |
3796 | } | |
3797 | ||
3798 | /* | |
3799 | * creates new preallocated space for locality group inodes belongs to | |
3800 | */ | |
4ddfef7b ES |
3801 | static noinline_for_stack int |
3802 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3803 | { |
3804 | struct super_block *sb = ac->ac_sb; | |
3805 | struct ext4_locality_group *lg; | |
3806 | struct ext4_prealloc_space *pa; | |
3807 | struct ext4_group_info *grp; | |
3808 | ||
3809 | /* preallocate only when found space is larger then requested */ | |
3810 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3811 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3812 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3813 | ||
3814 | BUG_ON(ext4_pspace_cachep == NULL); | |
3815 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3816 | if (pa == NULL) | |
3817 | return -ENOMEM; | |
3818 | ||
3819 | /* preallocation can change ac_b_ex, thus we store actually | |
3820 | * allocated blocks for history */ | |
3821 | ac->ac_f_ex = ac->ac_b_ex; | |
3822 | ||
3823 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3824 | pa->pa_lstart = pa->pa_pstart; | |
3825 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3826 | pa->pa_free = pa->pa_len; | |
3827 | atomic_set(&pa->pa_count, 1); | |
3828 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3829 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3830 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3831 | pa->pa_deleted = 0; |
cc0fb9ad | 3832 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3833 | |
6ba495e9 | 3834 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3835 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3836 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3837 | |
3838 | ext4_mb_use_group_pa(ac, pa); | |
3839 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3840 | ||
3841 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3842 | lg = ac->ac_lg; | |
3843 | BUG_ON(lg == NULL); | |
3844 | ||
3845 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3846 | pa->pa_inode = NULL; | |
3847 | ||
3848 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3849 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3850 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3851 | ||
6be2ded1 AK |
3852 | /* |
3853 | * We will later add the new pa to the right bucket | |
3854 | * after updating the pa_free in ext4_mb_release_context | |
3855 | */ | |
c9de560d AT |
3856 | return 0; |
3857 | } | |
3858 | ||
3859 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3860 | { | |
3861 | int err; | |
3862 | ||
3863 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3864 | err = ext4_mb_new_group_pa(ac); | |
3865 | else | |
3866 | err = ext4_mb_new_inode_pa(ac); | |
3867 | return err; | |
3868 | } | |
3869 | ||
3870 | /* | |
3871 | * finds all unused blocks in on-disk bitmap, frees them in | |
3872 | * in-core bitmap and buddy. | |
3873 | * @pa must be unlinked from inode and group lists, so that | |
3874 | * nobody else can find/use it. | |
3875 | * the caller MUST hold group/inode locks. | |
3876 | * TODO: optimize the case when there are no in-core structures yet | |
3877 | */ | |
4ddfef7b ES |
3878 | static noinline_for_stack int |
3879 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3880 | struct ext4_prealloc_space *pa) |
c9de560d | 3881 | { |
c9de560d AT |
3882 | struct super_block *sb = e4b->bd_sb; |
3883 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3884 | unsigned int end; |
3885 | unsigned int next; | |
c9de560d AT |
3886 | ext4_group_t group; |
3887 | ext4_grpblk_t bit; | |
ba80b101 | 3888 | unsigned long long grp_blk_start; |
c9de560d AT |
3889 | int err = 0; |
3890 | int free = 0; | |
3891 | ||
3892 | BUG_ON(pa->pa_deleted == 0); | |
3893 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3894 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3895 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3896 | end = bit + pa->pa_len; | |
3897 | ||
c9de560d | 3898 | while (bit < end) { |
ffad0a44 | 3899 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3900 | if (bit >= end) |
3901 | break; | |
ffad0a44 | 3902 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3903 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3904 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3905 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3906 | free += next - bit; |
3907 | ||
3e1e5f50 | 3908 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3909 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3910 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3911 | next - bit); |
c9de560d AT |
3912 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3913 | bit = next + 1; | |
3914 | } | |
3915 | if (free != pa->pa_free) { | |
9d8b9ec4 TT |
3916 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
3917 | "pa %p: logic %lu, phys. %lu, len %lu", | |
3918 | pa, (unsigned long) pa->pa_lstart, | |
3919 | (unsigned long) pa->pa_pstart, | |
3920 | (unsigned long) pa->pa_len); | |
e29136f8 | 3921 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3922 | free, pa->pa_free); |
e56eb659 AK |
3923 | /* |
3924 | * pa is already deleted so we use the value obtained | |
3925 | * from the bitmap and continue. | |
3926 | */ | |
c9de560d | 3927 | } |
c9de560d AT |
3928 | atomic_add(free, &sbi->s_mb_discarded); |
3929 | ||
3930 | return err; | |
3931 | } | |
3932 | ||
4ddfef7b ES |
3933 | static noinline_for_stack int |
3934 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3935 | struct ext4_prealloc_space *pa) |
c9de560d | 3936 | { |
c9de560d AT |
3937 | struct super_block *sb = e4b->bd_sb; |
3938 | ext4_group_t group; | |
3939 | ext4_grpblk_t bit; | |
3940 | ||
60e07cf5 | 3941 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
3942 | BUG_ON(pa->pa_deleted == 0); |
3943 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3944 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3945 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3946 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3947 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3948 | |
3949 | return 0; | |
3950 | } | |
3951 | ||
3952 | /* | |
3953 | * releases all preallocations in given group | |
3954 | * | |
3955 | * first, we need to decide discard policy: | |
3956 | * - when do we discard | |
3957 | * 1) ENOSPC | |
3958 | * - how many do we discard | |
3959 | * 1) how many requested | |
3960 | */ | |
4ddfef7b ES |
3961 | static noinline_for_stack int |
3962 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3963 | ext4_group_t group, int needed) |
3964 | { | |
3965 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3966 | struct buffer_head *bitmap_bh = NULL; | |
3967 | struct ext4_prealloc_space *pa, *tmp; | |
3968 | struct list_head list; | |
3969 | struct ext4_buddy e4b; | |
3970 | int err; | |
3971 | int busy = 0; | |
3972 | int free = 0; | |
3973 | ||
6ba495e9 | 3974 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3975 | |
3976 | if (list_empty(&grp->bb_prealloc_list)) | |
3977 | return 0; | |
3978 | ||
574ca174 | 3979 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
3980 | if (IS_ERR(bitmap_bh)) { |
3981 | err = PTR_ERR(bitmap_bh); | |
3982 | ext4_error(sb, "Error %d reading block bitmap for %u", | |
3983 | err, group); | |
ce89f46c | 3984 | return 0; |
c9de560d AT |
3985 | } |
3986 | ||
3987 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3988 | if (err) { |
12062ddd | 3989 | ext4_error(sb, "Error loading buddy information for %u", group); |
ce89f46c AK |
3990 | put_bh(bitmap_bh); |
3991 | return 0; | |
3992 | } | |
c9de560d AT |
3993 | |
3994 | if (needed == 0) | |
7137d7a4 | 3995 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3996 | |
c9de560d | 3997 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3998 | repeat: |
3999 | ext4_lock_group(sb, group); | |
