Commit | Line | Data |
---|---|---|
f5166768 | 1 | // SPDX-License-Identifier: GPL-2.0 |
c9de560d AT |
2 | /* |
3 | * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com | |
4 | * Written by Alex Tomas <alex@clusterfs.com> | |
c9de560d AT |
5 | */ |
6 | ||
7 | ||
8 | /* | |
9 | * mballoc.c contains the multiblocks allocation routines | |
10 | */ | |
11 | ||
18aadd47 | 12 | #include "ext4_jbd2.h" |
8f6e39a7 | 13 | #include "mballoc.h" |
28623c2f | 14 | #include <linux/log2.h> |
a0b30c12 | 15 | #include <linux/module.h> |
5a0e3ad6 | 16 | #include <linux/slab.h> |
1a5d5e5d | 17 | #include <linux/nospec.h> |
66114cad | 18 | #include <linux/backing-dev.h> |
9bffad1e TT |
19 | #include <trace/events/ext4.h> |
20 | ||
a0b30c12 TT |
21 | #ifdef CONFIG_EXT4_DEBUG |
22 | ushort ext4_mballoc_debug __read_mostly; | |
23 | ||
24 | module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); | |
25 | MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); | |
26 | #endif | |
27 | ||
c9de560d AT |
28 | /* |
29 | * MUSTDO: | |
30 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
31 | * - search for metadata in few groups | |
32 | * | |
33 | * TODO v4: | |
34 | * - normalization should take into account whether file is still open | |
35 | * - discard preallocations if no free space left (policy?) | |
36 | * - don't normalize tails | |
37 | * - quota | |
38 | * - reservation for superuser | |
39 | * | |
40 | * TODO v3: | |
41 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
42 | * - track min/max extents in each group for better group selection | |
43 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
44 | * - tree of groups sorted by number of free blocks | |
45 | * - error handling | |
46 | */ | |
47 | ||
48 | /* | |
49 | * The allocation request involve request for multiple number of blocks | |
50 | * near to the goal(block) value specified. | |
51 | * | |
b713a5ec TT |
52 | * During initialization phase of the allocator we decide to use the |
53 | * group preallocation or inode preallocation depending on the size of | |
54 | * the file. The size of the file could be the resulting file size we | |
55 | * would have after allocation, or the current file size, which ever | |
56 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
57 | * select to use the group preallocation. The default value of | |
58 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
59 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
60 | * terms of number of blocks. | |
c9de560d AT |
61 | * |
62 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 63 | * ensure that we have small files closer together on the disk. |
c9de560d | 64 | * |
b713a5ec TT |
65 | * First stage the allocator looks at the inode prealloc list, |
66 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
67 | * spaces for this particular inode. The inode prealloc space is | |
68 | * represented as: | |
c9de560d AT |
69 | * |
70 | * pa_lstart -> the logical start block for this prealloc space | |
71 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
72 | * pa_len -> length for this prealloc space (in clusters) |
73 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
74 | * |
75 | * The inode preallocation space is used looking at the _logical_ start | |
76 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
77 | * space we will consume the particular prealloc space. This makes sure that |
78 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
79 | * |
80 | * The important thing to be noted in case of inode prealloc space is that | |
81 | * we don't modify the values associated to inode prealloc space except | |
82 | * pa_free. | |
83 | * | |
84 | * If we are not able to find blocks in the inode prealloc space and if we | |
85 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 86 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
87 | * |
88 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
89 | * | |
90 | * The reason for having a per cpu locality group is to reduce the contention | |
91 | * between CPUs. It is possible to get scheduled at this point. | |
92 | * | |
93 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 94 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
95 | * |
96 | * If we can't allocate blocks via inode prealloc or/and locality group | |
97 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
98 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
99 | * mapped to the buddy and bitmap information regarding different | |
100 | * groups. The buddy information is attached to buddy cache inode so that | |
101 | * we can access them through the page cache. The information regarding | |
102 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
103 | * block bitmap and buddy information. The information are stored in the | |
104 | * inode as: | |
105 | * | |
106 | * { page } | |
c3a326a6 | 107 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
108 | * |
109 | * | |
110 | * one block each for bitmap and buddy information. So for each group we | |
ea1754a0 | 111 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / |
c9de560d AT |
112 | * blocksize) blocks. So it can have information regarding groups_per_page |
113 | * which is blocks_per_page/2 | |
114 | * | |
115 | * The buddy cache inode is not stored on disk. The inode is thrown | |
116 | * away when the filesystem is unmounted. | |
117 | * | |
118 | * We look for count number of blocks in the buddy cache. If we were able | |
119 | * to locate that many free blocks we return with additional information | |
120 | * regarding rest of the contiguous physical block available | |
121 | * | |
122 | * Before allocating blocks via buddy cache we normalize the request | |
123 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
124 | * blocks that we get after allocation is added to the respective prealloc | |
125 | * list. In case of inode preallocation we follow a list of heuristics | |
126 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
127 | * we are doing a group prealloc we try to normalize the request to | |
27baebb8 TT |
128 | * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is |
129 | * dependent on the cluster size; for non-bigalloc file systems, it is | |
c9de560d | 130 | * 512 blocks. This can be tuned via |
d7a1fee1 | 131 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
132 | * terms of number of blocks. If we have mounted the file system with -O |
133 | * stripe=<value> option the group prealloc request is normalized to the | |
d7a1fee1 DE |
134 | * the smallest multiple of the stripe value (sbi->s_stripe) which is |
135 | * greater than the default mb_group_prealloc. | |
c9de560d | 136 | * |
d7a1fee1 | 137 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 138 | * |
b713a5ec TT |
139 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
140 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
141 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 142 | * |
b713a5ec | 143 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
144 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
145 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 146 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 147 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
148 | * stripe size. This should result in better allocation on RAID setups. If |
149 | * not, we search in the specific group using bitmap for best extents. The | |
150 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 151 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 152 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
153 | * best extent in the found extents. Searching for the blocks starts with |
154 | * the group specified as the goal value in allocation context via | |
155 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 156 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
157 | * checked. |
158 | * | |
159 | * Both the prealloc space are getting populated as above. So for the first | |
160 | * request we will hit the buddy cache which will result in this prealloc | |
161 | * space getting filled. The prealloc space is then later used for the | |
162 | * subsequent request. | |
163 | */ | |
164 | ||
165 | /* | |
166 | * mballoc operates on the following data: | |
167 | * - on-disk bitmap | |
168 | * - in-core buddy (actually includes buddy and bitmap) | |
169 | * - preallocation descriptors (PAs) | |
170 | * | |
171 | * there are two types of preallocations: | |
172 | * - inode | |
173 | * assiged to specific inode and can be used for this inode only. | |
174 | * it describes part of inode's space preallocated to specific | |
175 | * physical blocks. any block from that preallocated can be used | |
176 | * independent. the descriptor just tracks number of blocks left | |
177 | * unused. so, before taking some block from descriptor, one must | |
178 | * make sure corresponded logical block isn't allocated yet. this | |
179 | * also means that freeing any block within descriptor's range | |
180 | * must discard all preallocated blocks. | |
181 | * - locality group | |
182 | * assigned to specific locality group which does not translate to | |
183 | * permanent set of inodes: inode can join and leave group. space | |
184 | * from this type of preallocation can be used for any inode. thus | |
185 | * it's consumed from the beginning to the end. | |
186 | * | |
187 | * relation between them can be expressed as: | |
188 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
189 | * | |
190 | * this mean blocks mballoc considers used are: | |
191 | * - allocated blocks (persistent) | |
192 | * - preallocated blocks (non-persistent) | |
193 | * | |
194 | * consistency in mballoc world means that at any time a block is either | |
195 | * free or used in ALL structures. notice: "any time" should not be read | |
196 | * literally -- time is discrete and delimited by locks. | |
197 | * | |
198 | * to keep it simple, we don't use block numbers, instead we count number of | |
199 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
200 | * | |
201 | * all operations can be expressed as: | |
202 | * - init buddy: buddy = on-disk + PAs | |
203 | * - new PA: buddy += N; PA = N | |
204 | * - use inode PA: on-disk += N; PA -= N | |
205 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
206 | * - use locality group PA on-disk += N; PA -= N | |
207 | * - discard locality group PA buddy -= PA; PA = 0 | |
208 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
209 | * is used in real operation because we can't know actual used | |
210 | * bits from PA, only from on-disk bitmap | |
211 | * | |
212 | * if we follow this strict logic, then all operations above should be atomic. | |
213 | * given some of them can block, we'd have to use something like semaphores | |
214 | * killing performance on high-end SMP hardware. let's try to relax it using | |
215 | * the following knowledge: | |
216 | * 1) if buddy is referenced, it's already initialized | |
217 | * 2) while block is used in buddy and the buddy is referenced, | |
218 | * nobody can re-allocate that block | |
219 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
220 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
221 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
222 | * block | |
223 | * | |
224 | * so, now we're building a concurrency table: | |
225 | * - init buddy vs. | |
226 | * - new PA | |
227 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
228 | * until PA is linked to allocation group to avoid concurrent buddy init | |
229 | * - use inode PA | |
230 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
231 | * given (3) we care that PA-=N operation doesn't interfere with init | |
232 | * - discard inode PA | |
233 | * the simplest way would be to have buddy initialized by the discard | |
234 | * - use locality group PA | |
235 | * again PA-=N must be serialized with init | |
236 | * - discard locality group PA | |
237 | * the simplest way would be to have buddy initialized by the discard | |
238 | * - new PA vs. | |
239 | * - use inode PA | |
240 | * i_data_sem serializes them | |
241 | * - discard inode PA | |
242 | * discard process must wait until PA isn't used by another process | |
243 | * - use locality group PA | |
244 | * some mutex should serialize them | |
245 | * - discard locality group PA | |
246 | * discard process must wait until PA isn't used by another process | |
247 | * - use inode PA | |
248 | * - use inode PA | |
249 | * i_data_sem or another mutex should serializes them | |
250 | * - discard inode PA | |
251 | * discard process must wait until PA isn't used by another process | |
252 | * - use locality group PA | |
253 | * nothing wrong here -- they're different PAs covering different blocks | |
254 | * - discard locality group PA | |
255 | * discard process must wait until PA isn't used by another process | |
256 | * | |
257 | * now we're ready to make few consequences: | |
258 | * - PA is referenced and while it is no discard is possible | |
259 | * - PA is referenced until block isn't marked in on-disk bitmap | |
260 | * - PA changes only after on-disk bitmap | |
261 | * - discard must not compete with init. either init is done before | |
262 | * any discard or they're serialized somehow | |
263 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
264 | * | |
265 | * a special case when we've used PA to emptiness. no need to modify buddy | |
266 | * in this case, but we should care about concurrent init | |
267 | * | |
268 | */ | |
269 | ||
270 | /* | |
271 | * Logic in few words: | |
272 | * | |
273 | * - allocation: | |
274 | * load group | |
275 | * find blocks | |
276 | * mark bits in on-disk bitmap | |
277 | * release group | |
278 | * | |
279 | * - use preallocation: | |
280 | * find proper PA (per-inode or group) | |
281 | * load group | |
282 | * mark bits in on-disk bitmap | |
283 | * release group | |
284 | * release PA | |
285 | * | |
286 | * - free: | |
287 | * load group | |
288 | * mark bits in on-disk bitmap | |
289 | * release group | |
290 | * | |
291 | * - discard preallocations in group: | |
292 | * mark PAs deleted | |
293 | * move them onto local list | |
294 | * load on-disk bitmap | |
295 | * load group | |
296 | * remove PA from object (inode or locality group) | |
297 | * mark free blocks in-core | |
298 | * | |
299 | * - discard inode's preallocations: | |
300 | */ | |
301 | ||
302 | /* | |
303 | * Locking rules | |
304 | * | |
305 | * Locks: | |
306 | * - bitlock on a group (group) | |
307 | * - object (inode/locality) (object) | |
308 | * - per-pa lock (pa) | |
309 | * | |
310 | * Paths: | |
311 | * - new pa | |
312 | * object | |
313 | * group | |
314 | * | |
315 | * - find and use pa: | |
316 | * pa | |
317 | * | |
318 | * - release consumed pa: | |
319 | * pa | |
320 | * group | |
321 | * object | |
322 | * | |
323 | * - generate in-core bitmap: | |
324 | * group | |
325 | * pa | |
326 | * | |
327 | * - discard all for given object (inode, locality group): | |
328 | * object | |
329 | * pa | |
330 | * group | |
331 | * | |
332 | * - discard all for given group: | |
333 | * group | |
334 | * pa | |
335 | * group | |
336 | * object | |
337 | * | |
338 | */ | |
c3a326a6 AK |
339 | static struct kmem_cache *ext4_pspace_cachep; |
340 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 341 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
342 | |
343 | /* We create slab caches for groupinfo data structures based on the | |
344 | * superblock block size. There will be one per mounted filesystem for | |
345 | * each unique s_blocksize_bits */ | |
2892c15d | 346 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
347 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
348 | ||
d6006186 | 349 | static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
2892c15d ES |
350 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", |
351 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
352 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
353 | }; | |
354 | ||
c3a326a6 AK |
355 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
356 | ext4_group_t group); | |
7a2fcbf7 AK |
357 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
358 | ext4_group_t group); | |
c3a326a6 | 359 | |
ffad0a44 AK |
360 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
361 | { | |
c9de560d | 362 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
363 | *bit += ((unsigned long) addr & 7UL) << 3; |
364 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 365 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
366 | *bit += ((unsigned long) addr & 3UL) << 3; |
367 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
368 | #else |
369 | #error "how many bits you are?!" | |
370 | #endif | |
ffad0a44 AK |
371 | return addr; |
372 | } | |
c9de560d AT |
373 | |
374 | static inline int mb_test_bit(int bit, void *addr) | |
375 | { | |
376 | /* | |
377 | * ext4_test_bit on architecture like powerpc | |
378 | * needs unsigned long aligned address | |
379 | */ | |
ffad0a44 | 380 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
381 | return ext4_test_bit(bit, addr); |
382 | } | |
383 | ||
384 | static inline void mb_set_bit(int bit, void *addr) | |
385 | { | |
ffad0a44 | 386 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
387 | ext4_set_bit(bit, addr); |
388 | } | |
389 | ||
c9de560d AT |
390 | static inline void mb_clear_bit(int bit, void *addr) |
391 | { | |
ffad0a44 | 392 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
393 | ext4_clear_bit(bit, addr); |
394 | } | |
395 | ||
eabe0444 AS |
396 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
397 | { | |
398 | addr = mb_correct_addr_and_bit(&bit, addr); | |
399 | return ext4_test_and_clear_bit(bit, addr); | |
400 | } | |
401 | ||
ffad0a44 AK |
402 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
403 | { | |
e7dfb246 | 404 | int fix = 0, ret, tmpmax; |
ffad0a44 | 405 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 406 | tmpmax = max + fix; |
ffad0a44 AK |
407 | start += fix; |
408 | ||
e7dfb246 AK |
409 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
410 | if (ret > max) | |
411 | return max; | |
412 | return ret; | |
ffad0a44 AK |
413 | } |
414 | ||
415 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
416 | { | |
e7dfb246 | 417 | int fix = 0, ret, tmpmax; |
ffad0a44 | 418 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 419 | tmpmax = max + fix; |
ffad0a44 AK |
420 | start += fix; |
421 | ||
e7dfb246 AK |
422 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
423 | if (ret > max) | |
424 | return max; | |
425 | return ret; | |
ffad0a44 AK |
426 | } |
427 | ||
c9de560d AT |
428 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
429 | { | |
430 | char *bb; | |
431 | ||
c5e8f3f3 | 432 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
433 | BUG_ON(max == NULL); |
434 | ||
435 | if (order > e4b->bd_blkbits + 1) { | |
436 | *max = 0; | |
437 | return NULL; | |
438 | } | |
439 | ||
440 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
441 | if (order == 0) { |
442 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 443 | return e4b->bd_bitmap; |
84b775a3 | 444 | } |
c9de560d | 445 | |
c5e8f3f3 | 446 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
447 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
448 | ||
449 | return bb; | |
450 | } | |
451 | ||
452 | #ifdef DOUBLE_CHECK | |
453 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
454 | int first, int count) | |
455 | { | |
456 | int i; | |
457 | struct super_block *sb = e4b->bd_sb; | |
458 | ||
459 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
460 | return; | |
bc8e6740 | 461 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
462 | for (i = 0; i < count; i++) { |
463 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
464 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
465 | |
466 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 467 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 468 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
469 | inode ? inode->i_ino : 0, |
470 | blocknr, | |
471 | "freeing block already freed " | |
472 | "(bit %u)", | |
473 | first + i); | |
736dedbb WS |
474 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
475 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
476 | } |
477 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
478 | } | |
479 | } | |
480 | ||
481 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
482 | { | |
483 | int i; | |
484 | ||
485 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
486 | return; | |
bc8e6740 | 487 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
488 | for (i = 0; i < count; i++) { |
489 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
490 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
491 | } | |
492 | } | |
493 | ||
494 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
495 | { | |
496 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
497 | unsigned char *b1, *b2; | |
498 | int i; | |
499 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
500 | b2 = (unsigned char *) bitmap; | |
501 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
502 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
503 | ext4_msg(e4b->bd_sb, KERN_ERR, |
504 | "corruption in group %u " | |
505 | "at byte %u(%u): %x in copy != %x " | |
506 | "on disk/prealloc", | |
507 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
508 | BUG(); |
509 | } | |
510 | } | |
511 | } | |
512 | } | |
513 | ||
514 | #else | |
515 | static inline void mb_free_blocks_double(struct inode *inode, | |
516 | struct ext4_buddy *e4b, int first, int count) | |
517 | { | |
518 | return; | |
519 | } | |
520 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
521 | int first, int count) | |
522 | { | |
523 | return; | |
524 | } | |
525 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
526 | { | |
527 | return; | |
528 | } | |
529 | #endif | |
530 | ||
531 | #ifdef AGGRESSIVE_CHECK | |
532 | ||
533 | #define MB_CHECK_ASSERT(assert) \ | |
534 | do { \ | |
535 | if (!(assert)) { \ | |
536 | printk(KERN_EMERG \ | |
537 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
538 | function, file, line, # assert); \ | |
539 | BUG(); \ | |
540 | } \ | |
541 | } while (0) | |
542 | ||
543 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
544 | const char *function, int line) | |
545 | { | |
546 | struct super_block *sb = e4b->bd_sb; | |
547 | int order = e4b->bd_blkbits + 1; | |
548 | int max; | |
549 | int max2; | |
550 | int i; | |
551 | int j; | |
552 | int k; | |
553 | int count; | |
554 | struct ext4_group_info *grp; | |
555 | int fragments = 0; | |
556 | int fstart; | |
557 | struct list_head *cur; | |
558 | void *buddy; | |
559 | void *buddy2; | |
560 | ||
c9de560d AT |
561 | { |
562 | static int mb_check_counter; | |
563 | if (mb_check_counter++ % 100 != 0) | |
564 | return 0; | |
565 | } | |
566 | ||
567 | while (order > 1) { | |
568 | buddy = mb_find_buddy(e4b, order, &max); | |
569 | MB_CHECK_ASSERT(buddy); | |
570 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
571 | MB_CHECK_ASSERT(buddy2); | |
572 | MB_CHECK_ASSERT(buddy != buddy2); | |
573 | MB_CHECK_ASSERT(max * 2 == max2); | |
574 | ||
575 | count = 0; | |
576 | for (i = 0; i < max; i++) { | |
577 | ||
578 | if (mb_test_bit(i, buddy)) { | |
579 | /* only single bit in buddy2 may be 1 */ | |
580 | if (!mb_test_bit(i << 1, buddy2)) { | |
581 | MB_CHECK_ASSERT( | |
582 | mb_test_bit((i<<1)+1, buddy2)); | |
583 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
584 | MB_CHECK_ASSERT( | |
585 | mb_test_bit(i << 1, buddy2)); | |
586 | } | |
587 | continue; | |
588 | } | |
589 | ||
0a10da73 | 590 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
591 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
592 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
593 | ||
594 | for (j = 0; j < (1 << order); j++) { | |
595 | k = (i * (1 << order)) + j; | |
596 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 597 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
598 | } |
599 | count++; | |
600 | } | |
601 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
602 | order--; | |
603 | } | |
604 | ||
605 | fstart = -1; | |
606 | buddy = mb_find_buddy(e4b, 0, &max); | |
607 | for (i = 0; i < max; i++) { | |
608 | if (!mb_test_bit(i, buddy)) { | |
609 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
610 | if (fstart == -1) { | |
611 | fragments++; | |
612 | fstart = i; | |
613 | } | |
614 | continue; | |
615 | } | |
616 | fstart = -1; | |
617 | /* check used bits only */ | |
618 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
619 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
620 | k = i >> j; | |
621 | MB_CHECK_ASSERT(k < max2); | |
622 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
623 | } | |
624 | } | |
625 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
626 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
627 | ||
628 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
629 | list_for_each(cur, &grp->bb_prealloc_list) { |
630 | ext4_group_t groupnr; | |
631 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
632 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
633 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 634 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 635 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
636 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
637 | } | |
638 | return 0; | |
639 | } | |
640 | #undef MB_CHECK_ASSERT | |
641 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 642 | __FILE__, __func__, __LINE__) |
c9de560d AT |
643 | #else |
644 | #define mb_check_buddy(e4b) | |
645 | #endif | |
646 | ||
7c786059 CL |
647 | /* |
648 | * Divide blocks started from @first with length @len into | |
