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 | ||
a3450215 RH |
514 | static void mb_group_bb_bitmap_alloc(struct super_block *sb, |
515 | struct ext4_group_info *grp, ext4_group_t group) | |
516 | { | |
517 | struct buffer_head *bh; | |
518 | ||
519 | grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); | |
520 | BUG_ON(grp->bb_bitmap == NULL); | |
521 | ||
522 | bh = ext4_read_block_bitmap(sb, group); | |
523 | BUG_ON(IS_ERR_OR_NULL(bh)); | |
524 | ||
525 | memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); | |
526 | put_bh(bh); | |
527 | } | |
528 | ||
529 | static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
530 | { | |
531 | kfree(grp->bb_bitmap); | |
532 | } | |
533 | ||
c9de560d AT |
534 | #else |
535 | static inline void mb_free_blocks_double(struct inode *inode, | |
536 | struct ext4_buddy *e4b, int first, int count) | |
537 | { | |
538 | return; | |
539 | } | |
540 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
541 | int first, int count) | |
542 | { | |
543 | return; | |
544 | } | |
545 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
546 | { | |
547 | return; | |
548 | } | |
a3450215 RH |
549 | |
550 | static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, | |
551 | struct ext4_group_info *grp, ext4_group_t group) | |
552 | { | |
553 | return; | |
554 | } | |
555 | ||
556 | static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
557 | { | |
558 | return; | |
559 | } | |
c9de560d AT |
560 | #endif |
561 | ||
562 | #ifdef AGGRESSIVE_CHECK | |
563 | ||
564 | #define MB_CHECK_ASSERT(assert) \ | |
565 | do { \ | |
566 | if (!(assert)) { \ | |
567 | printk(KERN_EMERG \ | |
568 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
569 | function, file, line, # assert); \ | |
570 | BUG(); \ | |
571 | } \ | |
572 | } while (0) | |
573 | ||
574 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
575 | const char *function, int line) | |
576 | { | |
577 | struct super_block *sb = e4b->bd_sb; | |
578 | int order = e4b->bd_blkbits + 1; | |
579 | int max; | |
580 | int max2; | |
581 | int i; | |
582 | int j; | |
583 | int k; | |
584 | int count; | |
585 | struct ext4_group_info *grp; | |
586 | int fragments = 0; | |
587 | int fstart; | |
588 | struct list_head *cur; | |
589 | void *buddy; | |
590 | void *buddy2; | |
591 | ||
c9de560d AT |
592 | { |
593 | static int mb_check_counter; | |
594 | if (mb_check_counter++ % 100 != 0) | |
595 | return 0; | |
596 | } | |
597 | ||
598 | while (order > 1) { | |
599 | buddy = mb_find_buddy(e4b, order, &max); | |
600 | MB_CHECK_ASSERT(buddy); | |
601 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
602 | MB_CHECK_ASSERT(buddy2); | |
603 | MB_CHECK_ASSERT(buddy != buddy2); | |
604 | MB_CHECK_ASSERT(max * 2 == max2); | |
605 | ||
606 | count = 0; | |
607 | for (i = 0; i < max; i++) { | |
608 | ||
609 | if (mb_test_bit(i, buddy)) { | |
610 | /* only single bit in buddy2 may be 1 */ | |
611 | if (!mb_test_bit(i << 1, buddy2)) { | |
612 | MB_CHECK_ASSERT( | |
613 | mb_test_bit((i<<1)+1, buddy2)); | |
614 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
615 | MB_CHECK_ASSERT( | |
616 | mb_test_bit(i << 1, buddy2)); | |
617 | } | |
618 | continue; | |
619 | } | |
620 | ||
0a10da73 | 621 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
622 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
623 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
624 | ||
625 | for (j = 0; j < (1 << order); j++) { | |
626 | k = (i * (1 << order)) + j; | |
627 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 628 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
629 | } |
630 | count++; | |
631 | } | |
632 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
633 | order--; | |
634 | } | |
635 | ||
636 | fstart = -1; | |
637 | buddy = mb_find_buddy(e4b, 0, &max); | |
638 | for (i = 0; i < max; i++) { | |
639 | if (!mb_test_bit(i, buddy)) { | |
640 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
641 | if (fstart == -1) { | |
642 | fragments++; | |
643 | fstart = i; | |
644 | } | |
645 | continue; | |
646 | } | |
647 | fstart = -1; | |
648 | /* check used bits only */ | |
649 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
650 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
651 | k = i >> j; | |
652 | MB_CHECK_ASSERT(k < max2); | |
653 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
654 | } | |
655 | } | |
656 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
657 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
658 | ||
659 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
660 | list_for_each(cur, &grp->bb_prealloc_list) { |
661 | ext4_group_t groupnr; | |
662 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
663 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
664 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 665 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 666 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
667 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
668 | } | |