4000 | list_for_each_entry_safe(pa, tmp, | |
4001 | &grp->bb_prealloc_list, pa_group_list) { | |
4002 | spin_lock(&pa->pa_lock); | |
4003 | if (atomic_read(&pa->pa_count)) { | |
4004 | spin_unlock(&pa->pa_lock); | |
4005 | busy = 1; | |
4006 | continue; | |
4007 | } | |
4008 | if (pa->pa_deleted) { | |
4009 | spin_unlock(&pa->pa_lock); | |
4010 | continue; | |
4011 | } | |
4012 | ||
4013 | /* seems this one can be freed ... */ | |
4014 | pa->pa_deleted = 1; | |
4015 | ||
4016 | /* we can trust pa_free ... */ | |
4017 | free += pa->pa_free; | |
4018 | ||
4019 | spin_unlock(&pa->pa_lock); | |
4020 | ||
4021 | list_del(&pa->pa_group_list); | |
4022 | list_add(&pa->u.pa_tmp_list, &list); | |
4023 | } | |
4024 | ||
4025 | /* if we still need more blocks and some PAs were used, try again */ | |
4026 | if (free < needed && busy) { | |
4027 | busy = 0; | |
4028 | ext4_unlock_group(sb, group); | |
bb8b20ed | 4029 | cond_resched(); |
c9de560d AT |
4030 | goto repeat; |
4031 | } | |
4032 | ||
4033 | /* found anything to free? */ | |
4034 | if (list_empty(&list)) { | |
4035 | BUG_ON(free != 0); | |
4036 | goto out; | |
4037 | } | |
4038 | ||
4039 | /* now free all selected PAs */ | |
4040 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
4041 | ||
4042 | /* remove from object (inode or locality group) */ | |
4043 | spin_lock(pa->pa_obj_lock); | |
4044 | list_del_rcu(&pa->pa_inode_list); | |
4045 | spin_unlock(pa->pa_obj_lock); | |
4046 | ||
cc0fb9ad | 4047 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 4048 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 4049 | else |
3e1e5f50 | 4050 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4051 | |
4052 | list_del(&pa->u.pa_tmp_list); | |
4053 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4054 | } | |
4055 | ||
4056 | out: | |
4057 | ext4_unlock_group(sb, group); | |
e39e07fd | 4058 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4059 | put_bh(bitmap_bh); |
4060 | return free; | |
4061 | } | |
4062 | ||
4063 | /* | |
4064 | * releases all non-used preallocated blocks for given inode | |
4065 | * | |
4066 | * It's important to discard preallocations under i_data_sem | |
4067 | * We don't want another block to be served from the prealloc | |
4068 | * space when we are discarding the inode prealloc space. | |
4069 | * | |
4070 | * FIXME!! Make sure it is valid at all the call sites | |
4071 | */ | |
c2ea3fde | 4072 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
4073 | { |
4074 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4075 | struct super_block *sb = inode->i_sb; | |
4076 | struct buffer_head *bitmap_bh = NULL; | |
4077 | struct ext4_prealloc_space *pa, *tmp; | |
4078 | ext4_group_t group = 0; | |
4079 | struct list_head list; | |
4080 | struct ext4_buddy e4b; | |
4081 | int err; | |
4082 | ||
c2ea3fde | 4083 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4084 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4085 | return; | |
4086 | } | |
4087 | ||
6ba495e9 | 4088 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 4089 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
4090 | |
4091 | INIT_LIST_HEAD(&list); | |
4092 | ||
4093 | repeat: | |
4094 | /* first, collect all pa's in the inode */ | |
4095 | spin_lock(&ei->i_prealloc_lock); | |
4096 | while (!list_empty(&ei->i_prealloc_list)) { | |
4097 | pa = list_entry(ei->i_prealloc_list.next, | |
4098 | struct ext4_prealloc_space, pa_inode_list); | |
4099 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4100 | spin_lock(&pa->pa_lock); | |
4101 | if (atomic_read(&pa->pa_count)) { | |
4102 | /* this shouldn't happen often - nobody should | |
4103 | * use preallocation while we're discarding it */ | |
4104 | spin_unlock(&pa->pa_lock); | |
4105 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
4106 | ext4_msg(sb, KERN_ERR, |
4107 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
4108 | WARN_ON(1); |
4109 | schedule_timeout_uninterruptible(HZ); | |
4110 | goto repeat; | |
4111 | ||
4112 | } | |
4113 | if (pa->pa_deleted == 0) { | |
4114 | pa->pa_deleted = 1; | |
4115 | spin_unlock(&pa->pa_lock); | |
4116 | list_del_rcu(&pa->pa_inode_list); | |
4117 | list_add(&pa->u.pa_tmp_list, &list); | |
4118 | continue; | |
4119 | } | |
4120 | ||
4121 | /* someone is deleting pa right now */ | |
4122 | spin_unlock(&pa->pa_lock); | |
4123 | spin_unlock(&ei->i_prealloc_lock); | |
4124 | ||
4125 | /* we have to wait here because pa_deleted | |
4126 | * doesn't mean pa is already unlinked from | |
4127 | * the list. as we might be called from | |
4128 | * ->clear_inode() the inode will get freed | |
4129 | * and concurrent thread which is unlinking | |
4130 | * pa from inode's list may access already | |
4131 | * freed memory, bad-bad-bad */ | |
4132 | ||
4133 | /* XXX: if this happens too often, we can | |
4134 | * add a flag to force wait only in case | |
4135 | * of ->clear_inode(), but not in case of | |
4136 | * regular truncate */ | |
4137 | schedule_timeout_uninterruptible(HZ); | |
4138 | goto repeat; | |
4139 | } | |
4140 | spin_unlock(&ei->i_prealloc_lock); | |
4141 | ||
4142 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4143 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 4144 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d AT |
4145 | |
4146 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 4147 | if (err) { |
12062ddd ES |
4148 | ext4_error(sb, "Error loading buddy information for %u", |
4149 | group); | |
ce89f46c AK |
4150 | continue; |
4151 | } | |
c9de560d | 4152 | |
574ca174 | 4153 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4154 | if (IS_ERR(bitmap_bh)) { |
4155 | err = PTR_ERR(bitmap_bh); | |
4156 | ext4_error(sb, "Error %d reading block bitmap for %u", | |
4157 | err, group); | |
e39e07fd | 4158 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 4159 | continue; |
c9de560d AT |
4160 | } |
4161 | ||
4162 | ext4_lock_group(sb, group); | |
4163 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4164 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4165 | ext4_unlock_group(sb, group); |
4166 | ||
e39e07fd | 4167 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4168 | put_bh(bitmap_bh); |
4169 | ||
4170 | list_del(&pa->u.pa_tmp_list); | |
4171 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4172 | } | |
4173 | } | |
4174 | ||
6ba495e9 | 4175 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
4176 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
4177 | { | |
4178 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 4179 | ext4_group_t ngroups, i; |
c9de560d | 4180 | |
a0b30c12 | 4181 | if (!ext4_mballoc_debug || |
4dd89fc6 | 4182 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) |
e3570639 ES |
4183 | return; |
4184 | ||
7f6a11e7 | 4185 | ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" |
9d8b9ec4 | 4186 | " Allocation context details:"); |
7f6a11e7 | 4187 | ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", |
c9de560d | 4188 | ac->ac_status, ac->ac_flags); |
7f6a11e7 | 4189 | ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " |
9d8b9ec4 TT |
4190 | "goal %lu/%lu/%lu@%lu, " |
4191 | "best %lu/%lu/%lu@%lu cr %d", | |
c9de560d AT |
4192 | (unsigned long)ac->ac_o_ex.fe_group, |
4193 | (unsigned long)ac->ac_o_ex.fe_start, | |
4194 | (unsigned long)ac->ac_o_ex.fe_len, | |
4195 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4196 | (unsigned long)ac->ac_g_ex.fe_group, | |
4197 | (unsigned long)ac->ac_g_ex.fe_start, | |
4198 | (unsigned long)ac->ac_g_ex.fe_len, | |
4199 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4200 | (unsigned long)ac->ac_b_ex.fe_group, | |
4201 | (unsigned long)ac->ac_b_ex.fe_start, | |
4202 | (unsigned long)ac->ac_b_ex.fe_len, | |
4203 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4204 | (int)ac->ac_criteria); | |
dc9ddd98 | 4205 | ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); |