649 | * smaller chunks with power of 2 blocks. | |
650 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
651 | * then increase bb_counters[] for corresponded chunk size. | |
652 | */ | |
c9de560d | 653 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 654 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
655 | struct ext4_group_info *grp) |
656 | { | |
657 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
658 | ext4_grpblk_t min; |
659 | ext4_grpblk_t max; | |
660 | ext4_grpblk_t chunk; | |
69e43e8c | 661 | unsigned int border; |
c9de560d | 662 | |
7137d7a4 | 663 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
664 | |
665 | border = 2 << sb->s_blocksize_bits; | |
666 | ||
667 | while (len > 0) { | |
668 | /* find how many blocks can be covered since this position */ | |
669 | max = ffs(first | border) - 1; | |
670 | ||
671 | /* find how many blocks of power 2 we need to mark */ | |
672 | min = fls(len) - 1; | |
673 | ||
674 | if (max < min) | |
675 | min = max; | |
676 | chunk = 1 << min; | |
677 | ||
678 | /* mark multiblock chunks only */ | |
679 | grp->bb_counters[min]++; | |
680 | if (min > 0) | |
681 | mb_clear_bit(first >> min, | |
682 | buddy + sbi->s_mb_offsets[min]); | |
683 | ||
684 | len -= chunk; | |
685 | first += chunk; | |
686 | } | |
687 | } | |
688 | ||
8a57d9d6 CW |
689 | /* |
690 | * Cache the order of the largest free extent we have available in this block | |
691 | * group. | |
692 | */ | |
693 | static void | |
694 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
695 | { | |
696 | int i; | |
697 | int bits; | |
698 | ||
699 | grp->bb_largest_free_order = -1; /* uninit */ | |
700 | ||
701 | bits = sb->s_blocksize_bits + 1; | |
702 | for (i = bits; i >= 0; i--) { | |
703 | if (grp->bb_counters[i] > 0) { | |
704 | grp->bb_largest_free_order = i; | |
705 | break; | |
706 | } | |
707 | } | |
708 | } | |
709 | ||
089ceecc ES |
710 | static noinline_for_stack |
711 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
712 | void *buddy, void *bitmap, ext4_group_t group) |
713 | { | |
714 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 715 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 716 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
717 | ext4_grpblk_t i = 0; |
718 | ext4_grpblk_t first; | |
719 | ext4_grpblk_t len; | |
c9de560d AT |
720 | unsigned free = 0; |
721 | unsigned fragments = 0; | |
722 | unsigned long long period = get_cycles(); | |
723 | ||
724 | /* initialize buddy from bitmap which is aggregation | |
725 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 726 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
727 | grp->bb_first_free = i; |
728 | while (i < max) { | |
729 | fragments++; | |
730 | first = i; | |
ffad0a44 | 731 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
732 | len = i - first; |
733 | free += len; | |
734 | if (len > 1) | |
735 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
736 | else | |
737 | grp->bb_counters[0]++; | |
738 | if (i < max) | |
ffad0a44 | 739 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
740 | } |
741 | grp->bb_fragments = fragments; | |
742 | ||
743 | if (free != grp->bb_free) { | |
e29136f8 | 744 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
745 | "block bitmap and bg descriptor " |
746 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 747 | free, grp->bb_free); |
e56eb659 | 748 | /* |
163a203d | 749 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
750 | * corrupt and update bb_free using bitmap value |
751 | */ | |
c9de560d | 752 | grp->bb_free = free; |
db79e6d1 WS |
753 | ext4_mark_group_bitmap_corrupted(sb, group, |
754 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 755 | } |
8a57d9d6 | 756 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
757 | |
758 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
759 | ||
760 | period = get_cycles() - period; | |
49598e04 JP |
761 | spin_lock(&sbi->s_bal_lock); |
762 | sbi->s_mb_buddies_generated++; | |
763 | sbi->s_mb_generation_time += period; | |
764 | spin_unlock(&sbi->s_bal_lock); | |
c9de560d AT |
765 | } |
766 | ||
eabe0444 AS |
767 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
768 | { | |
769 | int count; | |
770 | int order = 1; | |
771 | void *buddy; | |
772 | ||
773 | while ((buddy = mb_find_buddy(e4b, order++, &count))) { | |
774 | ext4_set_bits(buddy, 0, count); | |
775 | } | |
776 | e4b->bd_info->bb_fragments = 0; | |
777 | memset(e4b->bd_info->bb_counters, 0, | |
778 | sizeof(*e4b->bd_info->bb_counters) * | |
779 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
780 | ||
781 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
782 | e4b->bd_bitmap, e4b->bd_group); | |
783 | } | |
784 | ||
c9de560d AT |
785 | /* The buddy information is attached the buddy cache inode |
786 | * for convenience. The information regarding each group | |
787 | * is loaded via ext4_mb_load_buddy. The information involve | |
788 | * block bitmap and buddy information. The information are | |
789 | * stored in the inode as | |
790 | * | |
791 | * { page } | |
c3a326a6 | 792 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
793 | * |
794 | * | |
795 | * one block each for bitmap and buddy information. | |
796 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 797 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
798 | * So it can have information regarding groups_per_page which |
799 | * is blocks_per_page/2 | |
8a57d9d6 CW |
800 | * |
801 | * Locking note: This routine takes the block group lock of all groups | |
802 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
803 | */ |
804 | ||
adb7ef60 | 805 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 806 | { |
8df9675f | 807 | ext4_group_t ngroups; |
c9de560d AT |
808 | int blocksize; |
809 | int blocks_per_page; | |
810 | int groups_per_page; | |
811 | int err = 0; | |
812 | int i; | |
813e5727 | 813 | ext4_group_t first_group, group; |
c9de560d AT |
814 | int first_block; |
815 | struct super_block *sb; | |
816 | struct buffer_head *bhs; | |
fa77dcfa | 817 | struct buffer_head **bh = NULL; |
c9de560d AT |
818 | struct inode *inode; |
819 | char *data; | |
820 | char *bitmap; | |
9b8b7d35 | 821 | struct ext4_group_info *grinfo; |
c9de560d | 822 | |
6ba495e9 | 823 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
824 | |
825 | inode = page->mapping->host; | |
826 | sb = inode->i_sb; | |
8df9675f | 827 | ngroups = ext4_get_groups_count(sb); |
93407472 | 828 | blocksize = i_blocksize(inode); |
09cbfeaf | 829 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d AT |
830 | |
831 | groups_per_page = blocks_per_page >> 1; | |
832 | if (groups_per_page == 0) | |
833 | groups_per_page = 1; | |
834 | ||
835 | /* allocate buffer_heads to read bitmaps */ | |
836 | if (groups_per_page > 1) { | |
c9de560d | 837 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 838 | bh = kzalloc(i, gfp); |
813e5727 TT |
839 | if (bh == NULL) { |
840 | err = -ENOMEM; | |
c9de560d | 841 | goto out; |
813e5727 | 842 | } |
c9de560d AT |
843 | } else |
844 | bh = &bhs; | |
845 | ||
846 | first_group = page->index * blocks_per_page / 2; | |
847 | ||
848 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
849 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
850 | if (group >= ngroups) | |
c9de560d AT |
851 | break; |
852 | ||
813e5727 | 853 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
854 | /* |
855 | * If page is uptodate then we came here after online resize | |
856 | * which added some new uninitialized group info structs, so | |
857 | * we must skip all initialized uptodate buddies on the page, | |
858 | * which may be currently in use by an allocating task. | |
859 | */ | |
860 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
861 | bh[i] = NULL; | |
862 | continue; | |
863 | } | |
9008a58e DW |
864 | bh[i] = ext4_read_block_bitmap_nowait(sb, group); |
865 | if (IS_ERR(bh[i])) { | |
866 | err = PTR_ERR(bh[i]); | |
867 | bh[i] = NULL; | |
c9de560d | 868 | goto out; |
2ccb5fb9 | 869 | } |
813e5727 | 870 | mb_debug(1, "read bitmap for group %u\n", group); |
c9de560d AT |
871 | } |
872 | ||
873 | /* wait for I/O completion */ | |
813e5727 | 874 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
875 | int err2; |
876 | ||
877 | if (!bh[i]) | |
878 | continue; | |
879 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
880 | if (!err) | |
881 | err = err2; | |
813e5727 | 882 | } |
c9de560d AT |
883 | |
884 | first_block = page->index * blocks_per_page; | |
885 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 886 | group = (first_block + i) >> 1; |
8df9675f | 887 | if (group >= ngroups) |
c9de560d AT |
888 | break; |
889 | ||
9b8b7d35 AG |
890 | if (!bh[group - first_group]) |
891 | /* skip initialized uptodate buddy */ | |
892 | continue; | |
893 | ||
bbdc322f LC |
894 | if (!buffer_verified(bh[group - first_group])) |
895 | /* Skip faulty bitmaps */ | |
896 | continue; | |
897 | err = 0; | |
898 | ||
c9de560d AT |
899 | /* |
900 | * data carry information regarding this | |
901 | * particular group in the format specified | |
902 | * above | |
903 | * | |
904 | */ | |
905 | data = page_address(page) + (i * blocksize); | |
906 | bitmap = bh[group - first_group]->b_data; | |
907 | ||
908 | /* | |
909 | * We place the buddy block and bitmap block | |
910 | * close together | |
911 | */ | |
912 | if ((first_block + i) & 1) { | |
913 | /* this is block of buddy */ | |
914 | BUG_ON(incore == NULL); | |
6ba495e9 | 915 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 916 | group, page->index, i * blocksize); |
f307333e | 917 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
918 | grinfo = ext4_get_group_info(sb, group); |
919 | grinfo->bb_fragments = 0; | |
920 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
921 | sizeof(*grinfo->bb_counters) * |
922 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
923 | /* |
924 | * incore got set to the group block bitmap below | |
925 | */ | |
7a2fcbf7 | 926 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
927 | /* init the buddy */ |
928 | memset(data, 0xff, blocksize); | |
c9de560d | 929 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 930 | ext4_unlock_group(sb, group); |
c9de560d AT |
931 | incore = NULL; |
932 | } else { | |
933 | /* this is block of bitmap */ | |
934 | BUG_ON(incore != NULL); | |
6ba495e9 | 935 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 936 | group, page->index, i * blocksize); |
f307333e | 937 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
938 | |
939 | /* see comments in ext4_mb_put_pa() */ | |
940 | ext4_lock_group(sb, group); | |
941 | memcpy(data, bitmap, blocksize); | |
942 | ||
943 | /* mark all preallocated blks used in in-core bitmap */ | |
944 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 945 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
946 | ext4_unlock_group(sb, group); |
947 | ||
948 | /* set incore so that the buddy information can be | |
949 | * generated using this | |
950 | */ | |
951 | incore = data; | |
952 | } | |
953 | } | |
954 | SetPageUptodate(page); | |
955 | ||
956 | out: | |
957 | if (bh) { | |
9b8b7d35 | 958 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
959 | brelse(bh[i]); |
960 | if (bh != &bhs) | |
961 | kfree(bh); | |
962 | } | |
963 | return err; | |
964 | } | |
965 | ||
eee4adc7 | 966 | /* |
2de8807b AG |
967 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
968 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
969 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
970 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 971 | */ |
2de8807b | 972 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 973 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 974 | { |
2de8807b AG |
975 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
976 | int block, pnum, poff; | |
eee4adc7 | 977 | int blocks_per_page; |
2de8807b AG |
978 | struct page *page; |
979 | ||
980 | e4b->bd_buddy_page = NULL; | |
981 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 982 | |
09cbfeaf | 983 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
984 | /* |
985 | * the buddy cache inode stores the block bitmap | |
986 | * and buddy information in consecutive blocks. | |
987 | * So for each group we need two blocks. | |
988 | */ | |
989 | block = group * 2; | |
990 | pnum = block / blocks_per_page; | |
2de8807b | 991 | poff = block % blocks_per_page; |
adb7ef60 | 992 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 993 | if (!page) |
c57ab39b | 994 | return -ENOMEM; |
2de8807b AG |
995 | BUG_ON(page->mapping != inode->i_mapping); |
996 | e4b->bd_bitmap_page = page; | |
997 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
998 | ||
999 | if (blocks_per_page >= 2) { | |
1000 | /* buddy and bitmap are on the same page */ | |
1001 | return 0; | |
eee4adc7 | 1002 | } |
2de8807b AG |
1003 | |
1004 | block++; | |
1005 | pnum = block / blocks_per_page; | |
adb7ef60 | 1006 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1007 | if (!page) |
c57ab39b | 1008 | return -ENOMEM; |
2de8807b AG |
1009 | BUG_ON(page->mapping != inode->i_mapping); |
1010 | e4b->bd_buddy_page = page; | |
1011 | return 0; | |
eee4adc7 ES |
1012 | } |
1013 | ||
2de8807b | 1014 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1015 | { |
2de8807b AG |
1016 | if (e4b->bd_bitmap_page) { |
1017 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1018 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1019 | } |
1020 | if (e4b->bd_buddy_page) { | |
1021 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1022 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1023 | } |
eee4adc7 ES |
1024 | } |
1025 | ||
8a57d9d6 CW |
1026 | /* |
1027 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1028 | * block group lock of all groups for this page; do not hold the BG lock when | |
1029 | * calling this routine! | |
1030 | */ | |
b6a758ec | 1031 | static noinline_for_stack |
adb7ef60 | 1032 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1033 | { |
1034 | ||
b6a758ec | 1035 | struct ext4_group_info *this_grp; |
2de8807b AG |
1036 | struct ext4_buddy e4b; |
1037 | struct page *page; | |
1038 | int ret = 0; | |
b6a758ec | 1039 | |
b10a44c3 | 1040 | might_sleep(); |
b6a758ec | 1041 | mb_debug(1, "init group %u\n", group); |
b6a758ec AK |
1042 | this_grp = ext4_get_group_info(sb, group); |
1043 | /* | |
08c3a813 AK |
1044 | * This ensures that we don't reinit the buddy cache |
1045 | * page which map to the group from which we are already | |
1046 | * allocating. If we are looking at the buddy cache we would | |
1047 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1048 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1049 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1050 | * page accessed. | |
b6a758ec | 1051 | */ |
adb7ef60 | 1052 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1053 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1054 | /* |
1055 | * somebody initialized the group | |
1056 | * return without doing anything | |
1057 | */ | |
b6a758ec AK |
1058 | goto err; |
1059 | } | |
2de8807b AG |
1060 | |
1061 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1062 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1063 | if (ret) |
1064 | goto err; | |
1065 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1066 | ret = -EIO; |
1067 | goto err; | |
1068 | } | |
b6a758ec | 1069 | |
2de8807b | 1070 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1071 | /* |
1072 | * If both the bitmap and buddy are in | |
1073 | * the same page we don't need to force | |
1074 | * init the buddy | |
1075 | */ | |
2de8807b AG |
1076 | ret = 0; |
1077 | goto err; | |
b6a758ec | 1078 | } |
2de8807b AG |
1079 | /* init buddy cache */ |
1080 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1081 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1082 | if (ret) |
1083 | goto err; | |
1084 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1085 | ret = -EIO; |
1086 | goto err; | |
1087 | } | |
b6a758ec | 1088 | err: |
2de8807b | 1089 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1090 | return ret; |
1091 | } | |
1092 | ||
8a57d9d6 CW |
1093 | /* |
1094 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1095 | * block group lock of all groups for this page; do not hold the BG lock when | |
1096 | * calling this routine! | |
1097 | */ | |
4ddfef7b | 1098 | static noinline_for_stack int |
adb7ef60 KK |
1099 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1100 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1101 | { |
c9de560d AT |
1102 | int blocks_per_page; |
1103 | int block; | |
1104 | int pnum; | |
1105 | int poff; | |
1106 | struct page *page; | |
fdf6c7a7 | 1107 | int ret; |
920313a7 AK |
1108 | struct ext4_group_info *grp; |
1109 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1110 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1111 | |
b10a44c3 | 1112 | might_sleep(); |
6ba495e9 | 1113 | mb_debug(1, "load group %u\n", group); |
c9de560d | 1114 | |
09cbfeaf | 1115 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1116 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1117 | |
1118 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1119 | e4b->bd_info = grp; |
c9de560d AT |
1120 | e4b->bd_sb = sb; |
1121 | e4b->bd_group = group; | |
1122 | e4b->bd_buddy_page = NULL; | |
1123 | e4b->bd_bitmap_page = NULL; | |
1124 | ||
f41c0750 | 1125 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1126 | /* |
1127 | * we need full data about the group | |
1128 | * to make a good selection | |
1129 | */ | |
adb7ef60 | 1130 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1131 | if (ret) |
1132 | return ret; | |
f41c0750 AK |
1133 | } |
1134 | ||
c9de560d AT |
1135 | /* |
1136 | * the buddy cache inode stores the block bitmap | |
1137 | * and buddy information in consecutive blocks. | |
1138 | * So for each group we need two blocks. | |
1139 | */ | |
1140 | block = group * 2; | |
1141 | pnum = block / blocks_per_page; | |
1142 | poff = block % blocks_per_page; | |
1143 | ||
1144 | /* we could use find_or_create_page(), but it locks page | |
1145 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1146 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1147 | if (page == NULL || !PageUptodate(page)) { |
1148 | if (page) | |
920313a7 AK |
1149 | /* |
1150 | * drop the page reference and try | |
1151 | * to get the page with lock. If we | |
1152 | * are not uptodate that implies | |
1153 | * somebody just created the page but | |
1154 | * is yet to initialize the same. So | |
1155 | * wait for it to initialize. | |
1156 | */ | |
09cbfeaf | 1157 | put_page(page); |
adb7ef60 | 1158 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1159 | if (page) { |
1160 | BUG_ON(page->mapping != inode->i_mapping); | |
1161 | if (!PageUptodate(page)) { | |
adb7ef60 | 1162 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1163 | if (ret) { |
1164 | unlock_page(page); | |
1165 | goto err; | |
1166 | } | |
c9de560d AT |
1167 | mb_cmp_bitmaps(e4b, page_address(page) + |
1168 | (poff * sb->s_blocksize)); | |
1169 | } | |
1170 | unlock_page(page); | |
1171 | } | |
1172 | } | |
c57ab39b YL |
1173 | if (page == NULL) { |
1174 | ret = -ENOMEM; | |
1175 | goto err; | |
1176 | } | |
1177 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1178 | ret = -EIO; |
c9de560d | 1179 | goto err; |
fdf6c7a7 | 1180 | } |
2457aec6 MG |
1181 | |
1182 | /* Pages marked accessed already */ | |
c9de560d AT |
1183 | e4b->bd_bitmap_page = page; |
1184 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1185 | |
1186 | block++; | |
1187 | pnum = block / blocks_per_page; | |
1188 | poff = block % blocks_per_page; | |
1189 | ||
2457aec6 | 1190 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1191 | if (page == NULL || !PageUptodate(page)) { |
1192 | if (page) | |
09cbfeaf | 1193 | put_page(page); |
adb7ef60 | 1194 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1195 | if (page) { |
1196 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1197 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1198 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1199 | gfp); | |
fdf6c7a7 SF |
1200 | if (ret) { |
1201 | unlock_page(page); | |
1202 | goto err; | |
1203 | } | |
1204 | } | |
c9de560d AT |
1205 | unlock_page(page); |
1206 | } | |
1207 | } | |
c57ab39b YL |
1208 | if (page == NULL) { |
1209 | ret = -ENOMEM; | |
1210 | goto err; | |
1211 | } | |
1212 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1213 | ret = -EIO; |
c9de560d | 1214 | goto err; |
fdf6c7a7 | 1215 | } |
2457aec6 MG |
1216 | |
1217 | /* Pages marked accessed already */ | |
c9de560d AT |
1218 | e4b->bd_buddy_page = page; |
1219 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1220 | |
1221 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1222 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1223 | ||
1224 | return 0; | |
1225 | ||
1226 | err: | |
26626f11 | 1227 | if (page) |
09cbfeaf | 1228 | put_page(page); |
c9de560d | 1229 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1230 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1231 | if (e4b->bd_buddy_page) |
09cbfeaf | 1232 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1233 | e4b->bd_buddy = NULL; |
1234 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1235 | return ret; |
c9de560d AT |
1236 | } |
1237 | ||
adb7ef60 KK |
1238 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1239 | struct ext4_buddy *e4b) | |
1240 | { | |
1241 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1242 | } | |
1243 | ||
e39e07fd | 1244 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1245 | { |
1246 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1247 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1248 | if (e4b->bd_buddy_page) |
09cbfeaf | 1249 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1250 | } |
1251 | ||
1252 | ||
1253 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1254 | { | |
1255 | int order = 1; | |
b5cb316c | 1256 | int bb_incr = 1 << (e4b->bd_blkbits - 1); |
c9de560d AT |
1257 | void *bb; |
1258 | ||
c5e8f3f3 | 1259 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1260 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1261 | ||
c5e8f3f3 | 1262 | bb = e4b->bd_buddy; |
c9de560d AT |
1263 | while (order <= e4b->bd_blkbits + 1) { |
1264 | block = block >> 1; | |
1265 | if (!mb_test_bit(block, bb)) { | |
1266 | /* this block is part of buddy of order 'order' */ | |
1267 | return order; | |
1268 | } | |
b5cb316c NS |
1269 | bb += bb_incr; |
1270 | bb_incr >>= 1; | |
c9de560d AT |
1271 | order++; |
1272 | } | |
1273 | return 0; | |
1274 | } | |
1275 | ||
955ce5f5 | 1276 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1277 | { |
1278 | __u32 *addr; | |
1279 | ||
1280 | len = cur + len; | |
1281 | while (cur < len) { | |
1282 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1283 | /* fast path: clear whole word at once */ | |
1284 | addr = bm + (cur >> 3); | |
1285 | *addr = 0; | |
1286 | cur += 32; | |
1287 | continue; | |
1288 | } | |
955ce5f5 | 1289 | mb_clear_bit(cur, bm); |
c9de560d AT |
1290 | cur++; |
1291 | } | |
1292 | } | |
1293 | ||
eabe0444 AS |
1294 | /* clear bits in given range |
1295 | * will return first found zero bit if any, -1 otherwise | |
1296 | */ | |
1297 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1298 | { | |
1299 | __u32 *addr; | |
1300 | int zero_bit = -1; | |
1301 | ||
1302 | len = cur + len; | |
1303 | while (cur < len) { | |
1304 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1305 | /* fast path: clear whole word at once */ | |
1306 | addr = bm + (cur >> 3); | |
1307 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1308 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1309 | *addr = 0; | |
1310 | cur += 32; | |
1311 | continue; | |
1312 | } | |
1313 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1314 | zero_bit = cur; | |
1315 | cur++; | |
1316 | } | |
1317 | ||
1318 | return zero_bit; | |
1319 | } | |
1320 | ||
c3e94d1d | 1321 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1322 | { |
1323 | __u32 *addr; | |
1324 | ||
1325 | len = cur + len; | |
1326 | while (cur < len) { | |
1327 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1328 | /* fast path: set whole word at once */ | |
1329 | addr = bm + (cur >> 3); | |
1330 | *addr = 0xffffffff; | |
1331 | cur += 32; | |
1332 | continue; | |
1333 | } | |
955ce5f5 | 1334 | mb_set_bit(cur, bm); |
c9de560d AT |
1335 | cur++; |
1336 | } | |
1337 | } | |
1338 | ||
eabe0444 AS |
1339 | /* |
1340 | * _________________________________________________________________ */ | |
1341 | ||
1342 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) | |
1343 | { | |
1344 | if (mb_test_bit(*bit + side, bitmap)) { | |
1345 | mb_clear_bit(*bit, bitmap); | |
1346 | (*bit) -= side; | |
1347 | return 1; | |
1348 | } | |
1349 | else { | |
1350 | (*bit) += side; | |
1351 | mb_set_bit(*bit, bitmap); | |
1352 | return -1; | |
1353 | } | |
1354 | } | |
1355 | ||
1356 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1357 | { | |
1358 | int max; | |
1359 | int order = 1; | |
1360 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1361 | ||
1362 | while (buddy) { | |
1363 | void *buddy2; | |
1364 | ||
1365 | /* Bits in range [first; last] are known to be set since | |
1366 | * corresponding blocks were allocated. Bits in range | |
1367 | * (first; last) will stay set because they form buddies on | |
1368 | * upper layer. We just deal with borders if they don't | |
1369 | * align with upper layer and then go up. | |
1370 | * Releasing entire group is all about clearing | |
1371 | * single bit of highest order buddy. | |
1372 | */ | |
1373 | ||
1374 | /* Example: | |
1375 | * --------------------------------- | |
1376 | * | 1 | 1 | 1 | 1 | | |
1377 | * --------------------------------- | |