7f6a11e7 | 4206 | ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); |
8df9675f TT |
4207 | ngroups = ext4_get_groups_count(sb); |
4208 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
4209 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4210 | struct ext4_prealloc_space *pa; | |
4211 | ext4_grpblk_t start; | |
4212 | struct list_head *cur; | |
4213 | ext4_lock_group(sb, i); | |
4214 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4215 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4216 | pa_group_list); | |
4217 | spin_lock(&pa->pa_lock); | |
4218 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4219 | NULL, &start); | |
4220 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
4221 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
4222 | start, pa->pa_len); | |
c9de560d | 4223 | } |
60bd63d1 | 4224 | ext4_unlock_group(sb, i); |
c9de560d AT |
4225 | |
4226 | if (grp->bb_free == 0) | |
4227 | continue; | |
1c718505 | 4228 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
4229 | i, grp->bb_free, grp->bb_fragments); |
4230 | } | |
4231 | printk(KERN_ERR "\n"); | |
4232 | } | |
4233 | #else | |
4234 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4235 | { | |
4236 | return; | |
4237 | } | |
4238 | #endif | |
4239 | ||
4240 | /* | |
4241 | * We use locality group preallocation for small size file. The size of the | |
4242 | * file is determined by the current size or the resulting size after | |
4243 | * allocation which ever is larger | |
4244 | * | |
b713a5ec | 4245 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4246 | */ |
4247 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4248 | { | |
4249 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4250 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4251 | loff_t size, isize; | |
4252 | ||
4253 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4254 | return; | |
4255 | ||
4ba74d00 TT |
4256 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4257 | return; | |
4258 | ||
53accfa9 | 4259 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
4260 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4261 | >> bsbits; | |
c9de560d | 4262 | |
50797481 TT |
4263 | if ((size == isize) && |
4264 | !ext4_fs_is_busy(sbi) && | |
4265 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
4266 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
4267 | return; | |
4268 | } | |
4269 | ||
ebbe0277 RD |
4270 | if (sbi->s_mb_group_prealloc <= 0) { |
4271 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
4272 | return; | |
4273 | } | |
4274 | ||
c9de560d | 4275 | /* don't use group allocation for large files */ |
71780577 | 4276 | size = max(size, isize); |
cc483f10 | 4277 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 4278 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 4279 | return; |
4ba74d00 | 4280 | } |
c9de560d AT |
4281 | |
4282 | BUG_ON(ac->ac_lg != NULL); | |
4283 | /* | |
4284 | * locality group prealloc space are per cpu. The reason for having | |
4285 | * per cpu locality group is to reduce the contention between block | |
4286 | * request from multiple CPUs. | |
4287 | */ | |
a0b6bc63 | 4288 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
4289 | |
4290 | /* we're going to use group allocation */ | |
4291 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4292 | ||
4293 | /* serialize all allocations in the group */ | |
4294 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4295 | } | |
4296 | ||
4ddfef7b ES |
4297 | static noinline_for_stack int |
4298 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4299 | struct ext4_allocation_request *ar) |
4300 | { | |
4301 | struct super_block *sb = ar->inode->i_sb; | |
4302 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4303 | struct ext4_super_block *es = sbi->s_es; | |
4304 | ext4_group_t group; | |
498e5f24 TT |
4305 | unsigned int len; |
4306 | ext4_fsblk_t goal; | |
c9de560d AT |
4307 | ext4_grpblk_t block; |
4308 | ||
4309 | /* we can't allocate > group size */ | |
4310 | len = ar->len; | |
4311 | ||
4312 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
4313 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
4314 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4315 | |
4316 | /* start searching from the goal */ | |
4317 | goal = ar->goal; | |
4318 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4319 | goal >= ext4_blocks_count(es)) | |
4320 | goal = le32_to_cpu(es->s_first_data_block); | |
4321 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4322 | ||
4323 | /* set up allocation goals */ | |
f5a44db5 | 4324 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 4325 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4326 | ac->ac_sb = sb; |
4327 | ac->ac_inode = ar->inode; | |
53accfa9 | 4328 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4329 | ac->ac_o_ex.fe_group = group; |
4330 | ac->ac_o_ex.fe_start = block; | |
4331 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4332 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4333 | ac->ac_flags = ar->flags; |
c9de560d AT |
4334 | |
4335 | /* we have to define context: we'll we work with a file or | |
4336 | * locality group. this is a policy, actually */ | |
4337 | ext4_mb_group_or_file(ac); | |
4338 | ||
6ba495e9 | 4339 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4340 | "left: %u/%u, right %u/%u to %swritable\n", |
4341 | (unsigned) ar->len, (unsigned) ar->logical, | |
4342 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4343 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4344 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4345 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4346 | return 0; | |
4347 | ||
4348 | } | |
4349 | ||
6be2ded1 AK |
4350 | static noinline_for_stack void |
4351 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4352 | struct ext4_locality_group *lg, | |
4353 | int order, int total_entries) | |
4354 | { | |
4355 | ext4_group_t group = 0; | |
4356 | struct ext4_buddy e4b; | |
4357 | struct list_head discard_list; | |
4358 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4359 | |
6ba495e9 | 4360 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4361 | |
4362 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4363 | |
4364 | spin_lock(&lg->lg_prealloc_lock); | |
4365 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4366 | pa_inode_list) { | |
4367 | spin_lock(&pa->pa_lock); | |
4368 | if (atomic_read(&pa->pa_count)) { | |
4369 | /* | |
4370 | * This is the pa that we just used | |
4371 | * for block allocation. So don't | |
4372 | * free that | |
4373 | */ | |
4374 | spin_unlock(&pa->pa_lock); | |
4375 | continue; | |
4376 | } | |
4377 | if (pa->pa_deleted) { | |
4378 | spin_unlock(&pa->pa_lock); | |
4379 | continue; | |
4380 | } | |
4381 | /* only lg prealloc space */ | |
cc0fb9ad | 4382 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4383 | |
4384 | /* seems this one can be freed ... */ | |
4385 | pa->pa_deleted = 1; | |
4386 | spin_unlock(&pa->pa_lock); | |
4387 | ||
4388 | list_del_rcu(&pa->pa_inode_list); | |
4389 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4390 | ||
4391 | total_entries--; | |
4392 | if (total_entries <= 5) { | |
4393 | /* | |
4394 | * we want to keep only 5 entries | |
4395 | * allowing it to grow to 8. This | |
4396 | * mak sure we don't call discard | |
4397 | * soon for this list. | |
4398 | */ | |
4399 | break; | |
4400 | } | |
4401 | } | |
4402 | spin_unlock(&lg->lg_prealloc_lock); | |
4403 | ||
4404 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4405 | ||
bd86298e | 4406 | group = ext4_get_group_number(sb, pa->pa_pstart); |
6be2ded1 | 4407 | if (ext4_mb_load_buddy(sb, group, &e4b)) { |
12062ddd ES |
4408 | ext4_error(sb, "Error loading buddy information for %u", |
4409 | group); | |
6be2ded1 AK |
4410 | continue; |