1378 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1379 | * --------------------------------- | |
1380 | * 0 1 2 3 4 5 6 7 | |
1381 | * \_____________________/ | |
1382 | * | |
1383 | * Neither [1] nor [6] is aligned to above layer. | |
1384 | * Left neighbour [0] is free, so mark it busy, | |
1385 | * decrease bb_counters and extend range to | |
1386 | * [0; 6] | |
1387 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1388 | * mark [6] free, increase bb_counters and shrink range to | |
1389 | * [0; 5]. | |
1390 | * Then shift range to [0; 2], go up and do the same. | |
1391 | */ | |
1392 | ||
1393 | ||
1394 | if (first & 1) | |
1395 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1396 | if (!(last & 1)) | |
1397 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1398 | if (first > last) | |
1399 | break; | |
1400 | order++; | |
1401 | ||
1402 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1403 | mb_clear_bits(buddy, first, last - first + 1); | |
1404 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1405 | break; | |
1406 | } | |
1407 | first >>= 1; | |
1408 | last >>= 1; | |
1409 | buddy = buddy2; | |
1410 | } | |
1411 | } | |
1412 | ||
7e5a8cdd | 1413 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1414 | int first, int count) |
c9de560d | 1415 | { |
eabe0444 AS |
1416 | int left_is_free = 0; |
1417 | int right_is_free = 0; | |
1418 | int block; | |
1419 | int last = first + count - 1; | |
c9de560d AT |
1420 | struct super_block *sb = e4b->bd_sb; |
1421 | ||
c99d1e6e TT |
1422 | if (WARN_ON(count == 0)) |
1423 | return; | |
eabe0444 | 1424 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1425 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1426 | /* Don't bother if the block group is corrupt. */ |
1427 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1428 | return; | |
1429 | ||
c9de560d AT |
1430 | mb_check_buddy(e4b); |
1431 | mb_free_blocks_double(inode, e4b, first, count); | |
1432 | ||
1433 | e4b->bd_info->bb_free += count; | |
1434 | if (first < e4b->bd_info->bb_first_free) | |
1435 | e4b->bd_info->bb_first_free = first; | |
1436 | ||
eabe0444 AS |
1437 | /* access memory sequentially: check left neighbour, |
1438 | * clear range and then check right neighbour | |
1439 | */ | |
c9de560d | 1440 | if (first != 0) |
eabe0444 AS |
1441 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1442 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1443 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1444 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1445 | ||
1446 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1447 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1448 | ext4_fsblk_t blocknr; |
1449 | ||
1450 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1451 | blocknr += EXT4_C2B(sbi, block); |
eabe0444 AS |
1452 | ext4_grp_locked_error(sb, e4b->bd_group, |
1453 | inode ? inode->i_ino : 0, | |
1454 | blocknr, | |
1455 | "freeing already freed block " | |
163a203d DW |
1456 | "(bit %u); block bitmap corrupt.", |
1457 | block); | |
db79e6d1 WS |
1458 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1459 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
eabe0444 AS |
1460 | mb_regenerate_buddy(e4b); |
1461 | goto done; | |
1462 | } | |
1463 | ||
1464 | /* let's maintain fragments counter */ | |
1465 | if (left_is_free && right_is_free) | |
c9de560d | 1466 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1467 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1468 | e4b->bd_info->bb_fragments++; |
1469 | ||
eabe0444 AS |
1470 | /* buddy[0] == bd_bitmap is a special case, so handle |
1471 | * it right away and let mb_buddy_mark_free stay free of | |
1472 | * zero order checks. | |
1473 | * Check if neighbours are to be coaleasced, | |
1474 | * adjust bitmap bb_counters and borders appropriately. | |
1475 | */ | |
1476 | if (first & 1) { | |
1477 | first += !left_is_free; | |
1478 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1479 | } | |
1480 | if (!(last & 1)) { | |
1481 | last -= !right_is_free; | |
1482 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1483 | } | |
c9de560d | 1484 | |
eabe0444 AS |
1485 | if (first <= last) |
1486 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1487 | |
eabe0444 | 1488 | done: |
8a57d9d6 | 1489 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1490 | mb_check_buddy(e4b); |
c9de560d AT |
1491 | } |
1492 | ||
15c006a2 | 1493 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1494 | int needed, struct ext4_free_extent *ex) |
1495 | { | |
1496 | int next = block; | |
15c006a2 | 1497 | int max, order; |
c9de560d AT |
1498 | void *buddy; |
1499 | ||
bc8e6740 | 1500 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1501 | BUG_ON(ex == NULL); |
1502 | ||
15c006a2 | 1503 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1504 | BUG_ON(buddy == NULL); |
1505 | BUG_ON(block >= max); | |
1506 | if (mb_test_bit(block, buddy)) { | |
1507 | ex->fe_len = 0; | |
1508 | ex->fe_start = 0; | |
1509 | ex->fe_group = 0; | |
1510 | return 0; | |
1511 | } | |
1512 | ||
15c006a2 RD |
1513 | /* find actual order */ |
1514 | order = mb_find_order_for_block(e4b, block); | |
1515 | block = block >> order; | |
c9de560d AT |
1516 | |
1517 | ex->fe_len = 1 << order; | |
1518 | ex->fe_start = block << order; | |
1519 | ex->fe_group = e4b->bd_group; | |
1520 | ||
1521 | /* calc difference from given start */ | |
1522 | next = next - ex->fe_start; | |
1523 | ex->fe_len -= next; | |
1524 | ex->fe_start += next; | |
1525 | ||
1526 | while (needed > ex->fe_len && | |
d8ec0c39 | 1527 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1528 | |
1529 | if (block + 1 >= max) | |
1530 | break; | |
1531 | ||
1532 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1533 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1534 | break; |
1535 | ||
b051d8dc | 1536 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1537 | |
c9de560d AT |
1538 | block = next >> order; |
1539 | ex->fe_len += 1 << order; | |
1540 | } | |
1541 | ||
31562b95 | 1542 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1543 | /* Should never happen! (but apparently sometimes does?!?) */ |
1544 | WARN_ON(1); | |
1545 | ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent " | |
1546 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1547 | block, order, needed, ex->fe_group, ex->fe_start, | |
1548 | ex->fe_len, ex->fe_logical); | |
1549 | ex->fe_len = 0; | |
1550 | ex->fe_start = 0; | |
1551 | ex->fe_group = 0; | |
1552 | } | |
c9de560d AT |
1553 | return ex->fe_len; |
1554 | } | |
1555 | ||
1556 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1557 | { | |
1558 | int ord; | |
1559 | int mlen = 0; | |
1560 | int max = 0; | |
1561 | int cur; | |
1562 | int start = ex->fe_start; | |
1563 | int len = ex->fe_len; | |
1564 | unsigned ret = 0; | |
1565 | int len0 = len; | |
1566 | void *buddy; | |
1567 | ||
1568 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1569 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1570 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1571 | mb_check_buddy(e4b); |
1572 | mb_mark_used_double(e4b, start, len); | |
1573 | ||
1574 | e4b->bd_info->bb_free -= len; | |
1575 | if (e4b->bd_info->bb_first_free == start) | |
1576 | e4b->bd_info->bb_first_free += len; | |
1577 | ||
1578 | /* let's maintain fragments counter */ | |
1579 | if (start != 0) | |
c5e8f3f3 | 1580 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1581 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1582 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1583 | if (mlen && max) |
1584 | e4b->bd_info->bb_fragments++; | |
1585 | else if (!mlen && !max) | |
1586 | e4b->bd_info->bb_fragments--; | |
1587 | ||
1588 | /* let's maintain buddy itself */ | |
1589 | while (len) { | |
1590 | ord = mb_find_order_for_block(e4b, start); | |
1591 | ||
1592 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1593 | /* the whole chunk may be allocated at once! */ | |
1594 | mlen = 1 << ord; | |
1595 | buddy = mb_find_buddy(e4b, ord, &max); | |
1596 | BUG_ON((start >> ord) >= max); | |
1597 | mb_set_bit(start >> ord, buddy); | |
1598 | e4b->bd_info->bb_counters[ord]--; | |
1599 | start += mlen; | |
1600 | len -= mlen; | |
1601 | BUG_ON(len < 0); | |
1602 | continue; | |
1603 | } | |
1604 | ||
1605 | /* store for history */ | |
1606 | if (ret == 0) | |
1607 | ret = len | (ord << 16); | |
1608 | ||
1609 | /* we have to split large buddy */ | |
1610 | BUG_ON(ord <= 0); | |
1611 | buddy = mb_find_buddy(e4b, ord, &max); | |
1612 | mb_set_bit(start >> ord, buddy); | |
1613 | e4b->bd_info->bb_counters[ord]--; | |
1614 | ||
1615 | ord--; | |
1616 | cur = (start >> ord) & ~1U; | |
1617 | buddy = mb_find_buddy(e4b, ord, &max); | |
1618 | mb_clear_bit(cur, buddy); | |
1619 | mb_clear_bit(cur + 1, buddy); | |
1620 | e4b->bd_info->bb_counters[ord]++; | |
1621 | e4b->bd_info->bb_counters[ord]++; | |
1622 | } | |
8a57d9d6 | 1623 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1624 | |
c5e8f3f3 | 1625 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1626 | mb_check_buddy(e4b); |
1627 | ||
1628 | return ret; | |
1629 | } | |
1630 | ||
1631 | /* | |
1632 | * Must be called under group lock! | |
1633 | */ | |
1634 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1635 | struct ext4_buddy *e4b) | |
1636 | { | |
1637 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1638 | int ret; | |
1639 | ||
1640 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1641 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1642 | ||
1643 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1644 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1645 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1646 | ||
1647 | /* preallocation can change ac_b_ex, thus we store actually | |
1648 | * allocated blocks for history */ | |
1649 | ac->ac_f_ex = ac->ac_b_ex; | |
1650 | ||
1651 | ac->ac_status = AC_STATUS_FOUND; | |
1652 | ac->ac_tail = ret & 0xffff; | |
1653 | ac->ac_buddy = ret >> 16; | |
1654 | ||
c3a326a6 AK |
1655 | /* |
1656 | * take the page reference. We want the page to be pinned | |
1657 | * so that we don't get a ext4_mb_init_cache_call for this | |
1658 | * group until we update the bitmap. That would mean we | |
1659 | * double allocate blocks. The reference is dropped | |
1660 | * in ext4_mb_release_context | |
1661 | */ | |
c9de560d AT |
1662 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1663 | get_page(ac->ac_bitmap_page); | |
1664 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1665 | get_page(ac->ac_buddy_page); | |
c9de560d | 1666 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1667 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1668 | spin_lock(&sbi->s_md_lock); |
1669 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1670 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1671 | spin_unlock(&sbi->s_md_lock); | |
1672 | } | |
1673 | } | |
1674 | ||
1675 | /* | |
1676 | * regular allocator, for general purposes allocation | |
1677 | */ | |
1678 | ||
1679 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1680 | struct ext4_buddy *e4b, | |
1681 | int finish_group) | |
1682 | { | |
1683 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1684 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1685 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1686 | struct ext4_free_extent ex; | |
1687 | int max; | |
1688 | ||
032115fc AK |
1689 | if (ac->ac_status == AC_STATUS_FOUND) |
1690 | return; | |
c9de560d AT |
1691 | /* |
1692 | * We don't want to scan for a whole year | |
1693 | */ | |
1694 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1695 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1696 | ac->ac_status = AC_STATUS_BREAK; | |
1697 | return; | |
1698 | } | |
1699 | ||
1700 | /* | |
1701 | * Haven't found good chunk so far, let's continue | |
1702 | */ | |
1703 | if (bex->fe_len < gex->fe_len) | |
1704 | return; | |
1705 | ||
1706 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1707 | && bex->fe_group == e4b->bd_group) { | |
1708 | /* recheck chunk's availability - we don't know | |
1709 | * when it was found (within this lock-unlock | |
1710 | * period or not) */ | |
15c006a2 | 1711 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1712 | if (max >= gex->fe_len) { |
1713 | ext4_mb_use_best_found(ac, e4b); | |
1714 | return; | |
1715 | } | |
1716 | } | |
1717 | } | |
1718 | ||
1719 | /* | |
1720 | * The routine checks whether found extent is good enough. If it is, | |
1721 | * then the extent gets marked used and flag is set to the context | |
1722 | * to stop scanning. Otherwise, the extent is compared with the | |
1723 | * previous found extent and if new one is better, then it's stored | |
1724 | * in the context. Later, the best found extent will be used, if | |
1725 | * mballoc can't find good enough extent. | |
1726 | * | |
1727 | * FIXME: real allocation policy is to be designed yet! | |
1728 | */ | |
1729 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1730 | struct ext4_free_extent *ex, | |
1731 | struct ext4_buddy *e4b) | |
1732 | { | |
1733 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1734 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1735 | ||
1736 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1737 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1738 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1739 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1740 | ||
1741 | ac->ac_found++; | |
1742 | ||
1743 | /* | |
1744 | * The special case - take what you catch first | |
1745 | */ | |
1746 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1747 | *bex = *ex; | |
1748 | ext4_mb_use_best_found(ac, e4b); | |
1749 | return; | |
1750 | } | |
1751 | ||
1752 | /* | |
1753 | * Let's check whether the chuck is good enough | |
1754 | */ | |
1755 | if (ex->fe_len == gex->fe_len) { | |
1756 | *bex = *ex; | |
1757 | ext4_mb_use_best_found(ac, e4b); | |
1758 | return; | |
1759 | } | |
1760 | ||
1761 | /* | |
1762 | * If this is first found extent, just store it in the context | |
1763 | */ | |
1764 | if (bex->fe_len == 0) { | |
1765 | *bex = *ex; | |
1766 | return; | |
1767 | } | |
1768 | ||
1769 | /* | |
1770 | * If new found extent is better, store it in the context | |
1771 | */ | |
1772 | if (bex->fe_len < gex->fe_len) { | |
1773 | /* if the request isn't satisfied, any found extent | |
1774 | * larger than previous best one is better */ | |
1775 | if (ex->fe_len > bex->fe_len) | |
1776 | *bex = *ex; | |
1777 | } else if (ex->fe_len > gex->fe_len) { | |
1778 | /* if the request is satisfied, then we try to find | |
1779 | * an extent that still satisfy the request, but is | |
1780 | * smaller than previous one */ | |
1781 | if (ex->fe_len < bex->fe_len) | |
1782 | *bex = *ex; | |
1783 | } | |
1784 | ||
1785 | ext4_mb_check_limits(ac, e4b, 0); | |
1786 | } | |
1787 | ||
089ceecc ES |
1788 | static noinline_for_stack |
1789 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1790 | struct ext4_buddy *e4b) |
1791 | { | |
1792 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1793 | ext4_group_t group = ex.fe_group; | |
1794 | int max; | |
1795 | int err; | |
1796 | ||
1797 | BUG_ON(ex.fe_len <= 0); | |
1798 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1799 | if (err) | |
1800 | return err; | |
1801 | ||
1802 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1803 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1804 | |
1805 | if (max > 0) { | |
1806 | ac->ac_b_ex = ex; | |
1807 | ext4_mb_use_best_found(ac, e4b); | |
1808 | } | |
1809 | ||
1810 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1811 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1812 | |
1813 | return 0; | |
1814 | } | |
1815 | ||
089ceecc ES |
1816 | static noinline_for_stack |
1817 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1818 | struct ext4_buddy *e4b) |
1819 | { | |
1820 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1821 | int max; | |
1822 | int err; | |
1823 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1824 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1825 | struct ext4_free_extent ex; |
1826 | ||
1827 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1828 | return 0; | |
838cd0cf YY |
1829 | if (grp->bb_free == 0) |
1830 | return 0; | |
c9de560d AT |
1831 | |
1832 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1833 | if (err) | |
1834 | return err; | |
1835 | ||
163a203d DW |
1836 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
1837 | ext4_mb_unload_buddy(e4b); | |
1838 | return 0; | |
1839 | } | |
1840 | ||
c9de560d | 1841 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 1842 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 1843 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 1844 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
1845 | |
1846 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1847 | ext4_fsblk_t start; | |
1848 | ||
5661bd68 AM |
1849 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1850 | ex.fe_start; | |
c9de560d AT |
1851 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1852 | if (do_div(start, sbi->s_stripe) == 0) { | |
1853 | ac->ac_found++; | |
1854 | ac->ac_b_ex = ex; | |
1855 | ext4_mb_use_best_found(ac, e4b); | |
1856 | } | |
1857 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1858 | BUG_ON(ex.fe_len <= 0); | |
1859 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1860 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1861 | ac->ac_found++; | |
1862 | ac->ac_b_ex = ex; | |
1863 | ext4_mb_use_best_found(ac, e4b); | |
1864 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1865 | /* Sometimes, caller may want to merge even small | |
1866 | * number of blocks to an existing extent */ | |
1867 | BUG_ON(ex.fe_len <= 0); | |
1868 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1869 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1870 | ac->ac_found++; | |
1871 | ac->ac_b_ex = ex; | |
1872 | ext4_mb_use_best_found(ac, e4b); | |
1873 | } | |
1874 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1875 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1876 | |
1877 | return 0; | |
1878 | } | |
1879 | ||
1880 | /* | |
1881 | * The routine scans buddy structures (not bitmap!) from given order | |
1882 | * to max order and tries to find big enough chunk to satisfy the req | |
1883 | */ | |
089ceecc ES |
1884 | static noinline_for_stack |
1885 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1886 | struct ext4_buddy *e4b) |
1887 | { | |
1888 | struct super_block *sb = ac->ac_sb; | |
1889 | struct ext4_group_info *grp = e4b->bd_info; | |
1890 | void *buddy; | |
1891 | int i; | |
1892 | int k; | |
1893 | int max; | |
1894 | ||
1895 | BUG_ON(ac->ac_2order <= 0); | |
1896 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1897 | if (grp->bb_counters[i] == 0) | |
1898 | continue; | |
1899 | ||
1900 | buddy = mb_find_buddy(e4b, i, &max); | |
1901 | BUG_ON(buddy == NULL); | |
1902 | ||
ffad0a44 | 1903 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
1904 | if (k >= max) { |
1905 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
1906 | "%d free clusters of order %d. But found 0", | |
1907 | grp->bb_counters[i], i); | |
1908 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
1909 | e4b->bd_group, | |
1910 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
1911 | break; | |
1912 | } | |
c9de560d AT |
1913 | ac->ac_found++; |
1914 | ||
1915 | ac->ac_b_ex.fe_len = 1 << i; | |
1916 | ac->ac_b_ex.fe_start = k << i; | |
1917 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1918 | ||
1919 | ext4_mb_use_best_found(ac, e4b); | |
1920 | ||
1921 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1922 | ||
1923 | if (EXT4_SB(sb)->s_mb_stats) | |
1924 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1925 | ||
1926 | break; | |
1927 | } | |
1928 | } | |
1929 | ||
1930 | /* | |
1931 | * The routine scans the group and measures all found extents. | |
1932 | * In order to optimize scanning, caller must pass number of | |
1933 | * free blocks in the group, so the routine can know upper limit. | |
1934 | */ | |
089ceecc ES |
1935 | static noinline_for_stack |
1936 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1937 | struct ext4_buddy *e4b) |
1938 | { | |
1939 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 1940 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1941 | struct ext4_free_extent ex; |
1942 | int i; | |
1943 | int free; | |
1944 | ||
1945 | free = e4b->bd_info->bb_free; | |
1946 | BUG_ON(free <= 0); | |
1947 | ||
1948 | i = e4b->bd_info->bb_first_free; | |
1949 | ||
1950 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1951 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1952 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1953 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1954 | /* |
e56eb659 | 1955 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1956 | * free blocks even though group info says we |
1957 | * we have free blocks | |
1958 | */ | |
e29136f8 | 1959 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1960 | "%d free clusters as per " |
fde4d95a | 1961 | "group info. But bitmap says 0", |
26346ff6 | 1962 | free); |
736dedbb WS |
1963 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1964 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
1965 | break; |
1966 | } | |
1967 | ||
15c006a2 | 1968 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
c9de560d | 1969 | BUG_ON(ex.fe_len <= 0); |
26346ff6 | 1970 | if (free < ex.fe_len) { |
e29136f8 | 1971 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1972 | "%d free clusters as per " |
fde4d95a | 1973 | "group info. But got %d blocks", |
26346ff6 | 1974 | free, ex.fe_len); |
736dedbb WS |
1975 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1976 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
1977 | /* |
1978 | * The number of free blocks differs. This mostly | |
1979 | * indicate that the bitmap is corrupt. So exit | |
1980 | * without claiming the space. | |
1981 | */ | |
1982 | break; | |
26346ff6 | 1983 | } |
ab0c00fc | 1984 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
1985 | ext4_mb_measure_extent(ac, &ex, e4b); |
1986 | ||
1987 | i += ex.fe_len; | |
1988 | free -= ex.fe_len; | |
1989 | } | |
1990 | ||
1991 | ext4_mb_check_limits(ac, e4b, 1); | |
1992 | } | |
1993 | ||
1994 | /* | |
1995 | * This is a special case for storages like raid5 | |
506bf2d8 | 1996 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 1997 | */ |
089ceecc ES |
1998 | static noinline_for_stack |
1999 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2000 | struct ext4_buddy *e4b) |
2001 | { | |
2002 | struct super_block *sb = ac->ac_sb; | |
2003 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2004 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2005 | struct ext4_free_extent ex; |
2006 | ext4_fsblk_t first_group_block; | |
2007 | ext4_fsblk_t a; | |
2008 | ext4_grpblk_t i; | |
2009 | int max; | |
2010 | ||
2011 | BUG_ON(sbi->s_stripe == 0); | |
2012 | ||
2013 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2014 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2015 | ||
c9de560d AT |
2016 | a = first_group_block + sbi->s_stripe - 1; |
2017 | do_div(a, sbi->s_stripe); | |
2018 | i = (a * sbi->s_stripe) - first_group_block; | |
2019 | ||
7137d7a4 | 2020 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2021 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2022 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2023 | if (max >= sbi->s_stripe) { |
2024 | ac->ac_found++; | |
ab0c00fc | 2025 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2026 | ac->ac_b_ex = ex; |
2027 | ext4_mb_use_best_found(ac, e4b); | |
2028 | break; | |
2029 | } | |
2030 | } | |
2031 | i += sbi->s_stripe; | |
2032 | } | |
2033 | } | |
2034 | ||
42ac1848 LC |
2035 | /* |
2036 | * This is now called BEFORE we load the buddy bitmap. | |
2037 | * Returns either 1 or 0 indicating that the group is either suitable | |
2038 | * for the allocation or not. In addition it can also return negative | |
2039 | * error code when something goes wrong. | |
2040 | */ | |
c9de560d AT |
2041 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
2042 | ext4_group_t group, int cr) | |
2043 | { | |
2044 | unsigned free, fragments; | |
a4912123 | 2045 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2046 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2047 | ||
2048 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2049 | |
01fc48e8 TT |
2050 | free = grp->bb_free; |
2051 | if (free == 0) | |
2052 | return 0; | |
2053 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2054 | return 0; | |
2055 | ||
163a203d DW |
2056 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
2057 | return 0; | |
2058 | ||
8a57d9d6 CW |
2059 | /* We only do this if the grp has never been initialized */ |
2060 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 2061 | int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS); |
8a57d9d6 | 2062 | if (ret) |
42ac1848 | 2063 | return ret; |
8a57d9d6 | 2064 | } |
c9de560d | 2065 | |
c9de560d | 2066 | fragments = grp->bb_fragments; |
c9de560d AT |
2067 | if (fragments == 0) |
2068 | return 0; | |
2069 | ||
2070 | switch (cr) { | |
2071 | case 0: | |
2072 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2073 | |
a4912123 TT |
2074 | /* Avoid using the first bg of a flexgroup for data files */ |
2075 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2076 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2077 | ((group % flex_size) == 0)) | |
2078 | return 0; | |
2079 | ||
40ae3487 TT |
2080 | if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || |
2081 | (free / fragments) >= ac->ac_g_ex.fe_len) | |
2082 | return 1; | |
2083 | ||
2084 | if (grp->bb_largest_free_order < ac->ac_2order) | |