4411 | } | |
4412 | ext4_lock_group(sb, group); | |
4413 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4414 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4415 | ext4_unlock_group(sb, group); |
4416 | ||
e39e07fd | 4417 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4418 | list_del(&pa->u.pa_tmp_list); |
4419 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4420 | } | |
6be2ded1 AK |
4421 | } |
4422 | ||
4423 | /* | |
4424 | * We have incremented pa_count. So it cannot be freed at this | |
4425 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4426 | * possible from this lg. That means pa_free cannot be updated. | |
4427 | * | |
4428 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4429 | * which can cause the lg_prealloc_list to be updated. | |
4430 | */ | |
4431 | ||
4432 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4433 | { | |
4434 | int order, added = 0, lg_prealloc_count = 1; | |
4435 | struct super_block *sb = ac->ac_sb; | |
4436 | struct ext4_locality_group *lg = ac->ac_lg; | |
4437 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4438 | ||
4439 | order = fls(pa->pa_free) - 1; | |
4440 | if (order > PREALLOC_TB_SIZE - 1) | |
4441 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4442 | order = PREALLOC_TB_SIZE - 1; | |
4443 | /* Add the prealloc space to lg */ | |
f1167009 | 4444 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4445 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
4446 | pa_inode_list) { | |
4447 | spin_lock(&tmp_pa->pa_lock); | |
4448 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4449 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4450 | continue; |
4451 | } | |
4452 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4453 | /* Add to the tail of the previous entry */ | |
4454 | list_add_tail_rcu(&pa->pa_inode_list, | |
4455 | &tmp_pa->pa_inode_list); | |
4456 | added = 1; | |
4457 | /* | |
4458 | * we want to count the total | |
4459 | * number of entries in the list | |
4460 | */ | |
4461 | } | |
4462 | spin_unlock(&tmp_pa->pa_lock); | |
4463 | lg_prealloc_count++; | |
4464 | } | |
4465 | if (!added) | |
4466 | list_add_tail_rcu(&pa->pa_inode_list, | |
4467 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4468 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4469 | |
4470 | /* Now trim the list to be not more than 8 elements */ | |
4471 | if (lg_prealloc_count > 8) { | |
4472 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4473 | order, lg_prealloc_count); |
6be2ded1 AK |
4474 | return; |
4475 | } | |
4476 | return ; | |
4477 | } | |
4478 | ||
c9de560d AT |
4479 | /* |
4480 | * release all resource we used in allocation | |
4481 | */ | |
4482 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4483 | { | |
53accfa9 | 4484 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4485 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4486 | if (pa) { | |
cc0fb9ad | 4487 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4488 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4489 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4490 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4491 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4492 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4493 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4494 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4495 | } |
c9de560d | 4496 | } |
ba443916 AK |
4497 | if (pa) { |
4498 | /* | |
4499 | * We want to add the pa to the right bucket. | |
4500 | * Remove it from the list and while adding | |
4501 | * make sure the list to which we are adding | |
44183d42 | 4502 | * doesn't grow big. |
ba443916 | 4503 | */ |
cc0fb9ad | 4504 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4505 | spin_lock(pa->pa_obj_lock); |
4506 | list_del_rcu(&pa->pa_inode_list); | |
4507 | spin_unlock(pa->pa_obj_lock); | |
4508 | ext4_mb_add_n_trim(ac); | |
4509 | } | |
4510 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4511 | } | |
c9de560d AT |
4512 | if (ac->ac_bitmap_page) |
4513 | page_cache_release(ac->ac_bitmap_page); | |
4514 | if (ac->ac_buddy_page) | |
4515 | page_cache_release(ac->ac_buddy_page); | |
4516 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4517 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4518 | ext4_mb_collect_stats(ac); | |
4519 | return 0; | |
4520 | } | |
4521 | ||
4522 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4523 | { | |
8df9675f | 4524 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4525 | int ret; |
4526 | int freed = 0; | |
4527 | ||
9bffad1e | 4528 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4529 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4530 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4531 | freed += ret; | |
4532 | needed -= ret; | |
4533 | } | |
4534 | ||
4535 | return freed; | |
4536 | } | |
4537 | ||
4538 | /* | |
4539 | * Main entry point into mballoc to allocate blocks | |
4540 | * it tries to use preallocation first, then falls back | |
4541 | * to usual allocation | |
4542 | */ | |
4543 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4544 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4545 | { |
6bc6e63f | 4546 | int freed; |
256bdb49 | 4547 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4548 | struct ext4_sb_info *sbi; |
4549 | struct super_block *sb; | |
4550 | ext4_fsblk_t block = 0; | |
60e58e0f | 4551 | unsigned int inquota = 0; |
53accfa9 | 4552 | unsigned int reserv_clstrs = 0; |
c9de560d | 4553 | |
b10a44c3 | 4554 | might_sleep(); |
c9de560d AT |
4555 | sb = ar->inode->i_sb; |
4556 | sbi = EXT4_SB(sb); | |
4557 | ||
9bffad1e | 4558 | trace_ext4_request_blocks(ar); |
ba80b101 | 4559 | |
45dc63e7 DM |
4560 | /* Allow to use superuser reservation for quota file */ |
4561 | if (IS_NOQUOTA(ar->inode)) | |
4562 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; | |
4563 | ||
1cac41cb MB |
4564 | if (ext4_test_inode_flag(ar->inode, EXT4_INODE_CORE_FILE)) |
4565 | ar->flags |= EXT4_MB_USE_EXTRA_ROOT_BLOCKS; | |
4566 | ||
e3cf5d5d | 4567 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
4568 | /* Without delayed allocation we need to verify |
4569 | * there is enough free blocks to do block allocation | |
4570 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4571 | */ |
55f020db | 4572 | while (ar->len && |
e7d5f315 | 4573 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4574 | |
030ba6bc | 4575 | /* let others to free the space */ |
bb8b20ed | 4576 | cond_resched(); |
030ba6bc AK |
4577 | ar->len = ar->len >> 1; |
4578 | } | |
4579 | if (!ar->len) { | |
a30d542a AK |
4580 | *errp = -ENOSPC; |
4581 | return 0; | |
4582 | } | |
53accfa9 | 4583 | reserv_clstrs = ar->len; |
55f020db | 4584 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4585 | dquot_alloc_block_nofail(ar->inode, |
4586 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4587 | } else { |
4588 | while (ar->len && | |
53accfa9 TT |
4589 | dquot_alloc_block(ar->inode, |
4590 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4591 | |
4592 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4593 | ar->len--; | |
4594 | } | |
60e58e0f MC |
4595 | } |
4596 | inquota = ar->len; | |
4597 | if (ar->len == 0) { | |
4598 | *errp = -EDQUOT; | |
6c7a120a | 4599 | goto out; |
60e58e0f | 4600 | } |
07031431 | 4601 | } |
d2a17637 | 4602 | |
85556c9a | 4603 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4604 | if (!ac) { |
363d4251 | 4605 | ar->len = 0; |
256bdb49 | 4606 | *errp = -ENOMEM; |
6c7a120a | 4607 | goto out; |
256bdb49 ES |
4608 | } |
4609 | ||
256bdb49 | 4610 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4611 | if (*errp) { |
4612 | ar->len = 0; | |
6c7a120a | 4613 | goto out; |
c9de560d AT |
4614 | } |
4615 | ||
256bdb49 ES |