2085 | return 0; | |
2086 | ||
8a57d9d6 | 2087 | return 1; |
c9de560d AT |
2088 | case 1: |
2089 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
2090 | return 1; | |
2091 | break; | |
2092 | case 2: | |
2093 | if (free >= ac->ac_g_ex.fe_len) | |
2094 | return 1; | |
2095 | break; | |
2096 | case 3: | |
2097 | return 1; | |
2098 | default: | |
2099 | BUG(); | |
2100 | } | |
2101 | ||
2102 | return 0; | |
2103 | } | |
2104 | ||
4ddfef7b ES |
2105 | static noinline_for_stack int |
2106 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2107 | { |
8df9675f | 2108 | ext4_group_t ngroups, group, i; |
c9de560d | 2109 | int cr; |
42ac1848 | 2110 | int err = 0, first_err = 0; |
c9de560d AT |
2111 | struct ext4_sb_info *sbi; |
2112 | struct super_block *sb; | |
2113 | struct ext4_buddy e4b; | |
c9de560d AT |
2114 | |
2115 | sb = ac->ac_sb; | |
2116 | sbi = EXT4_SB(sb); | |
8df9675f | 2117 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2118 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2119 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2120 | ngroups = sbi->s_blockfile_groups; |
2121 | ||
c9de560d AT |
2122 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2123 | ||
2124 | /* first, try the goal */ | |
2125 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2126 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2127 | goto out; | |
2128 | ||
2129 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2130 | goto out; | |
2131 | ||
2132 | /* | |
2133 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
2134 | * if ac2_order is set we also set criteria to 0 so that we | |
2135 | * try exact allocation using buddy. | |
2136 | */ | |
2137 | i = fls(ac->ac_g_ex.fe_len); | |
2138 | ac->ac_2order = 0; | |
2139 | /* | |
2140 | * We search using buddy data only if the order of the request | |
2141 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2142 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2143 | * We also support searching for power-of-two requests only for |
2144 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2145 | */ |
d9b22cf9 | 2146 | if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) { |
c9de560d AT |
2147 | /* |
2148 | * This should tell if fe_len is exactly power of 2 | |
2149 | */ | |
2150 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d JC |
2151 | ac->ac_2order = array_index_nospec(i - 1, |
2152 | sb->s_blocksize_bits + 2); | |
c9de560d AT |
2153 | } |
2154 | ||
4ba74d00 TT |
2155 | /* if stream allocation is enabled, use global goal */ |
2156 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2157 | /* TBD: may be hot point */ |
2158 | spin_lock(&sbi->s_md_lock); | |
2159 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2160 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2161 | spin_unlock(&sbi->s_md_lock); | |
2162 | } | |
4ba74d00 | 2163 | |
c9de560d AT |
2164 | /* Let's just scan groups to find more-less suitable blocks */ |
2165 | cr = ac->ac_2order ? 0 : 1; | |
2166 | /* | |
2167 | * cr == 0 try to get exact allocation, | |
2168 | * cr == 3 try to get anything | |
2169 | */ | |
2170 | repeat: | |
2171 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2172 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2173 | /* |
2174 | * searching for the right group start | |
2175 | * from the goal value specified | |
2176 | */ | |
2177 | group = ac->ac_g_ex.fe_group; | |
2178 | ||
8df9675f | 2179 | for (i = 0; i < ngroups; group++, i++) { |
42ac1848 | 2180 | int ret = 0; |
2ed5724d | 2181 | cond_resched(); |
e6155736 LM |
2182 | /* |
2183 | * Artificially restricted ngroups for non-extent | |
2184 | * files makes group > ngroups possible on first loop. | |
2185 | */ | |
2186 | if (group >= ngroups) | |
c9de560d AT |
2187 | group = 0; |
2188 | ||
8a57d9d6 | 2189 | /* This now checks without needing the buddy page */ |
42ac1848 LC |
2190 | ret = ext4_mb_good_group(ac, group, cr); |
2191 | if (ret <= 0) { | |
2192 | if (!first_err) | |
2193 | first_err = ret; | |
c9de560d | 2194 | continue; |
42ac1848 | 2195 | } |
c9de560d | 2196 | |
c9de560d AT |
2197 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2198 | if (err) | |
2199 | goto out; | |
2200 | ||
2201 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2202 | |
2203 | /* | |
2204 | * We need to check again after locking the | |
2205 | * block group | |
2206 | */ | |
42ac1848 LC |
2207 | ret = ext4_mb_good_group(ac, group, cr); |
2208 | if (ret <= 0) { | |
c9de560d | 2209 | ext4_unlock_group(sb, group); |
e39e07fd | 2210 | ext4_mb_unload_buddy(&e4b); |
42ac1848 LC |
2211 | if (!first_err) |
2212 | first_err = ret; | |
c9de560d AT |
2213 | continue; |
2214 | } | |
2215 | ||
2216 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2217 | if (cr == 0) |
c9de560d | 2218 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2219 | else if (cr == 1 && sbi->s_stripe && |
2220 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2221 | ext4_mb_scan_aligned(ac, &e4b); |
2222 | else | |
2223 | ext4_mb_complex_scan_group(ac, &e4b); | |
2224 | ||
2225 | ext4_unlock_group(sb, group); | |
e39e07fd | 2226 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2227 | |
2228 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2229 | break; | |
2230 | } | |
2231 | } | |
2232 | ||
2233 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2234 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2235 | /* | |
2236 | * We've been searching too long. Let's try to allocate | |
2237 | * the best chunk we've found so far | |
2238 | */ | |
2239 | ||
2240 | ext4_mb_try_best_found(ac, &e4b); | |
2241 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2242 | /* | |
2243 | * Someone more lucky has already allocated it. | |
2244 | * The only thing we can do is just take first | |
2245 | * found block(s) | |
2246 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2247 | */ | |
2248 | ac->ac_b_ex.fe_group = 0; | |
2249 | ac->ac_b_ex.fe_start = 0; | |
2250 | ac->ac_b_ex.fe_len = 0; | |
2251 | ac->ac_status = AC_STATUS_CONTINUE; | |
2252 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2253 | cr = 3; | |
2254 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2255 | goto repeat; | |
2256 | } | |
2257 | } | |
2258 | out: | |
42ac1848 LC |
2259 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2260 | err = first_err; | |
c9de560d AT |
2261 | return err; |
2262 | } | |
2263 | ||
c9de560d AT |
2264 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2265 | { | |
247dbed8 | 2266 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2267 | ext4_group_t group; |
2268 | ||
8df9675f | 2269 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2270 | return NULL; |
c9de560d | 2271 | group = *pos + 1; |
a9df9a49 | 2272 | return (void *) ((unsigned long) group); |
c9de560d AT |
2273 | } |
2274 | ||
2275 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2276 | { | |
247dbed8 | 2277 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2278 | ext4_group_t group; |
2279 | ||
2280 | ++*pos; | |
8df9675f | 2281 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2282 | return NULL; |
2283 | group = *pos + 1; | |
a9df9a49 | 2284 | return (void *) ((unsigned long) group); |
c9de560d AT |
2285 | } |
2286 | ||
2287 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2288 | { | |
247dbed8 | 2289 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
a9df9a49 | 2290 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2291 | int i; |
1c8457ca | 2292 | int err, buddy_loaded = 0; |
c9de560d | 2293 | struct ext4_buddy e4b; |
1c8457ca | 2294 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2295 | unsigned char blocksize_bits = min_t(unsigned char, |
2296 | sb->s_blocksize_bits, | |
2297 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2298 | struct sg { |
2299 | struct ext4_group_info info; | |
b80b32b6 | 2300 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2301 | } sg; |
2302 | ||
2303 | group--; | |
2304 | if (group == 0) | |
97b4af2f RV |
2305 | seq_puts(seq, "#group: free frags first [" |
2306 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2307 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2308 | |
b80b32b6 TT |
2309 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2310 | sizeof(struct ext4_group_info); | |
2311 | ||
1c8457ca AK |
2312 | grinfo = ext4_get_group_info(sb, group); |
2313 | /* Load the group info in memory only if not already loaded. */ | |
2314 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2315 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2316 | if (err) { | |
2317 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2318 | return 0; | |
2319 | } | |
2320 | buddy_loaded = 1; | |
c9de560d | 2321 | } |
1c8457ca | 2322 | |
b80b32b6 | 2323 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2324 | |
2325 | if (buddy_loaded) | |
2326 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2327 | |
a9df9a49 | 2328 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2329 | sg.info.bb_fragments, sg.info.bb_first_free); |
2330 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2331 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d AT |
2332 | sg.info.bb_counters[i] : 0); |
2333 | seq_printf(seq, " ]\n"); | |
2334 | ||
2335 | return 0; | |
2336 | } | |
2337 | ||
2338 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2339 | { | |
2340 | } | |
2341 | ||
247dbed8 | 2342 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2343 | .start = ext4_mb_seq_groups_start, |
2344 | .next = ext4_mb_seq_groups_next, | |
2345 | .stop = ext4_mb_seq_groups_stop, | |
2346 | .show = ext4_mb_seq_groups_show, | |
2347 | }; | |
2348 | ||
fb1813f4 CW |
2349 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2350 | { | |
2351 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2352 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2353 | ||
2354 | BUG_ON(!cachep); | |
2355 | return cachep; | |
2356 | } | |
5f21b0e6 | 2357 | |
28623c2f TT |
2358 | /* |
2359 | * Allocate the top-level s_group_info array for the specified number | |
2360 | * of groups | |
2361 | */ | |
2362 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2363 | { | |
2364 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2365 | unsigned size; | |
df3da4ea | 2366 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
2367 | |
2368 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2369 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2370 | if (size <= sbi->s_group_info_size) | |
2371 | return 0; | |
2372 | ||
2373 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 2374 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
2375 | if (!new_groupinfo) { |
2376 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2377 | return -ENOMEM; | |
2378 | } | |
df3da4ea SJS |
2379 | rcu_read_lock(); |
2380 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
2381 | if (old_groupinfo) | |
2382 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 2383 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
2384 | rcu_read_unlock(); |
2385 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 2386 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
2387 | if (old_groupinfo) |
2388 | ext4_kvfree_array_rcu(old_groupinfo); | |
28623c2f TT |
2389 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
2390 | sbi->s_group_info_size); | |
2391 | return 0; | |
2392 | } | |
2393 | ||
5f21b0e6 | 2394 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2395 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2396 | struct ext4_group_desc *desc) |
2397 | { | |
fb1813f4 | 2398 | int i; |
5f21b0e6 | 2399 | int metalen = 0; |
df3da4ea | 2400 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
2401 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2402 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2403 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2404 | |
2405 | /* | |
2406 | * First check if this group is the first of a reserved block. | |
2407 | * If it's true, we have to allocate a new table of pointers | |
2408 | * to ext4_group_info structures | |
2409 | */ | |
2410 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2411 | metalen = sizeof(*meta_group_info) << | |
2412 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 2413 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 2414 | if (meta_group_info == NULL) { |
7f6a11e7 | 2415 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2416 | "for a buddy group"); |
5f21b0e6 FB |
2417 | goto exit_meta_group_info; |
2418 | } | |
df3da4ea SJS |
2419 | rcu_read_lock(); |
2420 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
2421 | rcu_read_unlock(); | |
5f21b0e6 FB |
2422 | } |
2423 | ||
df3da4ea | 2424 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
2425 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
2426 | ||
4fdb5543 | 2427 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 2428 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2429 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2430 | goto exit_group_info; |
2431 | } | |
2432 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2433 | &(meta_group_info[i]->bb_state)); | |
2434 | ||
2435 | /* | |
2436 | * initialize bb_free to be able to skip | |
2437 | * empty groups without initialization | |
2438 | */ | |
8844618d TT |
2439 | if (ext4_has_group_desc_csum(sb) && |
2440 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 2441 | meta_group_info[i]->bb_free = |
cff1dfd7 | 2442 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2443 | } else { |
2444 | meta_group_info[i]->bb_free = | |
021b65bb | 2445 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2446 | } |
2447 | ||
2448 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2449 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2450 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2451 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 FB |
2452 | |
2453 | #ifdef DOUBLE_CHECK | |
2454 | { | |
2455 | struct buffer_head *bh; | |
2456 | meta_group_info[i]->bb_bitmap = | |
4fdb5543 | 2457 | kmalloc(sb->s_blocksize, GFP_NOFS); |
5f21b0e6 FB |
2458 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); |
2459 | bh = ext4_read_block_bitmap(sb, group); | |
9008a58e | 2460 | BUG_ON(IS_ERR_OR_NULL(bh)); |
5f21b0e6 FB |
2461 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, |
2462 | sb->s_blocksize); | |
2463 | put_bh(bh); | |
2464 | } | |
2465 | #endif | |
2466 | ||
2467 | return 0; | |
2468 | ||
2469 | exit_group_info: | |
2470 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2471 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
2472 | struct ext4_group_info ***group_info; |
2473 | ||
2474 | rcu_read_lock(); | |
2475 | group_info = rcu_dereference(sbi->s_group_info); | |
2476 | kfree(group_info[idx]); | |
2477 | group_info[idx] = NULL; | |
2478 | rcu_read_unlock(); | |
caaf7a29 | 2479 | } |
5f21b0e6 FB |
2480 | exit_meta_group_info: |
2481 | return -ENOMEM; | |
2482 | } /* ext4_mb_add_groupinfo */ | |
2483 | ||
c9de560d AT |
2484 | static int ext4_mb_init_backend(struct super_block *sb) |
2485 | { | |
8df9675f | 2486 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2487 | ext4_group_t i; |
c9de560d | 2488 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2489 | int err; |
5f21b0e6 | 2490 | struct ext4_group_desc *desc; |
df3da4ea | 2491 | struct ext4_group_info ***group_info; |
fb1813f4 | 2492 | struct kmem_cache *cachep; |
5f21b0e6 | 2493 | |
28623c2f TT |
2494 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2495 | if (err) | |
2496 | return err; | |
c9de560d | 2497 | |
c9de560d AT |
2498 | sbi->s_buddy_cache = new_inode(sb); |
2499 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2500 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2501 | goto err_freesgi; |
2502 | } | |
48e6061b YJ |
2503 | /* To avoid potentially colliding with an valid on-disk inode number, |
2504 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2505 | * not in the inode hash, so it should never be found by iget(), but | |
2506 | * this will avoid confusion if it ever shows up during debugging. */ | |
2507 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2508 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2509 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 2510 | cond_resched(); |
c9de560d AT |
2511 | desc = ext4_get_group_desc(sb, i, NULL); |
2512 | if (desc == NULL) { | |
9d8b9ec4 | 2513 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2514 | goto err_freebuddy; |
2515 | } | |
5f21b0e6 FB |
2516 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2517 | goto err_freebuddy; | |
c9de560d AT |
2518 | } |
2519 | ||
2520 | return 0; | |
2521 | ||
2522 | err_freebuddy: | |
fb1813f4 | 2523 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2524 | while (i-- > 0) |
fb1813f4 | 2525 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2526 | i = sbi->s_group_info_size; |
df3da4ea SJS |
2527 | rcu_read_lock(); |
2528 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 2529 | while (i-- > 0) |
df3da4ea SJS |
2530 | kfree(group_info[i]); |
2531 | rcu_read_unlock(); | |
c9de560d AT |
2532 | iput(sbi->s_buddy_cache); |
2533 | err_freesgi: | |
df3da4ea SJS |
2534 | rcu_read_lock(); |
2535 | kvfree(rcu_dereference(sbi->s_group_info)); | |
2536 | rcu_read_unlock(); | |
c9de560d AT |
2537 | return -ENOMEM; |
2538 | } | |
2539 | ||
2892c15d ES |
2540 | static void ext4_groupinfo_destroy_slabs(void) |
2541 | { | |
2542 | int i; | |
2543 | ||
2544 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 2545 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
2546 | ext4_groupinfo_caches[i] = NULL; |
2547 | } | |
2548 | } | |
2549 | ||
2550 | static int ext4_groupinfo_create_slab(size_t size) | |
2551 | { | |
2552 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2553 | int slab_size; | |
2554 | int blocksize_bits = order_base_2(size); | |
2555 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2556 | struct kmem_cache *cachep; | |
2557 | ||
2558 | if (cache_index >= NR_GRPINFO_CACHES) | |
2559 | return -EINVAL; | |
2560 | ||
2561 | if (unlikely(cache_index < 0)) | |
2562 | cache_index = 0; | |
2563 | ||
2564 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2565 | if (ext4_groupinfo_caches[cache_index]) { | |
2566 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2567 | return 0; /* Already created */ | |
2568 | } | |
2569 | ||
2570 | slab_size = offsetof(struct ext4_group_info, | |
2571 | bb_counters[blocksize_bits + 2]); | |
2572 | ||
2573 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2574 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2575 | NULL); | |
2576 | ||
823ba01f TM |
2577 | ext4_groupinfo_caches[cache_index] = cachep; |
2578 | ||
2892c15d ES |
2579 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2580 | if (!cachep) { | |
9d8b9ec4 TT |
2581 | printk(KERN_EMERG |
2582 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2583 | return -ENOMEM; |
2584 | } | |
2585 | ||
2892c15d ES |
2586 | return 0; |
2587 | } | |
2588 | ||
9d99012f | 2589 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2590 | { |
2591 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2592 | unsigned i, j; |
935244cd | 2593 | unsigned offset, offset_incr; |
c9de560d | 2594 | unsigned max; |
74767c5a | 2595 | int ret; |
c9de560d | 2596 | |
1927805e | 2597 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2598 | |
2599 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2600 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2601 | ret = -ENOMEM; |
2602 | goto out; | |
c9de560d | 2603 | } |
ff7ef329 | 2604 | |
1927805e | 2605 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2606 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2607 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2608 | ret = -ENOMEM; |
2609 | goto out; | |
2610 | } | |
2611 | ||
2892c15d ES |
2612 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2613 | if (ret < 0) | |
2614 | goto out; | |
c9de560d AT |
2615 | |
2616 | /* order 0 is regular bitmap */ | |
2617 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2618 | sbi->s_mb_offsets[0] = 0; | |
2619 | ||
2620 | i = 1; | |
2621 | offset = 0; | |
935244cd | 2622 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
2623 | max = sb->s_blocksize << 2; |
2624 | do { | |
2625 | sbi->s_mb_offsets[i] = offset; | |
2626 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
2627 | offset += offset_incr; |
2628 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
2629 | max = max >> 1; |
2630 | i++; | |
2631 | } while (i <= sb->s_blocksize_bits + 1); | |
2632 | ||
c9de560d | 2633 | spin_lock_init(&sbi->s_md_lock); |
c9de560d | 2634 | spin_lock_init(&sbi->s_bal_lock); |
d08854f5 | 2635 | sbi->s_mb_free_pending = 0; |
a0154344 | 2636 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
c9de560d AT |
2637 | |
2638 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2639 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2640 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2641 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2642 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
2643 | /* |
2644 | * The default group preallocation is 512, which for 4k block | |
2645 | * sizes translates to 2 megabytes. However for bigalloc file | |
2646 | * systems, this is probably too big (i.e, if the cluster size | |
2647 | * is 1 megabyte, then group preallocation size becomes half a | |
2648 | * gigabyte!). As a default, we will keep a two megabyte | |
2649 | * group pralloc size for cluster sizes up to 64k, and after | |
2650 | * that, we will force a minimum group preallocation size of | |
2651 | * 32 clusters. This translates to 8 megs when the cluster | |
2652 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2653 | * which seems reasonable as a default. | |
2654 | */ | |
2655 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2656 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2657 | /* |
2658 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2659 | * to the lowest multiple of s_stripe which is bigger than | |
2660 | * the s_mb_group_prealloc as determined above. We want | |
2661 | * the preallocation size to be an exact multiple of the | |
2662 | * RAID stripe size so that preallocations don't fragment | |
2663 | * the stripes. | |
2664 | */ | |
2665 | if (sbi->s_stripe > 1) { | |
2666 | sbi->s_mb_group_prealloc = roundup( | |
2667 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2668 | } | |
c9de560d | 2669 | |
730c213c | 2670 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2671 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2672 | ret = -ENOMEM; |
029b10c5 | 2673 | goto out; |
c9de560d | 2674 | } |
730c213c | 2675 | for_each_possible_cpu(i) { |
c9de560d | 2676 | struct ext4_locality_group *lg; |
730c213c | 2677 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2678 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2679 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2680 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2681 | spin_lock_init(&lg->lg_prealloc_lock); |
2682 | } | |
2683 | ||
79a77c5a YJ |
2684 | /* init file for buddy data */ |
2685 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2686 | if (ret != 0) |
2687 | goto out_free_locality_groups; | |
79a77c5a | 2688 | |
7aa0baea TM |
2689 | return 0; |
2690 | ||
2691 | out_free_locality_groups: | |
2692 | free_percpu(sbi->s_locality_groups); | |
2693 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 2694 | out: |
7aa0baea TM |
2695 | kfree(sbi->s_mb_offsets); |
2696 | sbi->s_mb_offsets = NULL; | |
2697 | kfree(sbi->s_mb_maxs); | |
2698 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2699 | return ret; |
c9de560d AT |
2700 | } |
2701 | ||
955ce5f5 | 2702 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2703 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2704 | { | |
2705 | struct ext4_prealloc_space *pa; | |
2706 | struct list_head *cur, *tmp; | |
2707 | int count = 0; | |
2708 | ||
2709 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2710 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2711 | list_del(&pa->pa_group_list); | |
2712 | count++; | |
688f05a0 | 2713 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2714 | } |
2715 | if (count) | |
6ba495e9 | 2716 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2717 | |
2718 | } | |
2719 | ||
2720 | int ext4_mb_release(struct super_block *sb) | |
2721 | { | |
8df9675f | 2722 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2723 | ext4_group_t i; |
2724 | int num_meta_group_infos; | |
df3da4ea | 2725 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 2726 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 2727 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2728 | |
c9de560d | 2729 | if (sbi->s_group_info) { |
8df9675f | 2730 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 2731 | cond_resched(); |
c9de560d AT |
2732 | grinfo = ext4_get_group_info(sb, i); |
2733 | #ifdef DOUBLE_CHECK | |
2734 | kfree(grinfo->bb_bitmap); | |
2735 | #endif | |
2736 | ext4_lock_group(sb, i); | |
2737 | ext4_mb_cleanup_pa(grinfo); | |
2738 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2739 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2740 | } |
8df9675f | 2741 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2742 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2743 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
2744 | rcu_read_lock(); |
2745 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 2746 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
2747 | kfree(group_info[i]); |
2748 | kvfree(group_info); | |
2749 | rcu_read_unlock(); | |
c9de560d AT |
2750 | } |
2751 | kfree(sbi->s_mb_offsets); | |
2752 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 2753 | iput(sbi->s_buddy_cache); |