4616 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4617 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4618 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4619 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4620 | repeat: |
4621 | /* allocate space in core */ | |
6c7a120a | 4622 | *errp = ext4_mb_regular_allocator(ac); |
2c00ef3e AK |
4623 | if (*errp) |
4624 | goto discard_and_exit; | |
c9de560d AT |
4625 | |
4626 | /* as we've just preallocated more space than | |
2c00ef3e | 4627 | * user requested originally, we store allocated |
c9de560d | 4628 | * space in a special descriptor */ |
256bdb49 | 4629 | if (ac->ac_status == AC_STATUS_FOUND && |
2c00ef3e AK |
4630 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) |
4631 | *errp = ext4_mb_new_preallocation(ac); | |
4632 | if (*errp) { | |
4633 | discard_and_exit: | |
4634 | ext4_discard_allocated_blocks(ac); | |
4635 | goto errout; | |
4636 | } | |
c9de560d | 4637 | } |
256bdb49 | 4638 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4639 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
3a22cf0c | 4640 | if (*errp) { |
b844167e | 4641 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4642 | goto errout; |
4643 | } else { | |
519deca0 AK |
4644 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4645 | ar->len = ac->ac_b_ex.fe_len; | |
4646 | } | |
c9de560d | 4647 | } else { |
256bdb49 | 4648 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4649 | if (freed) |
4650 | goto repeat; | |
4651 | *errp = -ENOSPC; | |
6c7a120a AK |
4652 | } |
4653 | ||
6d138ced | 4654 | errout: |
6c7a120a | 4655 | if (*errp) { |
256bdb49 | 4656 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4657 | ar->len = 0; |
256bdb49 | 4658 | ext4_mb_show_ac(ac); |
c9de560d | 4659 | } |
256bdb49 | 4660 | ext4_mb_release_context(ac); |
6c7a120a AK |
4661 | out: |
4662 | if (ac) | |
4663 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4664 | if (inquota && ar->len < inquota) |
53accfa9 | 4665 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4666 | if (!ar->len) { |
e3cf5d5d | 4667 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 4668 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4669 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4670 | reserv_clstrs); |
0087d9fb | 4671 | } |
c9de560d | 4672 | |
9bffad1e | 4673 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4674 | |
c9de560d AT |
4675 | return block; |
4676 | } | |
c9de560d | 4677 | |
c894058d AK |
4678 | /* |
4679 | * We can merge two free data extents only if the physical blocks | |
4680 | * are contiguous, AND the extents were freed by the same transaction, | |
4681 | * AND the blocks are associated with the same group. | |
4682 | */ | |
1cac41cb MB |
4683 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
4684 | struct ext4_free_data *entry, | |
4685 | struct ext4_free_data *new_entry, | |
4686 | struct rb_root *entry_rb_root) | |
c894058d | 4687 | { |
1cac41cb MB |
4688 | if ((entry->efd_tid != new_entry->efd_tid) || |
4689 | (entry->efd_group != new_entry->efd_group)) | |
4690 | return; | |
4691 | if (entry->efd_start_cluster + entry->efd_count == | |
4692 | new_entry->efd_start_cluster) { | |
4693 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
4694 | new_entry->efd_count += entry->efd_count; | |
4695 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
4696 | entry->efd_start_cluster) { | |
4697 | new_entry->efd_count += entry->efd_count; | |
4698 | } else | |
4699 | return; | |
4700 | spin_lock(&sbi->s_md_lock); | |
4701 | list_del(&entry->efd_list); | |
4702 | spin_unlock(&sbi->s_md_lock); | |
4703 | rb_erase(&entry->efd_node, entry_rb_root); | |
4704 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
4705 | } |
4706 | ||
4ddfef7b ES |
4707 | static noinline_for_stack int |
4708 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4709 | struct ext4_free_data *new_entry) |
c9de560d | 4710 | { |
e29136f8 | 4711 | ext4_group_t group = e4b->bd_group; |
84130193 | 4712 | ext4_grpblk_t cluster; |
7a2fcbf7 | 4713 | struct ext4_free_data *entry; |
c9de560d AT |
4714 | struct ext4_group_info *db = e4b->bd_info; |
4715 | struct super_block *sb = e4b->bd_sb; | |
4716 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4717 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4718 | struct rb_node *parent = NULL, *new_node; | |
4719 | ||
0390131b | 4720 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4721 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4722 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4723 | ||
18aadd47 BJ |
4724 | new_node = &new_entry->efd_node; |
4725 | cluster = new_entry->efd_start_cluster; | |
c894058d | 4726 | |
c894058d AK |
4727 | if (!*n) { |
4728 | /* first free block exent. We need to | |
4729 | protect buddy cache from being freed, | |
4730 | * otherwise we'll refresh it from | |
4731 | * on-disk bitmap and lose not-yet-available | |
4732 | * blocks */ | |
4733 | page_cache_get(e4b->bd_buddy_page); | |
4734 | page_cache_get(e4b->bd_bitmap_page); | |
4735 | } | |
4736 | while (*n) { | |
4737 | parent = *n; | |
18aadd47 BJ |
4738 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
4739 | if (cluster < entry->efd_start_cluster) | |
c894058d | 4740 | n = &(*n)->rb_left; |
18aadd47 | 4741 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
4742 | n = &(*n)->rb_right; |
4743 | else { | |
e29136f8 | 4744 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
4745 | ext4_group_first_block_no(sb, group) + |
4746 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 4747 | "Block already on to-be-freed list"); |
c894058d | 4748 | return 0; |
c9de560d | 4749 | } |
c894058d | 4750 | } |
c9de560d | 4751 | |
c894058d AK |
4752 | rb_link_node(new_node, parent, n); |
4753 | rb_insert_color(new_node, &db->bb_free_root); | |
4754 | ||
4755 | /* Now try to see the extent can be merged to left and right */ | |
4756 | node = rb_prev(new_node); | |
4757 | if (node) { | |
18aadd47 | 4758 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
1cac41cb MB |
4759 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
4760 | &(db->bb_free_root)); | |
c894058d | 4761 | } |
c9de560d | 4762 | |
c894058d AK |
4763 | node = rb_next(new_node); |
4764 | if (node) { | |
18aadd47 | 4765 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
1cac41cb MB |
4766 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
4767 | &(db->bb_free_root)); | |
c9de560d | 4768 | } |
1cac41cb MB |
4769 | |
4770 | spin_lock(&sbi->s_md_lock); | |
4771 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); | |
4772 | spin_unlock(&sbi->s_md_lock); | |
4773 | ||
c9de560d AT |
4774 | return 0; |
4775 | } | |
4776 | ||
44338711 TT |
4777 | /** |
4778 | * ext4_free_blocks() -- Free given blocks and update quota | |
4779 | * @handle: handle for this transaction | |
4780 | * @inode: inode | |
4781 | * @block: start physical block to free | |
4782 | * @count: number of blocks to count | |
5def1360 | 4783 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4784 | */ |
44338711 | 4785 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4786 | struct buffer_head *bh, ext4_fsblk_t block, |
4787 | unsigned long count, int flags) | |
c9de560d | 4788 | { |
26346ff6 | 4789 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4790 | struct super_block *sb = inode->i_sb; |
c9de560d | 4791 | struct ext4_group_desc *gdp; |
498e5f24 | 4792 | unsigned int overflow; |
c9de560d AT |
4793 | ext4_grpblk_t bit; |
4794 | struct buffer_head *gd_bh; | |
4795 | ext4_group_t block_group; | |
4796 | struct ext4_sb_info *sbi; | |
4797 | struct ext4_buddy e4b; | |
84130193 | 4798 | unsigned int count_clusters; |
c9de560d AT |
4799 | int err = 0; |
4800 | int ret; | |
4801 | ||
b10a44c3 | 4802 | might_sleep(); |
e6362609 TT |
4803 | if (bh) { |
4804 | if (block) | |
4805 | BUG_ON(block != bh->b_blocknr); | |
4806 | else | |
4807 | block = bh->b_blocknr; | |
4808 | } | |
c9de560d | 4809 | |