c9de560d | 2754 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
2755 | ext4_msg(sb, KERN_INFO, |
2756 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2757 | atomic_read(&sbi->s_bal_allocated), |
2758 | atomic_read(&sbi->s_bal_reqs), | |
2759 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2760 | ext4_msg(sb, KERN_INFO, |
2761 | "mballoc: %u extents scanned, %u goal hits, " | |
2762 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2763 | atomic_read(&sbi->s_bal_ex_scanned), |
2764 | atomic_read(&sbi->s_bal_goals), | |
2765 | atomic_read(&sbi->s_bal_2orders), | |
2766 | atomic_read(&sbi->s_bal_breaks), | |
2767 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2768 | ext4_msg(sb, KERN_INFO, |
2769 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2770 | sbi->s_mb_buddies_generated, |
c9de560d | 2771 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2772 | ext4_msg(sb, KERN_INFO, |
2773 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2774 | atomic_read(&sbi->s_mb_preallocated), |
2775 | atomic_read(&sbi->s_mb_discarded)); | |
2776 | } | |
2777 | ||
730c213c | 2778 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2779 | |
2780 | return 0; | |
2781 | } | |
2782 | ||
77ca6cdf | 2783 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
2784 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
2785 | struct bio **biop) | |
5c521830 | 2786 | { |
5c521830 JZ |
2787 | ext4_fsblk_t discard_block; |
2788 | ||
84130193 TT |
2789 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
2790 | ext4_group_first_block_no(sb, block_group)); | |
2791 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
2792 | trace_ext4_discard_blocks(sb, |
2793 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
2794 | if (biop) { |
2795 | return __blkdev_issue_discard(sb->s_bdev, | |
2796 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
2797 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
2798 | GFP_NOFS, 0, biop); | |
2799 | } else | |
2800 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
2801 | } |
2802 | ||
a0154344 DJ |
2803 | static void ext4_free_data_in_buddy(struct super_block *sb, |
2804 | struct ext4_free_data *entry) | |
c9de560d | 2805 | { |
c9de560d | 2806 | struct ext4_buddy e4b; |
c894058d | 2807 | struct ext4_group_info *db; |
d9f34504 | 2808 | int err, count = 0, count2 = 0; |
c9de560d | 2809 | |
18aadd47 BJ |
2810 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
2811 | entry->efd_count, entry->efd_group, entry); | |
c9de560d | 2812 | |
18aadd47 BJ |
2813 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
2814 | /* we expect to find existing buddy because it's pinned */ | |
2815 | BUG_ON(err != 0); | |
b90f6870 | 2816 | |
d08854f5 TT |
2817 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
2818 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
2819 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 2820 | |
18aadd47 BJ |
2821 | db = e4b.bd_info; |
2822 | /* there are blocks to put in buddy to make them really free */ | |
2823 | count += entry->efd_count; | |
2824 | count2++; | |
2825 | ext4_lock_group(sb, entry->efd_group); | |
2826 | /* Take it out of per group rb tree */ | |
2827 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
2828 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 2829 | |
18aadd47 BJ |
2830 | /* |
2831 | * Clear the trimmed flag for the group so that the next | |
2832 | * ext4_trim_fs can trim it. | |
2833 | * If the volume is mounted with -o discard, online discard | |
2834 | * is supported and the free blocks will be trimmed online. | |
2835 | */ | |
2836 | if (!test_opt(sb, DISCARD)) | |
2837 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 2838 | |
18aadd47 BJ |
2839 | if (!db->bb_free_root.rb_node) { |
2840 | /* No more items in the per group rb tree | |
2841 | * balance refcounts from ext4_mb_free_metadata() | |
2842 | */ | |
09cbfeaf KS |
2843 | put_page(e4b.bd_buddy_page); |
2844 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 2845 | } |
18aadd47 BJ |
2846 | ext4_unlock_group(sb, entry->efd_group); |
2847 | kmem_cache_free(ext4_free_data_cachep, entry); | |
2848 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2849 | |
6ba495e9 | 2850 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2851 | } |
2852 | ||
a0154344 DJ |
2853 | /* |
2854 | * This function is called by the jbd2 layer once the commit has finished, | |
2855 | * so we know we can free the blocks that were released with that commit. | |
2856 | */ | |
2857 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
2858 | { | |
2859 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2860 | struct ext4_free_data *entry, *tmp; | |
2861 | struct bio *discard_bio = NULL; | |
2862 | struct list_head freed_data_list; | |
2863 | struct list_head *cut_pos = NULL; | |
2864 | int err; | |
2865 | ||
2866 | INIT_LIST_HEAD(&freed_data_list); | |
2867 | ||
2868 | spin_lock(&sbi->s_md_lock); | |
2869 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
2870 | if (entry->efd_tid != commit_tid) | |
2871 | break; | |
2872 | cut_pos = &entry->efd_list; | |
2873 | } | |
2874 | if (cut_pos) | |
2875 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
2876 | cut_pos); | |
2877 | spin_unlock(&sbi->s_md_lock); | |
2878 | ||
2879 | if (test_opt(sb, DISCARD)) { | |
2880 | list_for_each_entry(entry, &freed_data_list, efd_list) { | |
2881 | err = ext4_issue_discard(sb, entry->efd_group, | |
2882 | entry->efd_start_cluster, | |
2883 | entry->efd_count, | |
2884 | &discard_bio); | |
2885 | if (err && err != -EOPNOTSUPP) { | |
2886 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
2887 | " group:%d block:%d count:%d failed" | |
2888 | " with %d", entry->efd_group, | |
2889 | entry->efd_start_cluster, | |
2890 | entry->efd_count, err); | |
2891 | } else if (err == -EOPNOTSUPP) | |
2892 | break; | |
2893 | } | |
2894 | ||
e4510577 | 2895 | if (discard_bio) { |
a0154344 | 2896 | submit_bio_wait(discard_bio); |
e4510577 DJ |
2897 | bio_put(discard_bio); |
2898 | } | |
a0154344 DJ |
2899 | } |
2900 | ||
2901 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
2902 | ext4_free_data_in_buddy(sb, entry); | |
2903 | } | |
2904 | ||
5dabfc78 | 2905 | int __init ext4_init_mballoc(void) |
c9de560d | 2906 | { |
16828088 TT |
2907 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2908 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d AT |
2909 | if (ext4_pspace_cachep == NULL) |
2910 | return -ENOMEM; | |
2911 | ||
16828088 TT |
2912 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2913 | SLAB_RECLAIM_ACCOUNT); | |
256bdb49 ES |
2914 | if (ext4_ac_cachep == NULL) { |
2915 | kmem_cache_destroy(ext4_pspace_cachep); | |
2916 | return -ENOMEM; | |
2917 | } | |
c894058d | 2918 | |
18aadd47 BJ |
2919 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
2920 | SLAB_RECLAIM_ACCOUNT); | |
2921 | if (ext4_free_data_cachep == NULL) { | |
c894058d AK |
2922 | kmem_cache_destroy(ext4_pspace_cachep); |
2923 | kmem_cache_destroy(ext4_ac_cachep); | |
2924 | return -ENOMEM; | |
2925 | } | |
c9de560d AT |
2926 | return 0; |
2927 | } | |
2928 | ||
5dabfc78 | 2929 | void ext4_exit_mballoc(void) |
c9de560d | 2930 | { |
60e6679e | 2931 | /* |
3e03f9ca JDB |
2932 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2933 | * before destroying the slab cache. | |
2934 | */ | |
2935 | rcu_barrier(); | |
c9de560d | 2936 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2937 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 2938 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 2939 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
2940 | } |
2941 | ||
2942 | ||
2943 | /* | |
73b2c716 | 2944 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
2945 | * Returns 0 if success or error code |
2946 | */ | |
4ddfef7b ES |
2947 | static noinline_for_stack int |
2948 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 2949 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
2950 | { |
2951 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
2952 | struct ext4_group_desc *gdp; |
2953 | struct buffer_head *gdp_bh; | |
2954 | struct ext4_sb_info *sbi; | |
2955 | struct super_block *sb; | |
2956 | ext4_fsblk_t block; | |
519deca0 | 2957 | int err, len; |
c9de560d AT |
2958 | |
2959 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2960 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2961 | ||
2962 | sb = ac->ac_sb; | |
2963 | sbi = EXT4_SB(sb); | |
c9de560d | 2964 | |
574ca174 | 2965 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
2966 | if (IS_ERR(bitmap_bh)) { |
2967 | err = PTR_ERR(bitmap_bh); | |
2968 | bitmap_bh = NULL; | |
c9de560d | 2969 | goto out_err; |
9008a58e | 2970 | } |
c9de560d | 2971 | |
5d601255 | 2972 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
c9de560d AT |
2973 | err = ext4_journal_get_write_access(handle, bitmap_bh); |
2974 | if (err) | |
2975 | goto out_err; | |
2976 | ||
2977 | err = -EIO; | |
2978 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2979 | if (!gdp) | |
2980 | goto out_err; | |
2981 | ||
a9df9a49 | 2982 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 2983 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 2984 | |
5d601255 | 2985 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
c9de560d AT |
2986 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2987 | if (err) | |
2988 | goto out_err; | |
2989 | ||
bda00de7 | 2990 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 2991 | |
53accfa9 | 2992 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 2993 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 2994 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 2995 | "fs metadata", block, block+len); |
519deca0 | 2996 | /* File system mounted not to panic on error |
554a5ccc | 2997 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
2998 | * We leak some of the blocks here. |
2999 | */ | |
955ce5f5 | 3000 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
3001 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3002 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 3003 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3004 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3005 | if (!err) |
554a5ccc | 3006 | err = -EFSCORRUPTED; |
519deca0 | 3007 | goto out_err; |
c9de560d | 3008 | } |
955ce5f5 AK |
3009 | |
3010 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3011 | #ifdef AGGRESSIVE_CHECK |
3012 | { | |
3013 | int i; | |
3014 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3015 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3016 | bitmap_bh->b_data)); | |
3017 | } | |
3018 | } | |
3019 | #endif | |
c3e94d1d YY |
3020 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3021 | ac->ac_b_ex.fe_len); | |
8844618d TT |
3022 | if (ext4_has_group_desc_csum(sb) && |
3023 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3024 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3025 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3026 | ext4_free_clusters_after_init(sb, |
021b65bb | 3027 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3028 | } |
021b65bb TT |
3029 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3030 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3031 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3032 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3033 | |
3034 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3035 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3036 | /* |
6bc6e63f | 3037 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3038 | */ |
6bc6e63f AK |
3039 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3040 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3041 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3042 | reserv_clstrs); | |
c9de560d | 3043 | |
772cb7c8 JS |
3044 | if (sbi->s_log_groups_per_flex) { |
3045 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3046 | ac->ac_b_ex.fe_group); | |
90ba983f | 3047 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3048 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3049 | flex_group)->free_clusters); | |
772cb7c8 JS |
3050 | } |
3051 | ||
0390131b | 3052 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3053 | if (err) |
3054 | goto out_err; | |
0390131b | 3055 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3056 | |
3057 | out_err: | |
42a10add | 3058 | brelse(bitmap_bh); |
c9de560d AT |
3059 | return err; |
3060 | } | |
3061 | ||
3062 | /* | |
3063 | * here we normalize request for locality group | |
d7a1fee1 DE |
3064 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3065 | * s_strip if we set the same via mount option. | |
3066 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3067 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3068 | * |
3069 | * XXX: should we try to preallocate more than the group has now? | |
3070 | */ | |
3071 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3072 | { | |
3073 | struct super_block *sb = ac->ac_sb; | |
3074 | struct ext4_locality_group *lg = ac->ac_lg; | |
3075 | ||
3076 | BUG_ON(lg == NULL); | |
d7a1fee1 | 3077 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 3078 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
3079 | current->pid, ac->ac_g_ex.fe_len); |
3080 | } | |
3081 | ||
3082 | /* | |
3083 | * Normalization means making request better in terms of | |
3084 | * size and alignment | |
3085 | */ | |
4ddfef7b ES |
3086 | static noinline_for_stack void |
3087 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3088 | struct ext4_allocation_request *ar) |
3089 | { | |
53accfa9 | 3090 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3091 | int bsbits, max; |
3092 | ext4_lblk_t end; | |
1592d2c5 CW |
3093 | loff_t size, start_off; |
3094 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3095 | ext4_lblk_t start; |
c9de560d | 3096 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 3097 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3098 | |
3099 | /* do normalize only data requests, metadata requests | |
3100 | do not need preallocation */ | |
3101 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3102 | return; | |
3103 | ||
3104 | /* sometime caller may want exact blocks */ | |
3105 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3106 | return; | |
3107 | ||
3108 | /* caller may indicate that preallocation isn't | |
3109 | * required (it's a tail, for example) */ | |
3110 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3111 | return; | |
3112 | ||
3113 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3114 | ext4_mb_normalize_group_request(ac); | |
3115 | return ; | |
3116 | } | |
3117 | ||
3118 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3119 | ||
3120 | /* first, let's learn actual file size | |
3121 | * given current request is allocated */ | |
53accfa9 | 3122 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
3123 | size = size << bsbits; |
3124 | if (size < i_size_read(ac->ac_inode)) | |
3125 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 3126 | orig_size = size; |
c9de560d | 3127 | |
1930479c VC |
3128 | /* max size of free chunks */ |
3129 | max = 2 << bsbits; | |
c9de560d | 3130 | |
1930479c VC |
3131 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3132 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3133 | |
3134 | /* first, try to predict filesize */ | |
3135 | /* XXX: should this table be tunable? */ | |
3136 | start_off = 0; | |
3137 | if (size <= 16 * 1024) { | |
3138 | size = 16 * 1024; | |
3139 | } else if (size <= 32 * 1024) { | |
3140 | size = 32 * 1024; | |
3141 | } else if (size <= 64 * 1024) { | |
3142 | size = 64 * 1024; | |
3143 | } else if (size <= 128 * 1024) { | |
3144 | size = 128 * 1024; | |
3145 | } else if (size <= 256 * 1024) { | |
3146 | size = 256 * 1024; | |
3147 | } else if (size <= 512 * 1024) { | |
3148 | size = 512 * 1024; | |
3149 | } else if (size <= 1024 * 1024) { | |
3150 | size = 1024 * 1024; | |
1930479c | 3151 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3152 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3153 | (21 - bsbits)) << 21; |
3154 | size = 2 * 1024 * 1024; | |
3155 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3156 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3157 | (22 - bsbits)) << 22; | |
3158 | size = 4 * 1024 * 1024; | |
3159 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3160 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3161 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3162 | (23 - bsbits)) << 23; | |
3163 | size = 8 * 1024 * 1024; | |
3164 | } else { | |
b27b1535 XW |
3165 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
3166 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
3167 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 3168 | } |
5a0790c2 AK |
3169 | size = size >> bsbits; |
3170 | start = start_off >> bsbits; | |
c9de560d AT |
3171 | |
3172 | /* don't cover already allocated blocks in selected range */ | |
3173 | if (ar->pleft && start <= ar->lleft) { | |
3174 | size -= ar->lleft + 1 - start; | |
3175 | start = ar->lleft + 1; | |
3176 | } | |
3177 | if (ar->pright && start + size - 1 >= ar->lright) | |
3178 | size -= start + size - ar->lright; | |
3179 | ||
cd648b8a JK |
3180 | /* |
3181 | * Trim allocation request for filesystems with artificially small | |
3182 | * groups. | |
3183 | */ | |
3184 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
3185 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
3186 | ||
c9de560d AT |
3187 | end = start + size; |
3188 | ||
3189 | /* check we don't cross already preallocated blocks */ | |
3190 | rcu_read_lock(); | |
9a0762c5 | 3191 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3192 | ext4_lblk_t pa_end; |
c9de560d | 3193 | |
c9de560d AT |
3194 | if (pa->pa_deleted) |
3195 | continue; | |
3196 | spin_lock(&pa->pa_lock); | |
3197 | if (pa->pa_deleted) { | |
3198 | spin_unlock(&pa->pa_lock); | |
3199 | continue; | |
3200 | } | |
3201 | ||
53accfa9 TT |
3202 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3203 | pa->pa_len); | |
c9de560d AT |
3204 | |
3205 | /* PA must not overlap original request */ | |
3206 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3207 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3208 | ||
38877f4e ES |
3209 | /* skip PAs this normalized request doesn't overlap with */ |
3210 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3211 | spin_unlock(&pa->pa_lock); |
3212 | continue; | |
3213 | } | |
3214 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3215 | ||
38877f4e | 3216 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3217 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3218 | BUG_ON(pa_end < start); | |
3219 | start = pa_end; | |
38877f4e | 3220 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3221 | BUG_ON(pa->pa_lstart > end); |
3222 | end = pa->pa_lstart; | |
3223 | } | |
3224 | spin_unlock(&pa->pa_lock); | |
3225 | } | |
3226 | rcu_read_unlock(); | |
3227 | size = end - start; | |
3228 | ||
3229 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3230 | rcu_read_lock(); | |
9a0762c5 | 3231 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3232 | ext4_lblk_t pa_end; |
53accfa9 | 3233 | |
c9de560d AT |
3234 | spin_lock(&pa->pa_lock); |
3235 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3236 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3237 | pa->pa_len); | |
c9de560d AT |
3238 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3239 | } | |
3240 | spin_unlock(&pa->pa_lock); | |
3241 | } | |
3242 | rcu_read_unlock(); | |
3243 | ||
3244 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3245 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3246 | ext4_msg(ac->ac_sb, KERN_ERR, |
3247 | "start %lu, size %lu, fe_logical %lu", | |
3248 | (unsigned long) start, (unsigned long) size, | |
3249 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 3250 | BUG(); |
c9de560d | 3251 | } |
b5b60778 | 3252 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3253 | |
3254 | /* now prepare goal request */ | |
3255 | ||
3256 | /* XXX: is it better to align blocks WRT to logical | |
3257 | * placement or satisfy big request as is */ | |
3258 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3259 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3260 | |
3261 | /* define goal start in order to merge */ | |
3262 | if (ar->pright && (ar->lright == (start + size))) { | |
3263 | /* merge to the right */ | |
3264 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3265 | &ac->ac_f_ex.fe_group, | |
3266 | &ac->ac_f_ex.fe_start); | |
3267 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3268 | } | |
3269 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3270 | /* merge to the left */ | |
3271 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3272 | &ac->ac_f_ex.fe_group, | |
3273 | &ac->ac_f_ex.fe_start); | |
3274 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3275 | } | |
3276 | ||
6ba495e9 | 3277 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3278 | (unsigned) orig_size, (unsigned) start); |
3279 | } | |
3280 | ||
3281 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3282 | { | |
3283 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3284 | ||
3285 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3286 | atomic_inc(&sbi->s_bal_reqs); | |
3287 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3288 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3289 | atomic_inc(&sbi->s_bal_success); |
3290 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3291 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3292 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3293 | atomic_inc(&sbi->s_bal_goals); | |
3294 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3295 | atomic_inc(&sbi->s_bal_breaks); | |
3296 | } | |
3297 | ||
296c355c TT |
3298 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3299 | trace_ext4_mballoc_alloc(ac); | |
3300 | else | |
3301 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3302 | } |
3303 | ||
b844167e CW |
3304 | /* |
3305 | * Called on failure; free up any blocks from the inode PA for this | |
3306 | * context. We don't need this for MB_GROUP_PA because we only change | |
3307 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3308 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3309 | */ | |
3310 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3311 | { | |
3312 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
3313 | struct ext4_buddy e4b; |
3314 | int err; | |
b844167e | 3315 | |
86f0afd4 | 3316 | if (pa == NULL) { |
c99d1e6e TT |
3317 | if (ac->ac_f_ex.fe_len == 0) |
3318 | return; | |
86f0afd4 TT |
3319 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
3320 | if (err) { | |
3321 | /* | |
3322 | * This should never happen since we pin the | |
3323 | * pages in the ext4_allocation_context so | |
3324 | * ext4_mb_load_buddy() should never fail. | |
3325 | */ | |
3326 | WARN(1, "mb_load_buddy failed (%d)", err); | |
3327 | return; | |
3328 | } | |
3329 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
3330 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
3331 | ac->ac_f_ex.fe_len); | |
3332 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 3333 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
3334 | return; |
3335 | } | |
3336 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 3337 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
3338 | } |
3339 | ||
c9de560d AT |
3340 | /* |
3341 | * use blocks preallocated to inode | |
3342 | */ | |
3343 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3344 | struct ext4_prealloc_space *pa) | |
3345 | { | |
53accfa9 | 3346 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3347 | ext4_fsblk_t start; |
3348 | ext4_fsblk_t end; | |
3349 | int len; | |
3350 | ||
3351 | /* found preallocated blocks, use them */ | |
3352 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3353 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3354 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3355 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3356 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3357 | &ac->ac_b_ex.fe_start); | |
3358 | ac->ac_b_ex.fe_len = len; | |
3359 | ac->ac_status = AC_STATUS_FOUND; | |
3360 | ac->ac_pa = pa; | |
3361 | ||
3362 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3363 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3364 | BUG_ON(pa->pa_free < len); |
3365 | pa->pa_free -= len; | |
3366 | ||
6ba495e9 | 3367 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3368 | } |
3369 | ||
3370 | /* | |
3371 | * use blocks preallocated to locality group | |
3372 | */ | |
3373 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3374 | struct ext4_prealloc_space *pa) | |
3375 | { | |
03cddb80 | 3376 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3377 | |
c9de560d AT |
3378 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3379 | &ac->ac_b_ex.fe_group, | |
3380 | &ac->ac_b_ex.fe_start); | |
3381 | ac->ac_b_ex.fe_len = len; | |
3382 | ac->ac_status = AC_STATUS_FOUND; | |
3383 | ac->ac_pa = pa; | |
3384 | ||
3385 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3386 | * possible race when the group is being loaded concurrently |
c9de560d | 3387 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3388 | * in on-disk bitmap -- see ext4_mb_release_context() |
3389 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3390 | */ |
6ba495e9 | 3391 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3392 | } |
3393 | ||
5e745b04 AK |
3394 | /* |
3395 | * Return the prealloc space that have minimal distance | |
3396 | * from the goal block. @cpa is the prealloc | |
3397 | * space that is having currently known minimal distance | |
3398 | * from the goal block. | |
3399 | */ | |
3400 | static struct ext4_prealloc_space * | |
3401 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3402 | struct ext4_prealloc_space *pa, | |
3403 | struct ext4_prealloc_space *cpa) | |
3404 | { | |
3405 | ext4_fsblk_t cur_distance, new_distance; | |
3406 | ||
3407 | if (cpa == NULL) { | |
3408 | atomic_inc(&pa->pa_count); | |
3409 | return pa; | |
3410 | } | |
79211c8e AM |
3411 | cur_distance = abs(goal_block - cpa->pa_pstart); |
3412 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 3413 | |
5a54b2f1 | 3414 | if (cur_distance <= new_distance) |
5e745b04 AK |
3415 | return cpa; |
3416 | ||
3417 | /* drop the previous reference */ | |
3418 | atomic_dec(&cpa->pa_count); | |
3419 | atomic_inc(&pa->pa_count); | |
3420 | return pa; | |
3421 | } | |
3422 | ||
c9de560d AT |
3423 | /* |
3424 | * search goal blocks in preallocated space | |
3425 | */ | |
4ddfef7b ES |
3426 | static noinline_for_stack int |
3427 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3428 | { |
53accfa9 | 3429 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3430 | int order, i; |
c9de560d AT |
3431 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3432 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3433 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3434 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3435 | |
3436 | /* only data can be preallocated */ | |
3437 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3438 | return 0; | |
3439 | ||
3440 | /* first, try per-file preallocation */ | |
3441 | rcu_read_lock(); | |
9a0762c5 | 3442 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3443 | |
3444 | /* all fields in this condition don't change, | |
3445 | * so we can skip locking for them */ | |
3446 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3447 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3448 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3449 | continue; |
3450 | ||
fb0a387d | 3451 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3452 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3453 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3454 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3455 | continue; |
3456 | ||
c9de560d AT |
3457 | /* found preallocated blocks, use them */ |
3458 | spin_lock(&pa->pa_lock); | |
3459 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3460 | atomic_inc(&pa->pa_count); | |
3461 | ext4_mb_use_inode_pa(ac, pa); | |
3462 | spin_unlock(&pa->pa_lock); | |
3463 | ac->ac_criteria = 10; | |
3464 | rcu_read_unlock(); | |
3465 | return 1; | |
3466 | } | |
3467 | spin_unlock(&pa->pa_lock); | |
3468 | } | |
3469 | rcu_read_unlock(); | |
3470 | ||
3471 | /* can we use group allocation? */ | |
3472 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3473 | return 0; | |
3474 | ||
3475 | /* inode may have no locality group for some reason */ | |
3476 | lg = ac->ac_lg; | |
3477 | if (lg == NULL) | |
3478 | return 0; | |
6be2ded1 AK |
3479 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3480 | if (order > PREALLOC_TB_SIZE - 1) | |
3481 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3482 | order = PREALLOC_TB_SIZE - 1; | |
3483 | ||
bda00de7 | 3484 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3485 | /* |
3486 | * search for the prealloc space that is having | |
3487 | * minimal distance from the goal block. | |
3488 | */ | |
6be2ded1 AK |
3489 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3490 | rcu_read_lock(); | |
3491 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3492 | pa_inode_list) { | |
3493 | spin_lock(&pa->pa_lock); | |
3494 | if (pa->pa_deleted == 0 && | |
3495 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3496 | |
3497 | cpa = ext4_mb_check_group_pa(goal_block, | |
3498 | pa, cpa); | |
6be2ded1 | 3499 | } |
c9de560d | 3500 | spin_unlock(&pa->pa_lock); |
c9de560d | 3501 | } |
6be2ded1 | 3502 | rcu_read_unlock(); |
c9de560d | 3503 | } |
5e745b04 AK |
3504 | if (cpa) { |
3505 | ext4_mb_use_group_pa(ac, cpa); | |
3506 | ac->ac_criteria = 20; | |
3507 | return 1; | |
3508 | } | |
c9de560d AT |
3509 | return 0; |
3510 | } | |
3511 | ||
7a2fcbf7 AK |
3512 | /* |
3513 | * the function goes through all block freed in the group | |
3514 | * but not yet committed and marks them used in in-core bitmap. | |
3515 | * buddy must be generated from this bitmap | |
955ce5f5 | 3516 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3517 | */ |
3518 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3519 | ext4_group_t group) | |
3520 | { | |
3521 | struct rb_node *n; | |
3522 | struct ext4_group_info *grp; | |
3523 | struct ext4_free_data *entry; | |
3524 | ||
3525 | grp = ext4_get_group_info(sb, group); | |
3526 | n = rb_first(&(grp->bb_free_root)); | |
3527 | ||
3528 | while (n) { | |
18aadd47 BJ |
3529 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3530 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3531 | n = rb_next(n); |
3532 | } | |
3533 | return; | |
3534 | } | |
3535 | ||
c9de560d AT |
3536 | /* |
3537 | * the function goes through all preallocation in this group and marks them | |
3538 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3539 | * Need to be called with ext4 group lock held |
c9de560d | 3540 | */ |
089ceecc ES |
3541 | static noinline_for_stack |
3542 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3543 | ext4_group_t group) |
3544 | { | |
3545 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3546 | struct ext4_prealloc_space *pa; | |
3547 | struct list_head *cur; | |
3548 | ext4_group_t groupnr; | |
3549 | ext4_grpblk_t start; | |
3550 | int preallocated = 0; | |
c9de560d AT |
3551 | int len; |
3552 | ||
3553 | /* all form of preallocation discards first load group, | |
3554 | * so the only competing code is preallocation use. | |
3555 | * we don't need any locking here | |
3556 | * notice we do NOT ignore preallocations with pa_deleted | |
3557 | * otherwise we could leave used blocks available for | |
3558 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3559 | * is dropping preallocation | |
3560 | */ | |
3561 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3562 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3563 | spin_lock(&pa->pa_lock); | |
3564 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3565 | &groupnr, &start); | |
3566 | len = pa->pa_len; | |
3567 | spin_unlock(&pa->pa_lock); | |
3568 | if (unlikely(len == 0)) | |
3569 | continue; | |
3570 | BUG_ON(groupnr != group); | |
c3e94d1d | 3571 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3572 | preallocated += len; |
c9de560d | 3573 | } |
ff950156 | 3574 | mb_debug(1, "preallocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3575 | } |
3576 | ||
3577 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3578 | { | |
3579 | struct ext4_prealloc_space *pa; | |
3580 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
3581 | |
3582 | BUG_ON(atomic_read(&pa->pa_count)); | |
3583 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
3584 | kmem_cache_free(ext4_pspace_cachep, pa); |
3585 | } | |
3586 | ||
3587 | /* | |
3588 | * drops a reference to preallocated space descriptor | |
3589 | * if this was the last reference and the space is consumed | |
3590 | */ | |
3591 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3592 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3593 | { | |
a9df9a49 | 3594 | ext4_group_t grp; |
d33a1976 | 3595 | ext4_fsblk_t grp_blk; |
c9de560d | 3596 | |
c9de560d AT |
3597 | /* in this short window concurrent discard can set pa_deleted */ |
3598 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
3599 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
3600 | spin_unlock(&pa->pa_lock); | |
3601 | return; | |
3602 | } | |
3603 | ||
c9de560d AT |
3604 | if (pa->pa_deleted == 1) { |
3605 | spin_unlock(&pa->pa_lock); | |
3606 | return; | |
3607 | } | |
3608 | ||
3609 | pa->pa_deleted = 1; | |
3610 | spin_unlock(&pa->pa_lock); | |
3611 | ||
d33a1976 | 3612 | grp_blk = pa->pa_pstart; |
60e6679e | 3613 | /* |
cc0fb9ad AK |
3614 | * If doing group-based preallocation, pa_pstart may be in the |
3615 | * next group when pa is used up | |
3616 | */ | |
3617 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3618 | grp_blk--; |
3619 | ||
bd86298e | 3620 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
3621 | |
3622 | /* | |
3623 | * possible race: | |
3624 | * | |
3625 | * P1 (buddy init) P2 (regular allocation) | |
3626 | * find block B in PA | |
3627 | * copy on-disk bitmap to buddy | |
3628 | * mark B in on-disk bitmap | |
3629 | * drop PA from group | |
3630 | * mark all PAs in buddy | |
3631 | * | |
3632 | * thus, P1 initializes buddy with B available. to prevent this | |
3633 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3634 | * against that pair | |
3635 | */ | |
3636 | ext4_lock_group(sb, grp); | |
3637 | list_del(&pa->pa_group_list); | |
3638 | ext4_unlock_group(sb, grp); | |
3639 | ||
3640 | spin_lock(pa->pa_obj_lock); | |
3641 | list_del_rcu(&pa->pa_inode_list); | |
3642 | spin_unlock(pa->pa_obj_lock); | |
3643 | ||
3644 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3645 | } | |
3646 | ||
3647 | /* | |
3648 | * creates new preallocated space for given inode | |
3649 | */ | |
4ddfef7b ES |
3650 | static noinline_for_stack int |
3651 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3652 | { |
3653 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3654 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3655 | struct ext4_prealloc_space *pa; |
3656 | struct ext4_group_info *grp; | |
3657 | struct ext4_inode_info *ei; | |
3658 | ||
3659 | /* preallocate only when found space is larger then requested */ | |
3660 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3661 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3662 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3663 | ||
3664 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3665 | if (pa == NULL) | |
3666 | return -ENOMEM; | |
3667 | ||
3668 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3669 | int winl; | |
3670 | int wins; | |
3671 | int win; | |
3672 | int offs; | |
3673 | ||
3674 | /* we can't allocate as much as normalizer wants. | |
3675 | * so, found space must get proper lstart | |
3676 | * to cover original request */ | |
3677 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3678 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3679 | ||
3680 | /* we're limited by original request in that | |
3681 | * logical block must be covered any way | |
3682 | * winl is window we can move our chunk within */ | |
3683 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3684 | ||
3685 | /* also, we should cover whole original request */ | |
53accfa9 | 3686 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3687 | |
3688 | /* the smallest one defines real window */ | |
3689 | win = min(winl, wins); | |
3690 | ||
53accfa9 TT |
3691 | offs = ac->ac_o_ex.fe_logical % |
3692 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3693 | if (offs && offs < win) |
3694 | win = offs; | |
3695 | ||
53accfa9 | 3696 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3697 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3698 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3699 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3700 | } | |
3701 | ||
3702 | /* preallocation can change ac_b_ex, thus we store actually | |
3703 | * allocated blocks for history */ | |
3704 | ac->ac_f_ex = ac->ac_b_ex; | |
3705 | ||
3706 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3707 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3708 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3709 | pa->pa_free = pa->pa_len; | |
3710 | atomic_set(&pa->pa_count, 1); | |
3711 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3712 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3713 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3714 | pa->pa_deleted = 0; |
cc0fb9ad | 3715 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3716 | |
6ba495e9 | 3717 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3718 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3719 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3720 | |
3721 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3722 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3723 | |
3724 | ei = EXT4_I(ac->ac_inode); | |
3725 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3726 | ||
3727 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3728 | pa->pa_inode = ac->ac_inode; | |
3729 | ||
3730 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3731 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3732 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3733 | ||
3734 | spin_lock(pa->pa_obj_lock); | |
3735 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3736 | spin_unlock(pa->pa_obj_lock); | |
3737 | ||
3738 | return 0; | |
3739 | } | |
3740 | ||
3741 | /* | |
3742 | * creates new preallocated space for locality group inodes belongs to | |
3743 | */ | |
4ddfef7b ES |
3744 | static noinline_for_stack int |
3745 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3746 | { |
3747 | struct super_block *sb = ac->ac_sb; | |
3748 | struct ext4_locality_group *lg; | |
3749 | struct ext4_prealloc_space *pa; | |
3750 | struct ext4_group_info *grp; | |
3751 | ||
3752 | /* preallocate only when found space is larger then requested */ | |
3753 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3754 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3755 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3756 | ||
3757 | BUG_ON(ext4_pspace_cachep == NULL); | |
3758 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3759 | if (pa == NULL) | |
3760 | return -ENOMEM; | |
3761 | ||
3762 | /* preallocation can change ac_b_ex, thus we store actually | |
3763 | * allocated blocks for history */ | |
3764 | ac->ac_f_ex = ac->ac_b_ex; | |
3765 | ||
3766 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3767 | pa->pa_lstart = pa->pa_pstart; | |
3768 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3769 | pa->pa_free = pa->pa_len; | |
3770 | atomic_set(&pa->pa_count, 1); | |
3771 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3772 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3773 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3774 | pa->pa_deleted = 0; |
cc0fb9ad | 3775 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3776 | |
6ba495e9 | 3777 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3778 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3779 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3780 | |
3781 | ext4_mb_use_group_pa(ac, pa); | |
3782 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3783 | ||
3784 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3785 | lg = ac->ac_lg; | |
3786 | BUG_ON(lg == NULL); | |
3787 | ||
3788 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3789 | pa->pa_inode = NULL; | |
3790 | ||
3791 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3792 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3793 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3794 | ||
6be2ded1 AK |
3795 | /* |
3796 | * We will later add the new pa to the right bucket | |
3797 | * after updating the pa_free in ext4_mb_release_context | |
3798 | */ | |
c9de560d AT |
3799 | return 0; |
3800 | } | |
3801 | ||
3802 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3803 | { | |
3804 | int err; | |
3805 | ||
3806 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3807 | err = ext4_mb_new_group_pa(ac); | |
3808 | else | |
3809 | err = ext4_mb_new_inode_pa(ac); | |
3810 | return err; | |
3811 | } | |
3812 | ||
3813 | /* | |
3814 | * finds all unused blocks in on-disk bitmap, frees them in | |
3815 | * in-core bitmap and buddy. | |
3816 | * @pa must be unlinked from inode and group lists, so that | |
3817 | * nobody else can find/use it. | |
3818 | * the caller MUST hold group/inode locks. | |
3819 | * TODO: optimize the case when there are no in-core structures yet | |
3820 | */ | |
4ddfef7b ES |
3821 | static noinline_for_stack int |
3822 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3823 | struct ext4_prealloc_space *pa) |
c9de560d | 3824 | { |
c9de560d AT |
3825 | struct super_block *sb = e4b->bd_sb; |
3826 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3827 | unsigned int end; |
3828 | unsigned int next; | |
c9de560d AT |
3829 | ext4_group_t group; |
3830 | ext4_grpblk_t bit; | |
ba80b101 | 3831 | unsigned long long grp_blk_start; |
c9de560d AT |
3832 | int free = 0; |
3833 | ||
3834 | BUG_ON(pa->pa_deleted == 0); | |
3835 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3836 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3837 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3838 | end = bit + pa->pa_len; | |
3839 | ||
c9de560d | 3840 | while (bit < end) { |
ffad0a44 | 3841 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3842 | if (bit >= end) |
3843 | break; | |
ffad0a44 | 3844 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3845 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3846 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3847 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3848 | free += next - bit; |
3849 | ||
3e1e5f50 | 3850 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3851 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3852 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3853 | next - bit); |
c9de560d AT |
3854 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3855 | bit = next + 1; | |
3856 | } | |
3857 | if (free != pa->pa_free) { | |
9d8b9ec4 TT |
3858 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
3859 | "pa %p: logic %lu, phys. %lu, len %lu", | |
3860 | pa, (unsigned long) pa->pa_lstart, | |
3861 | (unsigned long) pa->pa_pstart, | |
3862 | (unsigned long) pa->pa_len); | |
e29136f8 | 3863 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3864 | free, pa->pa_free); |
e56eb659 AK |
3865 | /* |
3866 | * pa is already deleted so we use the value obtained | |
3867 | * from the bitmap and continue. | |
3868 | */ | |
c9de560d | 3869 | } |
c9de560d AT |
3870 | atomic_add(free, &sbi->s_mb_discarded); |
3871 | ||
863c37fc | 3872 | return 0; |
c9de560d AT |
3873 | } |
3874 | ||
4ddfef7b ES |
3875 | static noinline_for_stack int |
3876 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3877 | struct ext4_prealloc_space *pa) |
c9de560d | 3878 | { |
c9de560d AT |
3879 | struct super_block *sb = e4b->bd_sb; |
3880 | ext4_group_t group; | |
3881 | ext4_grpblk_t bit; | |
3882 | ||
60e07cf5 | 3883 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
3884 | BUG_ON(pa->pa_deleted == 0); |
3885 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3886 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3887 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3888 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3889 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3890 | |
3891 | return 0; | |
3892 | } | |
3893 | ||
3894 | /* | |
3895 | * releases all preallocations in given group | |
3896 | * | |
3897 | * first, we need to decide discard policy: | |
3898 | * - when do we discard | |
3899 | * 1) ENOSPC | |
3900 | * - how many do we discard | |
3901 | * 1) how many requested | |
3902 | */ | |
4ddfef7b ES |
3903 | static noinline_for_stack int |
3904 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3905 | ext4_group_t group, int needed) |
3906 | { | |
3907 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3908 | struct buffer_head *bitmap_bh = NULL; | |
3909 | struct ext4_prealloc_space *pa, *tmp; | |
3910 | struct list_head list; | |
3911 | struct ext4_buddy e4b; | |
3912 | int err; | |
3913 | int busy = 0; | |
3914 | int free = 0; | |
3915 | ||
6ba495e9 | 3916 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3917 | |
3918 | if (list_empty(&grp->bb_prealloc_list)) | |
3919 | return 0; | |
3920 | ||
574ca174 | 3921 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
3922 | if (IS_ERR(bitmap_bh)) { |
3923 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
3924 | ext4_error_err(sb, -err, |
3925 | "Error %d reading block bitmap for %u", | |
3926 | err, group); | |
ce89f46c | 3927 | return 0; |
c9de560d AT |
3928 | } |
3929 | ||
3930 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3931 | if (err) { |
9651e6b2 KK |
3932 | ext4_warning(sb, "Error %d loading buddy information for %u", |
3933 | err, group); | |
ce89f46c AK |
3934 | put_bh(bitmap_bh); |
3935 | return 0; | |
3936 | } | |
c9de560d AT |
3937 | |
3938 | if (needed == 0) | |
7137d7a4 | 3939 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3940 | |
c9de560d | 3941 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3942 | repeat: |
3943 | ext4_lock_group(sb, group); | |
3944 | list_for_each_entry_safe(pa, tmp, | |
3945 | &grp->bb_prealloc_list, pa_group_list) { | |
3946 | spin_lock(&pa->pa_lock); | |
3947 | if (atomic_read(&pa->pa_count)) { | |
3948 | spin_unlock(&pa->pa_lock); | |
3949 | busy = 1; | |
3950 | continue; | |
3951 | } | |
3952 | if (pa->pa_deleted) { | |
3953 | spin_unlock(&pa->pa_lock); | |
3954 | continue; | |
3955 | } | |
3956 | ||
3957 | /* seems this one can be freed ... */ | |
3958 | pa->pa_deleted = 1; | |
3959 | ||
3960 | /* we can trust pa_free ... */ | |
3961 | free += pa->pa_free; | |
3962 | ||
3963 | spin_unlock(&pa->pa_lock); | |
3964 | ||
3965 | list_del(&pa->pa_group_list); | |
3966 | list_add(&pa->u.pa_tmp_list, &list); | |
3967 | } | |
3968 | ||
3969 | /* if we still need more blocks and some PAs were used, try again */ | |
3970 | if (free < needed && busy) { | |
3971 | busy = 0; | |
3972 | ext4_unlock_group(sb, group); | |
bb8b20ed | 3973 | cond_resched(); |
c9de560d AT |
3974 | goto repeat; |
3975 | } | |
3976 | ||
3977 | /* found anything to free? */ | |
3978 | if (list_empty(&list)) { | |
3979 | BUG_ON(free != 0); | |
3980 | goto out; | |
3981 | } | |
3982 | ||
3983 | /* now free all selected PAs */ | |
3984 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3985 | ||
3986 | /* remove from object (inode or locality group) */ | |
3987 | spin_lock(pa->pa_obj_lock); | |
3988 | list_del_rcu(&pa->pa_inode_list); | |
3989 | spin_unlock(pa->pa_obj_lock); | |
3990 | ||
cc0fb9ad | 3991 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 3992 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 3993 | else |
3e1e5f50 | 3994 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3995 | |
3996 | list_del(&pa->u.pa_tmp_list); | |
3997 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3998 | } | |
3999 | ||
4000 | out: | |
4001 | ext4_unlock_group(sb, group); | |
e39e07fd | 4002 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4003 | put_bh(bitmap_bh); |
4004 | return free; | |
4005 | } | |
4006 | ||
4007 | /* | |
4008 | * releases all non-used preallocated blocks for given inode | |
4009 | * | |
4010 | * It's important to discard preallocations under i_data_sem | |
4011 | * We don't want another block to be served from the prealloc | |
4012 | * space when we are discarding the inode prealloc space. | |
4013 | * | |
4014 | * FIXME!! Make sure it is valid at all the call sites | |
4015 | */ | |
c2ea3fde | 4016 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
4017 | { |
4018 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4019 | struct super_block *sb = inode->i_sb; | |
4020 | struct buffer_head *bitmap_bh = NULL; | |
4021 | struct ext4_prealloc_space *pa, *tmp; | |
4022 | ext4_group_t group = 0; | |
4023 | struct list_head list; | |
4024 | struct ext4_buddy e4b; | |
4025 | int err; | |
4026 | ||
c2ea3fde | 4027 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4028 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4029 | return; | |
4030 | } | |
4031 | ||
6ba495e9 | 4032 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 4033 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
4034 | |
4035 | INIT_LIST_HEAD(&list); | |
4036 | ||
4037 | repeat: | |
4038 | /* first, collect all pa's in the inode */ | |
4039 | spin_lock(&ei->i_prealloc_lock); | |
4040 | while (!list_empty(&ei->i_prealloc_list)) { | |
4041 | pa = list_entry(ei->i_prealloc_list.next, | |
4042 | struct ext4_prealloc_space, pa_inode_list); | |
4043 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4044 | spin_lock(&pa->pa_lock); | |
4045 | if (atomic_read(&pa->pa_count)) { | |
4046 | /* this shouldn't happen often - nobody should | |
4047 | * use preallocation while we're discarding it */ | |
4048 | spin_unlock(&pa->pa_lock); | |
4049 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
4050 | ext4_msg(sb, KERN_ERR, |
4051 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
4052 | WARN_ON(1); |
4053 | schedule_timeout_uninterruptible(HZ); | |
4054 | goto repeat; | |
4055 | ||
4056 | } | |
4057 | if (pa->pa_deleted == 0) { | |
4058 | pa->pa_deleted = 1; | |
4059 | spin_unlock(&pa->pa_lock); | |
4060 | list_del_rcu(&pa->pa_inode_list); | |
4061 | list_add(&pa->u.pa_tmp_list, &list); | |
4062 | continue; | |
4063 | } | |
4064 | ||
4065 | /* someone is deleting pa right now */ | |
4066 | spin_unlock(&pa->pa_lock); | |
4067 | spin_unlock(&ei->i_prealloc_lock); | |
4068 | ||
4069 | /* we have to wait here because pa_deleted | |
4070 | * doesn't mean pa is already unlinked from | |
4071 | * the list. as we might be called from | |
4072 | * ->clear_inode() the inode will get freed | |
4073 | * and concurrent thread which is unlinking | |
4074 | * pa from inode's list may access already | |
4075 | * freed memory, bad-bad-bad */ | |
4076 | ||
4077 | /* XXX: if this happens too often, we can | |
4078 | * add a flag to force wait only in case | |
4079 | * of ->clear_inode(), but not in case of | |
4080 | * regular truncate */ | |
4081 | schedule_timeout_uninterruptible(HZ); | |
4082 | goto repeat; | |
4083 | } | |