c9de560d | 4810 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4811 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4812 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4813 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4814 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4815 | goto error_return; |
4816 | } | |
4817 | ||
0610b6e9 | 4818 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4819 | trace_ext4_free_blocks(inode, block, count, flags); |
4820 | ||
9c02ac97 DJ |
4821 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
4822 | BUG_ON(count > 1); | |
e6362609 | 4823 | |
9c02ac97 DJ |
4824 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
4825 | inode, bh, block); | |
e6362609 TT |
4826 | } |
4827 | ||
60e6679e | 4828 | /* |
e6362609 TT |
4829 | * We need to make sure we don't reuse the freed block until |
4830 | * after the transaction is committed, which we can do by | |
4831 | * treating the block as metadata, below. We make an | |
4832 | * exception if the inode is to be written in writeback mode | |
4833 | * since writeback mode has weak data consistency guarantees. | |
4834 | */ | |
4835 | if (!ext4_should_writeback_data(inode)) | |
4836 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
c9de560d | 4837 | |
84130193 TT |
4838 | /* |
4839 | * If the extent to be freed does not begin on a cluster | |
4840 | * boundary, we need to deal with partial clusters at the | |
4841 | * beginning and end of the extent. Normally we will free | |
4842 | * blocks at the beginning or the end unless we are explicitly | |
4843 | * requested to avoid doing so. | |
4844 | */ | |
f5a44db5 | 4845 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
4846 | if (overflow) { |
4847 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
4848 | overflow = sbi->s_cluster_ratio - overflow; | |
4849 | block += overflow; | |
4850 | if (count > overflow) | |
4851 | count -= overflow; | |
4852 | else | |
4853 | return; | |
4854 | } else { | |
4855 | block -= overflow; | |
4856 | count += overflow; | |
4857 | } | |
4858 | } | |
f5a44db5 | 4859 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
4860 | if (overflow) { |
4861 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
4862 | if (count > overflow) | |
4863 | count -= overflow; | |
4864 | else | |
4865 | return; | |
4866 | } else | |
4867 | count += sbi->s_cluster_ratio - overflow; | |
4868 | } | |
4869 | ||
9c02ac97 DJ |
4870 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
4871 | int i; | |
4872 | ||
4873 | for (i = 0; i < count; i++) { | |
4874 | cond_resched(); | |
4875 | bh = sb_find_get_block(inode->i_sb, block + i); | |
4876 | if (!bh) | |
4877 | continue; | |
4878 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
4879 | inode, bh, block + i); | |
4880 | } | |
4881 | } | |
4882 | ||
c9de560d AT |
4883 | do_more: |
4884 | overflow = 0; | |
4885 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4886 | ||
163a203d DW |
4887 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
4888 | ext4_get_group_info(sb, block_group)))) | |
4889 | return; | |
4890 | ||
c9de560d AT |
4891 | /* |
4892 | * Check to see if we are freeing blocks across a group | |
4893 | * boundary. | |
4894 | */ | |
84130193 TT |
4895 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4896 | overflow = EXT4_C2B(sbi, bit) + count - | |
4897 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
4898 | count -= overflow; |
4899 | } | |
810da240 | 4900 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 4901 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
4902 | if (IS_ERR(bitmap_bh)) { |
4903 | err = PTR_ERR(bitmap_bh); | |
4904 | bitmap_bh = NULL; | |
c9de560d | 4905 | goto error_return; |
ce89f46c | 4906 | } |
c9de560d | 4907 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4908 | if (!gdp) { |
4909 | err = -EIO; | |
c9de560d | 4910 | goto error_return; |
ce89f46c | 4911 | } |
c9de560d AT |
4912 | |
4913 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4914 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4915 | in_range(block, ext4_inode_table(sb, gdp), | |
84130193 | 4916 | EXT4_SB(sb)->s_itb_per_group) || |
c9de560d | 4917 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
84130193 | 4918 | EXT4_SB(sb)->s_itb_per_group)) { |
c9de560d | 4919 | |
12062ddd | 4920 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4921 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4922 | /* err = 0. ext4_std_error should be a no op */ |
4923 | goto error_return; | |
c9de560d AT |
4924 | } |
4925 | ||
4926 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4927 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4928 | if (err) | |
4929 | goto error_return; | |
4930 | ||
4931 | /* | |
4932 | * We are about to modify some metadata. Call the journal APIs | |
4933 | * to unshare ->b_data if a currently-committing transaction is | |
4934 | * using it | |
4935 | */ | |
4936 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4937 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4938 | if (err) | |
4939 | goto error_return; | |
c9de560d AT |
4940 | #ifdef AGGRESSIVE_CHECK |
4941 | { | |
4942 | int i; | |
84130193 | 4943 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
4944 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
4945 | } | |
4946 | #endif | |
84130193 | 4947 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 4948 | |
d8aafd0c KK |
4949 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
4950 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
4951 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
4952 | if (err) |
4953 | goto error_return; | |
e6362609 TT |
4954 | |
4955 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
7a2fcbf7 AK |
4956 | struct ext4_free_data *new_entry; |
4957 | /* | |
4958 | * blocks being freed are metadata. these blocks shouldn't | |
4959 | * be used until this transaction is committed | |
7444a072 MH |
4960 | * |
4961 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed | |
4962 | * to fail. | |
7a2fcbf7 | 4963 | */ |
7444a072 MH |
4964 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
4965 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
4966 | new_entry->efd_start_cluster = bit; |
4967 | new_entry->efd_group = block_group; | |
4968 | new_entry->efd_count = count_clusters; | |
4969 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4970 | |
7a2fcbf7 | 4971 | ext4_lock_group(sb, block_group); |
84130193 | 4972 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 4973 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4974 | } else { |
7a2fcbf7 AK |
4975 | /* need to update group_info->bb_free and bitmap |
4976 | * with group lock held. generate_buddy look at | |
4977 | * them with group lock_held | |
4978 | */ | |
d71c1ae2 | 4979 | if (test_opt(sb, DISCARD)) { |
9a878d1d | 4980 | err = ext4_issue_discard(sb, block_group, bit, count, |
1cac41cb | 4981 | 0, NULL); |
d71c1ae2 LC |
4982 | if (err && err != -EOPNOTSUPP) |
4983 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
4984 | " group:%d block:%d count:%lu failed" | |
4985 | " with %d", block_group, bit, count, | |
4986 | err); | |
8f9ff189 LC |
4987 | } else |
4988 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 4989 | |
955ce5f5 | 4990 | ext4_lock_group(sb, block_group); |
84130193 TT |
4991 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
4992 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
4993 | } |
4994 | ||
021b65bb TT |
4995 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
4996 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 4997 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 4998 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 4999 | ext4_unlock_group(sb, block_group); |
c9de560d | 5000 | |
772cb7c8 JS |
5001 | if (sbi->s_log_groups_per_flex) { |
5002 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f TT |