4084 | spin_unlock(&ei->i_prealloc_lock); | |
4085 | ||
4086 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4087 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 4088 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 4089 | |
9651e6b2 KK |
4090 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4091 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 4092 | if (err) { |
54d3adbc TT |
4093 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
4094 | err, group); | |
ce89f46c AK |
4095 | continue; |
4096 | } | |
c9de560d | 4097 | |
574ca174 | 4098 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4099 | if (IS_ERR(bitmap_bh)) { |
4100 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4101 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
4102 | err, group); | |
e39e07fd | 4103 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 4104 | continue; |
c9de560d AT |
4105 | } |
4106 | ||
4107 | ext4_lock_group(sb, group); | |
4108 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4109 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4110 | ext4_unlock_group(sb, group); |
4111 | ||
e39e07fd | 4112 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4113 | put_bh(bitmap_bh); |
4114 | ||
4115 | list_del(&pa->u.pa_tmp_list); | |
4116 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4117 | } | |
4118 | } | |
4119 | ||
6ba495e9 | 4120 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
4121 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
4122 | { | |
4123 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 4124 | ext4_group_t ngroups, i; |
c9de560d | 4125 | |
a0b30c12 | 4126 | if (!ext4_mballoc_debug || |
4dd89fc6 | 4127 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) |
e3570639 ES |
4128 | return; |
4129 | ||
7f6a11e7 | 4130 | ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" |
9d8b9ec4 | 4131 | " Allocation context details:"); |
7f6a11e7 | 4132 | ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", |
c9de560d | 4133 | ac->ac_status, ac->ac_flags); |
7f6a11e7 | 4134 | ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " |
9d8b9ec4 TT |
4135 | "goal %lu/%lu/%lu@%lu, " |
4136 | "best %lu/%lu/%lu@%lu cr %d", | |
c9de560d AT |
4137 | (unsigned long)ac->ac_o_ex.fe_group, |
4138 | (unsigned long)ac->ac_o_ex.fe_start, | |
4139 | (unsigned long)ac->ac_o_ex.fe_len, | |
4140 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4141 | (unsigned long)ac->ac_g_ex.fe_group, | |
4142 | (unsigned long)ac->ac_g_ex.fe_start, | |
4143 | (unsigned long)ac->ac_g_ex.fe_len, | |
4144 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4145 | (unsigned long)ac->ac_b_ex.fe_group, | |
4146 | (unsigned long)ac->ac_b_ex.fe_start, | |
4147 | (unsigned long)ac->ac_b_ex.fe_len, | |
4148 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4149 | (int)ac->ac_criteria); | |
dc9ddd98 | 4150 | ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); |
7f6a11e7 | 4151 | ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); |
8df9675f TT |
4152 | ngroups = ext4_get_groups_count(sb); |
4153 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
4154 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4155 | struct ext4_prealloc_space *pa; | |
4156 | ext4_grpblk_t start; | |
4157 | struct list_head *cur; | |
4158 | ext4_lock_group(sb, i); | |
4159 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4160 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4161 | pa_group_list); | |
4162 | spin_lock(&pa->pa_lock); | |
4163 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4164 | NULL, &start); | |
4165 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
4166 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
4167 | start, pa->pa_len); | |
c9de560d | 4168 | } |
60bd63d1 | 4169 | ext4_unlock_group(sb, i); |
c9de560d AT |
4170 | |
4171 | if (grp->bb_free == 0) | |
4172 | continue; | |
1c718505 | 4173 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
4174 | i, grp->bb_free, grp->bb_fragments); |
4175 | } | |
4176 | printk(KERN_ERR "\n"); | |
4177 | } | |
4178 | #else | |
4179 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4180 | { | |
4181 | return; | |
4182 | } | |
4183 | #endif | |
4184 | ||
4185 | /* | |
4186 | * We use locality group preallocation for small size file. The size of the | |
4187 | * file is determined by the current size or the resulting size after | |
4188 | * allocation which ever is larger | |
4189 | * | |
b713a5ec | 4190 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4191 | */ |
4192 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4193 | { | |
4194 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4195 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4196 | loff_t size, isize; | |
4197 | ||
4198 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4199 | return; | |
4200 | ||
4ba74d00 TT |
4201 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4202 | return; | |
4203 | ||
53accfa9 | 4204 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
4205 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4206 | >> bsbits; | |
c9de560d | 4207 | |
82dd124c NB |
4208 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
4209 | !inode_is_open_for_write(ac->ac_inode)) { | |
50797481 TT |
4210 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; |
4211 | return; | |
4212 | } | |
4213 | ||
ebbe0277 RD |
4214 | if (sbi->s_mb_group_prealloc <= 0) { |
4215 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
4216 | return; | |
4217 | } | |
4218 | ||
c9de560d | 4219 | /* don't use group allocation for large files */ |
71780577 | 4220 | size = max(size, isize); |
cc483f10 | 4221 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 4222 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 4223 | return; |
4ba74d00 | 4224 | } |
c9de560d AT |
4225 | |
4226 | BUG_ON(ac->ac_lg != NULL); | |
4227 | /* | |
4228 | * locality group prealloc space are per cpu. The reason for having | |
4229 | * per cpu locality group is to reduce the contention between block | |
4230 | * request from multiple CPUs. | |
4231 | */ | |
a0b6bc63 | 4232 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
4233 | |
4234 | /* we're going to use group allocation */ | |
4235 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4236 | ||
4237 | /* serialize all allocations in the group */ | |
4238 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4239 | } | |
4240 | ||
4ddfef7b ES |
4241 | static noinline_for_stack int |
4242 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4243 | struct ext4_allocation_request *ar) |
4244 | { | |
4245 | struct super_block *sb = ar->inode->i_sb; | |
4246 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4247 | struct ext4_super_block *es = sbi->s_es; | |
4248 | ext4_group_t group; | |
498e5f24 TT |
4249 | unsigned int len; |
4250 | ext4_fsblk_t goal; | |
c9de560d AT |
4251 | ext4_grpblk_t block; |
4252 | ||
4253 | /* we can't allocate > group size */ | |
4254 | len = ar->len; | |
4255 | ||
4256 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
4257 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
4258 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4259 | |
4260 | /* start searching from the goal */ | |
4261 | goal = ar->goal; | |
4262 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4263 | goal >= ext4_blocks_count(es)) | |
4264 | goal = le32_to_cpu(es->s_first_data_block); | |
4265 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4266 | ||
4267 | /* set up allocation goals */ | |
f5a44db5 | 4268 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 4269 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4270 | ac->ac_sb = sb; |
4271 | ac->ac_inode = ar->inode; | |
53accfa9 | 4272 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4273 | ac->ac_o_ex.fe_group = group; |
4274 | ac->ac_o_ex.fe_start = block; | |
4275 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4276 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4277 | ac->ac_flags = ar->flags; |
c9de560d AT |
4278 | |
4279 | /* we have to define context: we'll we work with a file or | |
4280 | * locality group. this is a policy, actually */ | |
4281 | ext4_mb_group_or_file(ac); | |
4282 | ||
6ba495e9 | 4283 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4284 | "left: %u/%u, right %u/%u to %swritable\n", |
4285 | (unsigned) ar->len, (unsigned) ar->logical, | |
4286 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4287 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4288 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 4289 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
4290 | return 0; |
4291 | ||
4292 | } | |
4293 | ||
6be2ded1 AK |
4294 | static noinline_for_stack void |
4295 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4296 | struct ext4_locality_group *lg, | |
4297 | int order, int total_entries) | |
4298 | { | |
4299 | ext4_group_t group = 0; | |
4300 | struct ext4_buddy e4b; | |
4301 | struct list_head discard_list; | |
4302 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4303 | |
6ba495e9 | 4304 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4305 | |
4306 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4307 | |
4308 | spin_lock(&lg->lg_prealloc_lock); | |
4309 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
92e9c58c MB |
4310 | pa_inode_list, |
4311 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
4312 | spin_lock(&pa->pa_lock); |
4313 | if (atomic_read(&pa->pa_count)) { | |
4314 | /* | |
4315 | * This is the pa that we just used | |
4316 | * for block allocation. So don't | |
4317 | * free that | |
4318 | */ | |
4319 | spin_unlock(&pa->pa_lock); | |
4320 | continue; | |
4321 | } | |
4322 | if (pa->pa_deleted) { | |
4323 | spin_unlock(&pa->pa_lock); | |
4324 | continue; | |
4325 | } | |
4326 | /* only lg prealloc space */ | |
cc0fb9ad | 4327 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4328 | |
4329 | /* seems this one can be freed ... */ | |
4330 | pa->pa_deleted = 1; | |
4331 | spin_unlock(&pa->pa_lock); | |
4332 | ||
4333 | list_del_rcu(&pa->pa_inode_list); | |
4334 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4335 | ||
4336 | total_entries--; | |
4337 | if (total_entries <= 5) { | |
4338 | /* | |
4339 | * we want to keep only 5 entries | |
4340 | * allowing it to grow to 8. This | |
4341 | * mak sure we don't call discard | |
4342 | * soon for this list. | |
4343 | */ | |
4344 | break; | |
4345 | } | |
4346 | } | |
4347 | spin_unlock(&lg->lg_prealloc_lock); | |
4348 | ||
4349 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 4350 | int err; |
6be2ded1 | 4351 | |
bd86298e | 4352 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
4353 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4354 | GFP_NOFS|__GFP_NOFAIL); | |
4355 | if (err) { | |
54d3adbc TT |
4356 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
4357 | err, group); | |
6be2ded1 AK |
4358 | continue; |
4359 | } | |
4360 | ext4_lock_group(sb, group); | |
4361 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4362 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4363 | ext4_unlock_group(sb, group); |
4364 | ||
e39e07fd | 4365 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4366 | list_del(&pa->u.pa_tmp_list); |
4367 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4368 | } | |
6be2ded1 AK |
4369 | } |
4370 | ||
4371 | /* | |
4372 | * We have incremented pa_count. So it cannot be freed at this | |
4373 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4374 | * possible from this lg. That means pa_free cannot be updated. | |
4375 | * | |
4376 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4377 | * which can cause the lg_prealloc_list to be updated. | |
4378 | */ | |
4379 | ||
4380 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4381 | { | |
4382 | int order, added = 0, lg_prealloc_count = 1; | |
4383 | struct super_block *sb = ac->ac_sb; | |
4384 | struct ext4_locality_group *lg = ac->ac_lg; | |
4385 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4386 | ||
4387 | order = fls(pa->pa_free) - 1; | |
4388 | if (order > PREALLOC_TB_SIZE - 1) | |
4389 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4390 | order = PREALLOC_TB_SIZE - 1; | |
4391 | /* Add the prealloc space to lg */ | |
f1167009 | 4392 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 4393 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
92e9c58c MB |
4394 | pa_inode_list, |
4395 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
4396 | spin_lock(&tmp_pa->pa_lock); |
4397 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4398 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4399 | continue; |
4400 | } | |
4401 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4402 | /* Add to the tail of the previous entry */ | |
4403 | list_add_tail_rcu(&pa->pa_inode_list, | |
4404 | &tmp_pa->pa_inode_list); | |
4405 | added = 1; | |
4406 | /* | |
4407 | * we want to count the total | |
4408 | * number of entries in the list | |
4409 | */ | |
4410 | } | |
4411 | spin_unlock(&tmp_pa->pa_lock); | |
4412 | lg_prealloc_count++; | |
4413 | } | |
4414 | if (!added) | |
4415 | list_add_tail_rcu(&pa->pa_inode_list, | |
4416 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4417 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4418 | |
4419 | /* Now trim the list to be not more than 8 elements */ | |
4420 | if (lg_prealloc_count > 8) { | |
4421 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4422 | order, lg_prealloc_count); |
6be2ded1 AK |
4423 | return; |
4424 | } | |
4425 | return ; | |
4426 | } | |
4427 | ||
c9de560d AT |
4428 | /* |
4429 | * release all resource we used in allocation | |
4430 | */ | |
4431 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4432 | { | |
53accfa9 | 4433 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4434 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4435 | if (pa) { | |
cc0fb9ad | 4436 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4437 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4438 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4439 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4440 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4441 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4442 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4443 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4444 | } |
c9de560d | 4445 | } |
ba443916 AK |
4446 | if (pa) { |
4447 | /* | |
4448 | * We want to add the pa to the right bucket. | |
4449 | * Remove it from the list and while adding | |
4450 | * make sure the list to which we are adding | |
44183d42 | 4451 | * doesn't grow big. |
ba443916 | 4452 | */ |
cc0fb9ad | 4453 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4454 | spin_lock(pa->pa_obj_lock); |
4455 | list_del_rcu(&pa->pa_inode_list); | |
4456 | spin_unlock(pa->pa_obj_lock); | |
4457 | ext4_mb_add_n_trim(ac); | |
4458 | } | |
4459 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4460 | } | |
c9de560d | 4461 | if (ac->ac_bitmap_page) |
09cbfeaf | 4462 | put_page(ac->ac_bitmap_page); |
c9de560d | 4463 | if (ac->ac_buddy_page) |
09cbfeaf | 4464 | put_page(ac->ac_buddy_page); |
c9de560d AT |
4465 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
4466 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4467 | ext4_mb_collect_stats(ac); | |
4468 | return 0; | |
4469 | } | |
4470 | ||
4471 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4472 | { | |
8df9675f | 4473 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4474 | int ret; |
4475 | int freed = 0; | |
4476 | ||
9bffad1e | 4477 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4478 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4479 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4480 | freed += ret; | |
4481 | needed -= ret; | |
4482 | } | |
4483 | ||
4484 | return freed; | |
4485 | } | |
4486 | ||
4487 | /* | |
4488 | * Main entry point into mballoc to allocate blocks | |
4489 | * it tries to use preallocation first, then falls back | |
4490 | * to usual allocation | |
4491 | */ | |
4492 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4493 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4494 | { |
6bc6e63f | 4495 | int freed; |
256bdb49 | 4496 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4497 | struct ext4_sb_info *sbi; |
4498 | struct super_block *sb; | |
4499 | ext4_fsblk_t block = 0; | |
60e58e0f | 4500 | unsigned int inquota = 0; |
53accfa9 | 4501 | unsigned int reserv_clstrs = 0; |
c9de560d | 4502 | |
b10a44c3 | 4503 | might_sleep(); |
c9de560d AT |
4504 | sb = ar->inode->i_sb; |
4505 | sbi = EXT4_SB(sb); | |
4506 | ||
9bffad1e | 4507 | trace_ext4_request_blocks(ar); |
ba80b101 | 4508 | |
45dc63e7 | 4509 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 4510 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
4511 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
4512 | ||
e3cf5d5d | 4513 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
4514 | /* Without delayed allocation we need to verify |
4515 | * there is enough free blocks to do block allocation | |
4516 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4517 | */ |
55f020db | 4518 | while (ar->len && |
e7d5f315 | 4519 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4520 | |
030ba6bc | 4521 | /* let others to free the space */ |
bb8b20ed | 4522 | cond_resched(); |
030ba6bc AK |
4523 | ar->len = ar->len >> 1; |
4524 | } | |
4525 | if (!ar->len) { | |
a30d542a AK |
4526 | *errp = -ENOSPC; |
4527 | return 0; | |
4528 | } | |
53accfa9 | 4529 | reserv_clstrs = ar->len; |
55f020db | 4530 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4531 | dquot_alloc_block_nofail(ar->inode, |
4532 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4533 | } else { |
4534 | while (ar->len && | |
53accfa9 TT |
4535 | dquot_alloc_block(ar->inode, |
4536 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4537 | |
4538 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4539 | ar->len--; | |
4540 | } | |
60e58e0f MC |
4541 | } |
4542 | inquota = ar->len; | |
4543 | if (ar->len == 0) { | |
4544 | *errp = -EDQUOT; | |
6c7a120a | 4545 | goto out; |
60e58e0f | 4546 | } |
07031431 | 4547 | } |
d2a17637 | 4548 | |
85556c9a | 4549 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4550 | if (!ac) { |
363d4251 | 4551 | ar->len = 0; |
256bdb49 | 4552 | *errp = -ENOMEM; |
6c7a120a | 4553 | goto out; |
256bdb49 ES |
4554 | } |
4555 | ||
256bdb49 | 4556 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4557 | if (*errp) { |
4558 | ar->len = 0; | |
6c7a120a | 4559 | goto out; |
c9de560d AT |
4560 | } |
4561 | ||
256bdb49 ES |
4562 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4563 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4564 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4565 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4566 | repeat: |
4567 | /* allocate space in core */ | |
6c7a120a | 4568 | *errp = ext4_mb_regular_allocator(ac); |
2c00ef3e AK |
4569 | if (*errp) |
4570 | goto discard_and_exit; | |
c9de560d AT |
4571 | |
4572 | /* as we've just preallocated more space than | |
2c00ef3e | 4573 | * user requested originally, we store allocated |
c9de560d | 4574 | * space in a special descriptor */ |
256bdb49 | 4575 | if (ac->ac_status == AC_STATUS_FOUND && |
2c00ef3e AK |
4576 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) |
4577 | *errp = ext4_mb_new_preallocation(ac); | |
4578 | if (*errp) { | |
4579 | discard_and_exit: | |
4580 | ext4_discard_allocated_blocks(ac); | |
4581 | goto errout; | |
4582 | } | |
c9de560d | 4583 | } |
256bdb49 | 4584 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4585 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 4586 | if (*errp) { |
b844167e | 4587 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4588 | goto errout; |
4589 | } else { | |
519deca0 AK |
4590 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4591 | ar->len = ac->ac_b_ex.fe_len; | |
4592 | } | |
c9de560d | 4593 | } else { |
256bdb49 | 4594 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4595 | if (freed) |
4596 | goto repeat; | |
4597 | *errp = -ENOSPC; | |
6c7a120a AK |
4598 | } |
4599 | ||
6d138ced | 4600 | errout: |
6c7a120a | 4601 | if (*errp) { |
256bdb49 | 4602 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4603 | ar->len = 0; |
256bdb49 | 4604 | ext4_mb_show_ac(ac); |
c9de560d | 4605 | } |
256bdb49 | 4606 | ext4_mb_release_context(ac); |
6c7a120a AK |
4607 | out: |
4608 | if (ac) | |
4609 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4610 | if (inquota && ar->len < inquota) |
53accfa9 | 4611 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4612 | if (!ar->len) { |
e3cf5d5d | 4613 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 4614 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4615 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4616 | reserv_clstrs); |
0087d9fb | 4617 | } |
c9de560d | 4618 | |
9bffad1e | 4619 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4620 | |
c9de560d AT |
4621 | return block; |
4622 | } | |
c9de560d | 4623 | |
c894058d AK |
4624 | /* |
4625 | * We can merge two free data extents only if the physical blocks | |
4626 | * are contiguous, AND the extents were freed by the same transaction, | |
4627 | * AND the blocks are associated with the same group. | |
4628 | */ | |
a0154344 DJ |
4629 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
4630 | struct ext4_free_data *entry, | |
4631 | struct ext4_free_data *new_entry, | |
4632 | struct rb_root *entry_rb_root) | |
c894058d | 4633 | { |
a0154344 DJ |
4634 | if ((entry->efd_tid != new_entry->efd_tid) || |
4635 | (entry->efd_group != new_entry->efd_group)) | |
4636 | return; | |
4637 | if (entry->efd_start_cluster + entry->efd_count == | |
4638 | new_entry->efd_start_cluster) { | |
4639 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
4640 | new_entry->efd_count += entry->efd_count; | |
4641 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
4642 | entry->efd_start_cluster) { | |
4643 | new_entry->efd_count += entry->efd_count; | |
4644 | } else | |
4645 | return; | |
4646 | spin_lock(&sbi->s_md_lock); | |
4647 | list_del(&entry->efd_list); | |
4648 | spin_unlock(&sbi->s_md_lock); | |
4649 | rb_erase(&entry->efd_node, entry_rb_root); | |
4650 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
4651 | } |
4652 | ||
4ddfef7b ES |
4653 | static noinline_for_stack int |
4654 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4655 | struct ext4_free_data *new_entry) |
c9de560d | 4656 | { |
e29136f8 | 4657 | ext4_group_t group = e4b->bd_group; |
84130193 | 4658 | ext4_grpblk_t cluster; |
d08854f5 | 4659 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 4660 | struct ext4_free_data *entry; |
c9de560d AT |
4661 | struct ext4_group_info *db = e4b->bd_info; |
4662 | struct super_block *sb = e4b->bd_sb; | |
4663 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4664 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4665 | struct rb_node *parent = NULL, *new_node; | |
4666 | ||
0390131b | 4667 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4668 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4669 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4670 | ||
18aadd47 BJ |
4671 | new_node = &new_entry->efd_node; |
4672 | cluster = new_entry->efd_start_cluster; | |
c894058d | 4673 | |
c894058d AK |
4674 | if (!*n) { |
4675 | /* first free block exent. We need to | |
4676 | protect buddy cache from being freed, | |
4677 | * otherwise we'll refresh it from | |
4678 | * on-disk bitmap and lose not-yet-available | |
4679 | * blocks */ | |
09cbfeaf KS |
4680 | get_page(e4b->bd_buddy_page); |
4681 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
4682 | } |
4683 | while (*n) { | |
4684 | parent = *n; | |
18aadd47 BJ |
4685 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
4686 | if (cluster < entry->efd_start_cluster) | |
c894058d | 4687 | n = &(*n)->rb_left; |
18aadd47 | 4688 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
4689 | n = &(*n)->rb_right; |
4690 | else { | |
e29136f8 | 4691 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
4692 | ext4_group_first_block_no(sb, group) + |
4693 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 4694 | "Block already on to-be-freed list"); |
c894058d | 4695 | return 0; |
c9de560d | 4696 | } |
c894058d | 4697 | } |
c9de560d | 4698 | |
c894058d AK |
4699 | rb_link_node(new_node, parent, n); |
4700 | rb_insert_color(new_node, &db->bb_free_root); | |
4701 | ||
4702 | /* Now try to see the extent can be merged to left and right */ | |
4703 | node = rb_prev(new_node); | |
4704 | if (node) { | |
18aadd47 | 4705 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
4706 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
4707 | &(db->bb_free_root)); | |
c894058d | 4708 | } |
c9de560d | 4709 | |
c894058d AK |
4710 | node = rb_next(new_node); |
4711 | if (node) { | |
18aadd47 | 4712 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
4713 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
4714 | &(db->bb_free_root)); | |
c9de560d | 4715 | } |
a0154344 | 4716 | |
d08854f5 | 4717 | spin_lock(&sbi->s_md_lock); |
a0154344 | 4718 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
4719 | sbi->s_mb_free_pending += clusters; |
4720 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
4721 | return 0; |
4722 | } | |
4723 | ||
44338711 TT |
4724 | /** |
4725 | * ext4_free_blocks() -- Free given blocks and update quota | |
4726 | * @handle: handle for this transaction | |
4727 | * @inode: inode | |
c60990b3 TT |
4728 | * @bh: optional buffer of the block to be freed |
4729 | * @block: starting physical block to be freed | |
4730 | * @count: number of blocks to be freed | |
5def1360 | 4731 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4732 | */ |
44338711 | 4733 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4734 | struct buffer_head *bh, ext4_fsblk_t block, |
4735 | unsigned long count, int flags) | |
c9de560d | 4736 | { |
26346ff6 | 4737 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4738 | struct super_block *sb = inode->i_sb; |
c9de560d | 4739 | struct ext4_group_desc *gdp; |
498e5f24 | 4740 | unsigned int overflow; |
c9de560d AT |
4741 | ext4_grpblk_t bit; |
4742 | struct buffer_head *gd_bh; | |
4743 | ext4_group_t block_group; | |
4744 | struct ext4_sb_info *sbi; | |
4745 | struct ext4_buddy e4b; | |
84130193 | 4746 | unsigned int count_clusters; |
c9de560d AT |
4747 | int err = 0; |
4748 | int ret; | |
4749 | ||
b10a44c3 | 4750 | might_sleep(); |
e6362609 TT |
4751 | if (bh) { |
4752 | if (block) | |
4753 | BUG_ON(block != bh->b_blocknr); | |
4754 | else | |
4755 | block = bh->b_blocknr; | |
4756 | } | |
c9de560d | 4757 | |
c9de560d | 4758 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4759 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4760 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4761 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4762 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4763 | goto error_return; |
4764 | } | |
4765 | ||
0610b6e9 | 4766 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4767 | trace_ext4_free_blocks(inode, block, count, flags); |
4768 | ||
9c02ac97 DJ |
4769 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
4770 | BUG_ON(count > 1); | |
e6362609 | 4771 | |
9c02ac97 DJ |
4772 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
4773 | inode, bh, block); | |
e6362609 TT |
4774 | } |
4775 | ||
84130193 TT |
4776 | /* |
4777 | * If the extent to be freed does not begin on a cluster | |
4778 | * boundary, we need to deal with partial clusters at the | |
4779 | * beginning and end of the extent. Normally we will free | |
4780 | * blocks at the beginning or the end unless we are explicitly | |
4781 | * requested to avoid doing so. | |
4782 | */ | |
f5a44db5 | 4783 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
4784 | if (overflow) { |
4785 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
4786 | overflow = sbi->s_cluster_ratio - overflow; | |
4787 | block += overflow; | |
4788 | if (count > overflow) | |
4789 | count -= overflow; | |
4790 | else | |
4791 | return; | |
4792 | } else { | |
4793 | block -= overflow; | |
4794 | count += overflow; | |
4795 | } | |
4796 | } | |
f5a44db5 | 4797 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
4798 | if (overflow) { |
4799 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
4800 | if (count > overflow) | |
4801 | count -= overflow; | |
4802 | else | |
4803 | return; | |
4804 | } else | |
4805 | count += sbi->s_cluster_ratio - overflow; | |
4806 | } | |
4807 | ||
9c02ac97 DJ |
4808 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
4809 | int i; | |
f96c450d | 4810 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; |
9c02ac97 DJ |
4811 | |
4812 | for (i = 0; i < count; i++) { | |
4813 | cond_resched(); | |
f96c450d DJ |
4814 | if (is_metadata) |
4815 | bh = sb_find_get_block(inode->i_sb, block + i); | |
4816 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
9c02ac97 DJ |
4817 | } |
4818 | } | |
4819 | ||
c9de560d AT |
4820 | do_more: |
4821 | overflow = 0; | |
4822 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4823 | ||
163a203d DW |
4824 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
4825 | ext4_get_group_info(sb, block_group)))) | |
4826 | return; | |
4827 | ||
c9de560d AT |
4828 | /* |
4829 | * Check to see if we are freeing blocks across a group | |
4830 | * boundary. | |
4831 | */ | |
84130193 TT |
4832 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4833 | overflow = EXT4_C2B(sbi, bit) + count - | |
4834 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
4835 | count -= overflow; |
4836 | } | |
810da240 | 4837 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 4838 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
4839 | if (IS_ERR(bitmap_bh)) { |
4840 | err = PTR_ERR(bitmap_bh); | |
4841 | bitmap_bh = NULL; | |
c9de560d | 4842 | goto error_return; |
ce89f46c | 4843 | } |
c9de560d | 4844 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4845 | if (!gdp) { |
4846 | err = -EIO; | |
c9de560d | 4847 | goto error_return; |
ce89f46c | 4848 | } |
c9de560d AT |
4849 | |
4850 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4851 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4852 | in_range(block, ext4_inode_table(sb, gdp), | |
49598e04 | 4853 | sbi->s_itb_per_group) || |
c9de560d | 4854 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
49598e04 | 4855 | sbi->s_itb_per_group)) { |
c9de560d | 4856 | |
12062ddd | 4857 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4858 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4859 | /* err = 0. ext4_std_error should be a no op */ |
4860 | goto error_return; | |
c9de560d AT |
4861 | } |
4862 | ||
4863 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4864 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4865 | if (err) | |
4866 | goto error_return; | |
4867 | ||
4868 | /* | |
4869 | * We are about to modify some metadata. Call the journal APIs | |
4870 | * to unshare ->b_data if a currently-committing transaction is | |
4871 | * using it | |
4872 | */ | |
4873 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4874 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4875 | if (err) | |
4876 | goto error_return; | |
c9de560d AT |
4877 | #ifdef AGGRESSIVE_CHECK |
4878 | { | |
4879 | int i; | |
84130193 | 4880 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
4881 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
4882 | } | |
4883 | #endif | |
84130193 | 4884 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 4885 | |
adb7ef60 KK |
4886 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
4887 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
4888 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
4889 | if (err) |
4890 | goto error_return; | |
e6362609 | 4891 | |
f96c450d DJ |
4892 | /* |
4893 | * We need to make sure we don't reuse the freed block until after the | |
4894 | * transaction is committed. We make an exception if the inode is to be | |
4895 | * written in writeback mode since writeback mode has weak data | |
4896 | * consistency guarantees. | |
4897 | */ | |
4898 | if (ext4_handle_valid(handle) && | |
4899 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
4900 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
4901 | struct ext4_free_data *new_entry; |
4902 | /* | |
7444a072 MH |
4903 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
4904 | * to fail. | |
7a2fcbf7 | 4905 | */ |
7444a072 MH |
4906 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
4907 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
4908 | new_entry->efd_start_cluster = bit; |
4909 | new_entry->efd_group = block_group; | |
4910 | new_entry->efd_count = count_clusters; | |
4911 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4912 | |
7a2fcbf7 | 4913 | ext4_lock_group(sb, block_group); |
84130193 | 4914 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 4915 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4916 | } else { |
7a2fcbf7 AK |
4917 | /* need to update group_info->bb_free and bitmap |
4918 | * with group lock held. generate_buddy look at | |
4919 | * them with group lock_held | |
4920 | */ | |
d71c1ae2 | 4921 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
4922 | err = ext4_issue_discard(sb, block_group, bit, count, |
4923 | NULL); | |
d71c1ae2 LC |
4924 | if (err && err != -EOPNOTSUPP) |
4925 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
4926 | " group:%d block:%d count:%lu failed" | |
4927 | " with %d", block_group, bit, count, | |
4928 | err); | |
8f9ff189 LC |
4929 | } else |
4930 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 4931 | |
955ce5f5 | 4932 | ext4_lock_group(sb, block_group); |
84130193 TT |
4933 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
4934 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
4935 | } |
4936 | ||
021b65bb TT |
4937 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
4938 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 4939 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 4940 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 4941 | ext4_unlock_group(sb, block_group); |
c9de560d | 4942 | |
772cb7c8 JS |
4943 | if (sbi->s_log_groups_per_flex) { |
4944 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 4945 | atomic64_add(count_clusters, |
7c990728 SJS |
4946 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
4947 | flex_group)->free_clusters); | |
772cb7c8 JS |
4948 | } |
4949 | ||
9fe67149 EW |
4950 | /* |
4951 | * on a bigalloc file system, defer the s_freeclusters_counter | |
4952 | * update to the caller (ext4_remove_space and friends) so they | |
4953 | * can determine if a cluster freed here should be rereserved | |
4954 | */ | |
4955 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
4956 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
4957 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
4958 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
4959 | count_clusters); | |
4960 | } | |
7d734532 JK |
4961 | |
4962 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 4963 | |
7a2fcbf7 AK |
4964 | /* We dirtied the bitmap block */ |
4965 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4966 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4967 | ||
c9de560d AT |
4968 | /* And the group descriptor block */ |
4969 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4970 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4971 | if (!err) |
4972 | err = ret; | |
4973 | ||
4974 | if (overflow && !err) { | |
4975 | block += count; | |
4976 | count = overflow; | |
4977 | put_bh(bitmap_bh); | |
4978 | goto do_more; | |
4979 | } | |
c9de560d AT |
4980 | error_return: |
4981 | brelse(bitmap_bh); | |
4982 | ext4_std_error(sb, err); | |
4983 | return; | |
4984 | } | |
7360d173 | 4985 | |
2846e820 | 4986 | /** |
0529155e | 4987 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
4988 | * @handle: handle to this transaction |
4989 | * @sb: super block | |
4907cb7b | 4990 | * @block: start physical block to add to the block group |
2846e820 AG |
4991 | * @count: number of blocks to free |
4992 | * | |
e73a347b | 4993 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 4994 | */ |
cc7365df | 4995 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
4996 | ext4_fsblk_t block, unsigned long count) |
4997 | { | |
4998 | struct buffer_head *bitmap_bh = NULL; | |
4999 | struct buffer_head *gd_bh; | |
5000 | ext4_group_t block_group; | |
5001 | ext4_grpblk_t bit; | |
5002 | unsigned int i; | |
5003 | struct ext4_group_desc *desc; | |
5004 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 5005 | struct ext4_buddy e4b; |
d77147ff | 5006 | int err = 0, ret, free_clusters_count; |
5007 | ext4_grpblk_t clusters_freed; | |
5008 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
5009 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
5010 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
5011 | |
5012 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
5013 | ||
4740b830 YY |
5014 | if (count == 0) |
5015 | return 0; | |
5016 | ||
2846e820 | 5017 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
5018 | /* |
5019 | * Check to see if we are freeing blocks across a group | |
5020 | * boundary. | |
5021 | */ | |
d77147ff | 5022 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
5023 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
5024 | block_group); |
5025 | err = -EINVAL; | |
2846e820 | 5026 | goto error_return; |
cc7365df | 5027 | } |
2cd05cc3 | 5028 | |
2846e820 | 5029 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5030 | if (IS_ERR(bitmap_bh)) { |
5031 | err = PTR_ERR(bitmap_bh); | |
5032 | bitmap_bh = NULL; | |
2846e820 | 5033 | goto error_return; |
cc7365df YY |
5034 | } |
5035 | ||
2846e820 | 5036 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
5037 | if (!desc) { |
5038 | err = -EIO; | |
2846e820 | 5039 | goto error_return; |
cc7365df | 5040 | } |
2846e820 AG |
5041 | |
5042 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
5043 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
5044 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
5045 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
5046 | sbi->s_itb_per_group)) { | |
5047 | ext4_error(sb, "Adding blocks in system zones - " | |
5048 | "Block = %llu, count = %lu", | |
5049 | block, count); | |
cc7365df | 5050 | err = -EINVAL; |
2846e820 AG |
5051 | goto error_return; |
5052 | } | |
5053 | ||
2cd05cc3 TT |
5054 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
5055 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
5056 | if (err) |
5057 | goto error_return; | |
5058 | ||
5059 | /* | |
5060 | * We are about to modify some metadata. Call the journal APIs | |
5061 | * to unshare ->b_data if a currently-committing transaction is | |
5062 | * using it | |
5063 | */ | |
5064 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
5065 | err = ext4_journal_get_write_access(handle, gd_bh); | |
5066 | if (err) | |
5067 | goto error_return; | |
e73a347b | 5068 | |
d77147ff | 5069 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 5070 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 5071 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
5072 | ext4_error(sb, "bit already cleared for block %llu", |
5073 | (ext4_fsblk_t)(block + i)); | |
5074 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
5075 | } else { | |
d77147ff | 5076 | clusters_freed++; |
2846e820 AG |
5077 | } |
5078 | } | |
e73a347b AG |
5079 | |
5080 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
5081 | if (err) | |
5082 | goto error_return; | |
5083 | ||
5084 | /* | |
5085 | * need to update group_info->bb_free and bitmap | |
5086 | * with group lock held. generate_buddy look at | |
5087 | * them with group lock_held | |
5088 | */ | |
2846e820 | 5089 | ext4_lock_group(sb, block_group); |
d77147ff | 5090 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
5091 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
5092 | free_clusters_count = clusters_freed + | |
5093 | ext4_free_group_clusters(sb, desc); | |
5094 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
79f1ba49 | 5095 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 5096 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 5097 | ext4_unlock_group(sb, block_group); |
57042651 | 5098 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 5099 | clusters_freed); |
2846e820 AG |
5100 | |
5101 | if (sbi->s_log_groups_per_flex) { | |
5102 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 5103 | atomic64_add(clusters_freed, |
7c990728 SJS |
5104 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
5105 | flex_group)->free_clusters); | |
2846e820 | 5106 | } |
e73a347b AG |
5107 | |
5108 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
5109 | |
5110 | /* We dirtied the bitmap block */ | |
5111 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5112 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5113 | ||
5114 | /* And the group descriptor block */ | |
5115 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
5116 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
5117 | if (!err) | |
5118 | err = ret; | |
5119 | ||
5120 | error_return: | |
5121 | brelse(bitmap_bh); | |
5122 | ext4_std_error(sb, err); | |
cc7365df | 5123 | return err; |
2846e820 AG |
5124 | } |
5125 | ||
7360d173 LC |
5126 | /** |
5127 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
5128 | * @sb: super block for the file system | |
5129 | * @start: starting block of the free extent in the alloc. group | |
5130 | * @count: number of blocks to TRIM | |
5131 | * @group: alloc. group we are working with | |
5132 | * @e4b: ext4 buddy for the group | |
5133 | * | |
5134 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
5135 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
5136 | * be called with under the group lock. | |
5137 | */ | |
d71c1ae2 | 5138 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
d9f34504 | 5139 | ext4_group_t group, struct ext4_buddy *e4b) |
e2cbd587 | 5140 | __releases(bitlock) |
5141 | __acquires(bitlock) | |
7360d173 LC |
5142 | { |
5143 | struct ext4_free_extent ex; | |
d71c1ae2 | 5144 | int ret = 0; |
7360d173 | 5145 | |
b3d4c2b1 TM |
5146 | trace_ext4_trim_extent(sb, group, start, count); |
5147 | ||
7360d173 LC |
5148 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
5149 | ||
5150 | ex.fe_start = start; | |
5151 | ex.fe_group = group; | |
5152 | ex.fe_len = count; | |
5153 | ||
5154 | /* | |
5155 | * Mark blocks used, so no one can reuse them while | |
5156 | * being trimmed. | |
5157 | */ | |
5158 | mb_mark_used(e4b, &ex); | |
5159 | ext4_unlock_group(sb, group); | |
a0154344 | 5160 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
5161 | ext4_lock_group(sb, group); |
5162 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 5163 | return ret; |
7360d173 LC |
5164 | } |
5165 | ||
5166 | /** | |
5167 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
5168 | * @sb: super block for file system | |
22612283 | 5169 | * @group: group to be trimmed |
7360d173 LC |
5170 | * @start: first group block to examine |
5171 | * @max: last group block to examine | |
5172 | * @minblocks: minimum extent block count | |
5173 | * | |
5174 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
5175 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
5176 | * the extent. | |
5177 | * | |
5178 | * | |
5179 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
5180 | * extents. When the free extent is found, mark it as used in group buddy | |
5181 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
5182 | * the group buddy bitmap. This is done until whole group is scanned. | |
5183 | */ | |
0b75a840 | 5184 | static ext4_grpblk_t |
78944086 LC |
5185 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
5186 | ext4_grpblk_t start, ext4_grpblk_t max, | |
5187 | ext4_grpblk_t minblocks) | |
7360d173 LC |
5188 | { |
5189 | void *bitmap; | |
169ddc3e | 5190 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 5191 | struct ext4_buddy e4b; |
d71c1ae2 | 5192 | int ret = 0; |
7360d173 | 5193 | |
b3d4c2b1 TM |
5194 | trace_ext4_trim_all_free(sb, group, start, max); |
5195 | ||
78944086 LC |
5196 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
5197 | if (ret) { | |
9651e6b2 KK |
5198 | ext4_warning(sb, "Error %d loading buddy information for %u", |
5199 | ret, group); | |
78944086 LC |
5200 | return ret; |
5201 | } | |
78944086 | 5202 | bitmap = e4b.bd_bitmap; |
28739eea LC |
5203 | |
5204 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
5205 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
5206 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
5207 | goto out; | |
5208 | ||
78944086 LC |
5209 | start = (e4b.bd_info->bb_first_free > start) ? |
5210 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 5211 | |
913eed83 LC |
5212 | while (start <= max) { |
5213 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
5214 | if (start > max) | |
7360d173 | 5215 | break; |
913eed83 | 5216 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
5217 | |
5218 | if ((next - start) >= minblocks) { | |
d71c1ae2 LC |
5219 | ret = ext4_trim_extent(sb, start, |
5220 | next - start, group, &e4b); | |
5221 | if (ret && ret != -EOPNOTSUPP) | |
5222 | break; | |
5223 | ret = 0; | |
7360d173 LC |
5224 | count += next - start; |
5225 | } | |
169ddc3e | 5226 | free_count += next - start; |
7360d173 LC |
5227 | start = next + 1; |
5228 | ||
5229 | if (fatal_signal_pending(current)) { | |
5230 | count = -ERESTARTSYS; | |
5231 | break; | |
5232 | } | |
5233 | ||
5234 | if (need_resched()) { | |
5235 | ext4_unlock_group(sb, group); | |
5236 | cond_resched(); | |
5237 | ext4_lock_group(sb, group); | |
5238 | } | |
5239 | ||
169ddc3e | 5240 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
5241 | break; |
5242 | } | |
3d56b8d2 | 5243 | |
d71c1ae2 LC |
5244 | if (!ret) { |
5245 | ret = count; | |
3d56b8d2 | 5246 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 5247 | } |
3d56b8d2 | 5248 | out: |
7360d173 | 5249 | ext4_unlock_group(sb, group); |
78944086 | 5250 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
5251 | |
5252 | ext4_debug("trimmed %d blocks in the group %d\n", | |
5253 | count, group); | |
5254 | ||
d71c1ae2 | 5255 | return ret; |
7360d173 LC |
5256 | } |
5257 | ||
5258 | /** | |
5259 | * ext4_trim_fs() -- trim ioctl handle function | |
5260 | * @sb: superblock for filesystem | |
5261 | * @range: fstrim_range structure | |
5262 | * | |
5263 | * start: First Byte to trim | |
5264 | * len: number of Bytes to trim from start | |
5265 | * minlen: minimum extent length in Bytes | |
5266 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
5267 | * start to start+len. For each such a group ext4_trim_all_free function | |
5268 | * is invoked to trim all free space. | |
5269 | */ | |
5270 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
5271 | { | |
78944086 | 5272 | struct ext4_group_info *grp; |
913eed83 | 5273 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 5274 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 5275 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
5276 | ext4_fsblk_t first_data_blk = |
5277 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 5278 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
5279 | int ret = 0; |
5280 | ||
5281 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 5282 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
5283 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
5284 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 5285 | |
5de35e8d LC |
5286 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
5287 | start >= max_blks || | |
5288 | range->len < sb->s_blocksize) | |
7360d173 | 5289 | return -EINVAL; |
913eed83 LC |
5290 | if (end >= max_blks) |
5291 | end = max_blks - 1; | |
5292 | if (end <= first_data_blk) | |
22f10457 | 5293 | goto out; |
913eed83 | 5294 | if (start < first_data_blk) |
0f0a25bf | 5295 | start = first_data_blk; |
7360d173 | 5296 | |
913eed83 | 5297 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5298 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5299 | &first_group, &first_cluster); |
913eed83 | 5300 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5301 | &last_group, &last_cluster); |
7360d173 | 5302 | |
913eed83 LC |
5303 | /* end now represents the last cluster to discard in this group */ |
5304 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5305 | |
5306 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5307 | grp = ext4_get_group_info(sb, group); |
5308 | /* We only do this if the grp has never been initialized */ | |
5309 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 5310 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
5311 | if (ret) |
5312 | break; | |
7360d173 LC |
5313 | } |
5314 | ||
0ba08517 | 5315 | /* |
913eed83 LC |
5316 | * For all the groups except the last one, last cluster will |
5317 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5318 | * change it for the last group, note that last_cluster is | |
5319 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5320 | */ |
913eed83 LC |
5321 | if (group == last_group) |
5322 | end = last_cluster; | |
7360d173 | 5323 | |
78944086 | 5324 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5325 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 5326 | end, minlen); |
7360d173 LC |
5327 | if (cnt < 0) { |
5328 | ret = cnt; | |
7360d173 LC |
5329 | break; |
5330 | } | |
21e7fd22 | 5331 | trimmed += cnt; |
7360d173 | 5332 | } |
913eed83 LC |
5333 | |
5334 | /* | |
5335 | * For every group except the first one, we are sure | |
5336 | * that the first cluster to discard will be cluster #0. | |
5337 | */ | |
7137d7a4 | 5338 | first_cluster = 0; |
7360d173 | 5339 | } |
7360d173 | 5340 | |
3d56b8d2 TM |
5341 | if (!ret) |
5342 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5343 | ||
22f10457 | 5344 | out: |
aaf7d73e | 5345 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5346 | return ret; |
5347 | } | |
0c9ec4be DW |
5348 | |
5349 | /* Iterate all the free extents in the group. */ | |
5350 | int | |
5351 | ext4_mballoc_query_range( | |
5352 | struct super_block *sb, | |
5353 | ext4_group_t group, | |
5354 | ext4_grpblk_t start, | |
5355 | ext4_grpblk_t end, | |
5356 | ext4_mballoc_query_range_fn formatter, | |
5357 | void *priv) | |
5358 | { | |
5359 | void *bitmap; | |
5360 | ext4_grpblk_t next; | |
5361 | struct ext4_buddy e4b; | |
5362 | int error; | |
5363 | ||
5364 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
5365 | if (error) | |
5366 | return error; | |
5367 | bitmap = e4b.bd_bitmap; | |
5368 | ||
5369 | ext4_lock_group(sb, group); | |
5370 | ||
5371 | start = (e4b.bd_info->bb_first_free > start) ? | |
5372 | e4b.bd_info->bb_first_free : start; | |
5373 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
5374 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
5375 | ||
5376 | while (start <= end) { | |
5377 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
5378 | if (start > end) | |
5379 | break; | |
5380 | next = mb_find_next_bit(bitmap, end + 1, start); | |
5381 | ||
5382 | ext4_unlock_group(sb, group); | |
5383 | error = formatter(sb, group, start, next - start, priv); | |
5384 | if (error) | |
5385 | goto out_unload; | |
5386 | ext4_lock_group(sb, group); | |
5387 | ||
5388 | start = next + 1; | |
5389 | } | |
5390 | ||
5391 | ext4_unlock_group(sb, group); | |
5392 | out_unload: | |
5393 | ext4_mb_unload_buddy(&e4b); | |
5394 | ||
5395 | return error; | |
5396 | } |