5003 | atomic64_add(count_clusters, |
5004 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
5005 | } |
5006 | ||
71d4f7d0 | 5007 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) |
7b415bf6 | 5008 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); |
7d734532 JK |
5009 | percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); |
5010 | ||
5011 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 5012 | |
7a2fcbf7 AK |
5013 | /* We dirtied the bitmap block */ |
5014 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5015 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5016 | ||
c9de560d AT |
5017 | /* And the group descriptor block */ |
5018 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 5019 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
5020 | if (!err) |
5021 | err = ret; | |
5022 | ||
5023 | if (overflow && !err) { | |
5024 | block += count; | |
5025 | count = overflow; | |
5026 | put_bh(bitmap_bh); | |
5027 | goto do_more; | |
5028 | } | |
c9de560d AT |
5029 | error_return: |
5030 | brelse(bitmap_bh); | |
5031 | ext4_std_error(sb, err); | |
5032 | return; | |
5033 | } | |
7360d173 | 5034 | |
2846e820 | 5035 | /** |
0529155e | 5036 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
5037 | * @handle: handle to this transaction |
5038 | * @sb: super block | |
4907cb7b | 5039 | * @block: start physical block to add to the block group |
2846e820 AG |
5040 | * @count: number of blocks to free |
5041 | * | |
e73a347b | 5042 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 5043 | */ |
cc7365df | 5044 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
5045 | ext4_fsblk_t block, unsigned long count) |
5046 | { | |
5047 | struct buffer_head *bitmap_bh = NULL; | |
5048 | struct buffer_head *gd_bh; | |
5049 | ext4_group_t block_group; | |
5050 | ext4_grpblk_t bit; | |
5051 | unsigned int i; | |
5052 | struct ext4_group_desc *desc; | |
5053 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 5054 | struct ext4_buddy e4b; |
2846e820 AG |
5055 | int err = 0, ret, blk_free_count; |
5056 | ext4_grpblk_t blocks_freed; | |
2846e820 AG |
5057 | |
5058 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
5059 | ||
4740b830 YY |
5060 | if (count == 0) |
5061 | return 0; | |
5062 | ||
2846e820 | 5063 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
5064 | /* |
5065 | * Check to see if we are freeing blocks across a group | |
5066 | * boundary. | |
5067 | */ | |
cc7365df YY |
5068 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
5069 | ext4_warning(sb, "too much blocks added to group %u\n", | |
5070 | block_group); | |
5071 | err = -EINVAL; | |
2846e820 | 5072 | goto error_return; |
cc7365df | 5073 | } |
2cd05cc3 | 5074 | |
2846e820 | 5075 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5076 | if (IS_ERR(bitmap_bh)) { |
5077 | err = PTR_ERR(bitmap_bh); | |
5078 | bitmap_bh = NULL; | |
2846e820 | 5079 | goto error_return; |
cc7365df YY |
5080 | } |
5081 | ||
2846e820 | 5082 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
5083 | if (!desc) { |
5084 | err = -EIO; | |
2846e820 | 5085 | goto error_return; |
cc7365df | 5086 | } |
2846e820 AG |
5087 | |
5088 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
5089 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
5090 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
5091 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
5092 | sbi->s_itb_per_group)) { | |
5093 | ext4_error(sb, "Adding blocks in system zones - " | |
5094 | "Block = %llu, count = %lu", | |
5095 | block, count); | |
cc7365df | 5096 | err = -EINVAL; |
2846e820 AG |
5097 | goto error_return; |
5098 | } | |
5099 | ||
2cd05cc3 TT |
5100 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
5101 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
5102 | if (err) |
5103 | goto error_return; | |
5104 | ||
5105 | /* | |
5106 | * We are about to modify some metadata. Call the journal APIs | |
5107 | * to unshare ->b_data if a currently-committing transaction is | |
5108 | * using it | |
5109 | */ | |
5110 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
5111 | err = ext4_journal_get_write_access(handle, gd_bh); | |
5112 | if (err) | |
5113 | goto error_return; | |
e73a347b | 5114 | |
2846e820 AG |
5115 | for (i = 0, blocks_freed = 0; i < count; i++) { |
5116 | BUFFER_TRACE(bitmap_bh, "clear bit"); | |
e73a347b | 5117 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
5118 | ext4_error(sb, "bit already cleared for block %llu", |
5119 | (ext4_fsblk_t)(block + i)); | |
5120 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
5121 | } else { | |
5122 | blocks_freed++; | |
5123 | } | |
5124 | } | |
e73a347b AG |
5125 | |
5126 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
5127 | if (err) | |
5128 | goto error_return; | |
5129 | ||
5130 | /* | |
5131 | * need to update group_info->bb_free and bitmap | |
5132 | * with group lock held. generate_buddy look at | |
5133 | * them with group lock_held | |
5134 | */ | |
2846e820 | 5135 | ext4_lock_group(sb, block_group); |
e73a347b AG |
5136 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
5137 | mb_free_blocks(NULL, &e4b, bit, count); | |
021b65bb TT |
5138 | blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); |
5139 | ext4_free_group_clusters_set(sb, desc, blk_free_count); | |
79f1ba49 | 5140 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 5141 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 5142 | ext4_unlock_group(sb, block_group); |
57042651 | 5143 | percpu_counter_add(&sbi->s_freeclusters_counter, |
810da240 | 5144 | EXT4_NUM_B2C(sbi, blocks_freed)); |
2846e820 AG |
5145 | |
5146 | if (sbi->s_log_groups_per_flex) { | |
5147 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f TT |
5148 | atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed), |
5149 | &sbi->s_flex_groups[flex_group].free_clusters); | |
2846e820 | 5150 | } |
e73a347b AG |
5151 | |
5152 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
5153 | |
5154 | /* We dirtied the bitmap block */ | |
5155 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5156 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5157 | ||
5158 | /* And the group descriptor block */ | |
5159 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
5160 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
5161 | if (!err) | |
5162 | err = ret; | |
5163 | ||
5164 | error_return: | |
5165 | brelse(bitmap_bh); | |
5166 | ext4_std_error(sb, err); | |
cc7365df | 5167 | return err; |
2846e820 AG |
5168 | } |
5169 | ||
7360d173 LC |
5170 | /** |
5171 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
5172 | * @sb: super block for the file system | |
5173 | * @start: starting block of the free extent in the alloc. group | |
5174 | * @count: number of blocks to TRIM | |
5175 | * @group: alloc. group we are working with | |
5176 | * @e4b: ext4 buddy for the group | |
9a878d1d | 5177 | * @blkdev_flags: flags for the block device |
7360d173 LC |
5178 | * |
5179 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
5180 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
5181 | * be called with under the group lock. | |
5182 | */ | |
d71c1ae2 | 5183 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
9a878d1d JA |
5184 | ext4_group_t group, struct ext4_buddy *e4b, |
5185 | unsigned long blkdev_flags) | |
e2cbd587 | 5186 | __releases(bitlock) |
5187 | __acquires(bitlock) | |
7360d173 LC |
5188 | { |
5189 | struct ext4_free_extent ex; | |
d71c1ae2 | 5190 | int ret = 0; |
7360d173 | 5191 | |
b3d4c2b1 TM |
5192 | trace_ext4_trim_extent(sb, group, start, count); |
5193 | ||
7360d173 LC |
5194 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
5195 | ||
5196 | ex.fe_start = start; | |
5197 | ex.fe_group = group; | |
5198 | ex.fe_len = count; | |
5199 | ||
5200 | /* | |
5201 | * Mark blocks used, so no one can reuse them while | |
5202 | * being trimmed. | |
5203 | */ | |
5204 | mb_mark_used(e4b, &ex); | |
5205 | ext4_unlock_group(sb, group); | |
1cac41cb | 5206 | ret = ext4_issue_discard(sb, group, start, count, blkdev_flags, NULL); |
7360d173 LC |
5207 | ext4_lock_group(sb, group); |
5208 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 5209 | return ret; |
7360d173 LC |
5210 | } |
5211 | ||
5212 | /** | |
5213 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
5214 | * @sb: super block for file system | |
22612283 | 5215 | * @group: group to be trimmed |
7360d173 LC |
5216 | * @start: first group block to examine |
5217 | * @max: last group block to examine | |
5218 | * @minblocks: minimum extent block count | |
9a878d1d | 5219 | * @blkdev_flags: flags for the block device |
7360d173 LC |
5220 | * |
5221 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
5222 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
5223 | * the extent. | |
5224 | * | |
5225 | * | |
5226 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
5227 | * extents. When the free extent is found, mark it as used in group buddy | |
5228 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
5229 | * the group buddy bitmap. This is done until whole group is scanned. | |
5230 | */ | |
0b75a840 | 5231 | static ext4_grpblk_t |
78944086 LC |
5232 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
5233 | ext4_grpblk_t start, ext4_grpblk_t max, | |
9a878d1d | 5234 | ext4_grpblk_t minblocks, unsigned long blkdev_flags) |
7360d173 LC |
5235 | { |
5236 | void *bitmap; | |
169ddc3e | 5237 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 5238 | struct ext4_buddy e4b; |
d71c1ae2 | 5239 | int ret = 0; |
7360d173 | 5240 | |
b3d4c2b1 TM |
5241 | trace_ext4_trim_all_free(sb, group, start, max); |
5242 | ||
78944086 LC |
5243 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
5244 | if (ret) { | |
5245 | ext4_error(sb, "Error in loading buddy " | |
5246 | "information for %u", group); | |
5247 | return ret; | |
5248 | } | |
78944086 | 5249 | bitmap = e4b.bd_bitmap; |
28739eea LC |
5250 | |
5251 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
5252 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
5253 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
5254 | goto out; | |
5255 | ||
78944086 LC |
5256 | start = (e4b.bd_info->bb_first_free > start) ? |
5257 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 5258 | |
913eed83 LC |
5259 | while (start <= max) { |
5260 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
5261 | if (start > max) | |
7360d173 | 5262 | break; |
913eed83 | 5263 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
5264 | |
5265 | if ((next - start) >= minblocks) { | |
d71c1ae2 | 5266 | ret = ext4_trim_extent(sb, start, |
9a878d1d JA |
5267 | next - start, group, &e4b, |
5268 | blkdev_flags); | |
d71c1ae2 LC |
5269 | if (ret && ret != -EOPNOTSUPP) |
5270 | break; | |
5271 | ret = 0; | |
7360d173 LC |
5272 | count += next - start; |
5273 | } | |
169ddc3e | 5274 | free_count += next - start; |
7360d173 LC |
5275 | start = next + 1; |
5276 | ||
5277 | if (fatal_signal_pending(current)) { | |
5278 | count = -ERESTARTSYS; | |
5279 | break; | |
5280 | } | |
5281 | ||
5282 | if (need_resched()) { | |
5283 | ext4_unlock_group(sb, group); | |
5284 | cond_resched(); | |
5285 | ext4_lock_group(sb, group); | |
5286 | } | |
5287 | ||
169ddc3e | 5288 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
5289 | break; |
5290 | } | |
3d56b8d2 | 5291 | |
d71c1ae2 LC |
5292 | if (!ret) { |
5293 | ret = count; | |
3d56b8d2 | 5294 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 5295 | } |
3d56b8d2 | 5296 | out: |
7360d173 | 5297 | ext4_unlock_group(sb, group); |
78944086 | 5298 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
5299 | |
5300 | ext4_debug("trimmed %d blocks in the group %d\n", | |
5301 | count, group); | |
5302 | ||
d71c1ae2 | 5303 | return ret; |
7360d173 LC |
5304 | } |
5305 | ||
5306 | /** | |
5307 | * ext4_trim_fs() -- trim ioctl handle function | |
5308 | * @sb: superblock for filesystem | |
5309 | * @range: fstrim_range structure | |
9a878d1d | 5310 | * @blkdev_flags: flags for the block device |
7360d173 LC |
5311 | * |
5312 | * start: First Byte to trim | |
5313 | * len: number of Bytes to trim from start | |
5314 | * minlen: minimum extent length in Bytes | |
5315 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
5316 | * start to start+len. For each such a group ext4_trim_all_free function | |
5317 | * is invoked to trim all free space. | |
5318 | */ | |
9a878d1d JA |
5319 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range, |
5320 | unsigned long blkdev_flags) | |
7360d173 | 5321 | { |
78944086 | 5322 | struct ext4_group_info *grp; |
913eed83 | 5323 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 5324 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 5325 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
5326 | ext4_fsblk_t first_data_blk = |
5327 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 5328 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
5329 | int ret = 0; |
5330 | ||
5331 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 5332 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
5333 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
5334 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 5335 | |
5de35e8d LC |
5336 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
5337 | start >= max_blks || | |
5338 | range->len < sb->s_blocksize) | |
7360d173 | 5339 | return -EINVAL; |
913eed83 LC |
5340 | if (end >= max_blks) |
5341 | end = max_blks - 1; | |
5342 | if (end <= first_data_blk) | |
22f10457 | 5343 | goto out; |
913eed83 | 5344 | if (start < first_data_blk) |
0f0a25bf | 5345 | start = first_data_blk; |
7360d173 | 5346 | |
913eed83 | 5347 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5348 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5349 | &first_group, &first_cluster); |
913eed83 | 5350 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5351 | &last_group, &last_cluster); |
7360d173 | 5352 | |
913eed83 LC |
5353 | /* end now represents the last cluster to discard in this group */ |
5354 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5355 | |
5356 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5357 | grp = ext4_get_group_info(sb, group); |
5358 | /* We only do this if the grp has never been initialized */ | |
5359 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
d8aafd0c | 5360 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
5361 | if (ret) |
5362 | break; | |
7360d173 LC |
5363 | } |
5364 | ||
0ba08517 | 5365 | /* |
913eed83 LC |
5366 | * For all the groups except the last one, last cluster will |
5367 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5368 | * change it for the last group, note that last_cluster is | |
5369 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5370 | */ |
913eed83 LC |
5371 | if (group == last_group) |
5372 | end = last_cluster; | |
7360d173 | 5373 | |
78944086 | 5374 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5375 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
9a878d1d | 5376 | end, minlen, blkdev_flags); |
7360d173 LC |
5377 | if (cnt < 0) { |
5378 | ret = cnt; | |
7360d173 LC |
5379 | break; |
5380 | } | |
21e7fd22 | 5381 | trimmed += cnt; |
7360d173 | 5382 | } |
913eed83 LC |
5383 | |
5384 | /* | |
5385 | * For every group except the first one, we are sure | |
5386 | * that the first cluster to discard will be cluster #0. | |
5387 | */ | |
7137d7a4 | 5388 | first_cluster = 0; |
7360d173 | 5389 | } |
7360d173 | 5390 | |
3d56b8d2 TM |
5391 | if (!ret) |
5392 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5393 | ||
22f10457 | 5394 | out: |
aaf7d73e | 5395 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5396 | return ret; |
5397 | } |