669 | return 0; | |
670 | } | |
671 | #undef MB_CHECK_ASSERT | |
672 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 673 | __FILE__, __func__, __LINE__) |
c9de560d AT |
674 | #else |
675 | #define mb_check_buddy(e4b) | |
676 | #endif | |
677 | ||
7c786059 CL |
678 | /* |
679 | * Divide blocks started from @first with length @len into | |
680 | * smaller chunks with power of 2 blocks. | |
681 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
682 | * then increase bb_counters[] for corresponded chunk size. | |
683 | */ | |
c9de560d | 684 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 685 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
686 | struct ext4_group_info *grp) |
687 | { | |
688 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
689 | ext4_grpblk_t min; |
690 | ext4_grpblk_t max; | |
691 | ext4_grpblk_t chunk; | |
69e43e8c | 692 | unsigned int border; |
c9de560d | 693 | |
7137d7a4 | 694 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
695 | |
696 | border = 2 << sb->s_blocksize_bits; | |
697 | ||
698 | while (len > 0) { | |
699 | /* find how many blocks can be covered since this position */ | |
700 | max = ffs(first | border) - 1; | |
701 | ||
702 | /* find how many blocks of power 2 we need to mark */ | |
703 | min = fls(len) - 1; | |
704 | ||
705 | if (max < min) | |
706 | min = max; | |
707 | chunk = 1 << min; | |
708 | ||
709 | /* mark multiblock chunks only */ | |
710 | grp->bb_counters[min]++; | |
711 | if (min > 0) | |
712 | mb_clear_bit(first >> min, | |
713 | buddy + sbi->s_mb_offsets[min]); | |
714 | ||
715 | len -= chunk; | |
716 | first += chunk; | |
717 | } | |
718 | } | |
719 | ||
8a57d9d6 CW |
720 | /* |
721 | * Cache the order of the largest free extent we have available in this block | |
722 | * group. | |
723 | */ | |
724 | static void | |
725 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
726 | { | |
727 | int i; | |
728 | int bits; | |
729 | ||
730 | grp->bb_largest_free_order = -1; /* uninit */ | |
731 | ||
732 | bits = sb->s_blocksize_bits + 1; | |
733 | for (i = bits; i >= 0; i--) { | |
734 | if (grp->bb_counters[i] > 0) { | |
735 | grp->bb_largest_free_order = i; | |
736 | break; | |
737 | } | |
738 | } | |
739 | } | |
740 | ||
089ceecc ES |
741 | static noinline_for_stack |
742 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
743 | void *buddy, void *bitmap, ext4_group_t group) |
744 | { | |
745 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 746 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 747 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
748 | ext4_grpblk_t i = 0; |
749 | ext4_grpblk_t first; | |
750 | ext4_grpblk_t len; | |
c9de560d AT |
751 | unsigned free = 0; |
752 | unsigned fragments = 0; | |
753 | unsigned long long period = get_cycles(); | |
754 | ||
755 | /* initialize buddy from bitmap which is aggregation | |
756 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 757 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
758 | grp->bb_first_free = i; |
759 | while (i < max) { | |
760 | fragments++; | |
761 | first = i; | |
ffad0a44 | 762 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
763 | len = i - first; |
764 | free += len; | |
765 | if (len > 1) | |
766 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
767 | else | |
768 | grp->bb_counters[0]++; | |
769 | if (i < max) | |
ffad0a44 | 770 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
771 | } |
772 | grp->bb_fragments = fragments; | |
773 | ||
774 | if (free != grp->bb_free) { | |
e29136f8 | 775 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
776 | "block bitmap and bg descriptor " |
777 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 778 | free, grp->bb_free); |
e56eb659 | 779 | /* |
163a203d | 780 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
781 | * corrupt and update bb_free using bitmap value |
782 | */ | |
c9de560d | 783 | grp->bb_free = free; |
db79e6d1 WS |
784 | ext4_mark_group_bitmap_corrupted(sb, group, |
785 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 786 | } |
8a57d9d6 | 787 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
788 | |
789 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
790 | ||
791 | period = get_cycles() - period; | |
49598e04 JP |
792 | spin_lock(&sbi->s_bal_lock); |
793 | sbi->s_mb_buddies_generated++; | |
794 | sbi->s_mb_generation_time += period; | |
795 | spin_unlock(&sbi->s_bal_lock); | |
c9de560d AT |
796 | } |
797 | ||
eabe0444 AS |
798 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
799 | { | |
800 | int count; | |
801 | int order = 1; | |
802 | void *buddy; | |
803 | ||
804 | while ((buddy = mb_find_buddy(e4b, order++, &count))) { | |
805 | ext4_set_bits(buddy, 0, count); | |
806 | } | |
807 | e4b->bd_info->bb_fragments = 0; | |
808 | memset(e4b->bd_info->bb_counters, 0, | |
809 | sizeof(*e4b->bd_info->bb_counters) * | |
810 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
811 | ||
812 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
813 | e4b->bd_bitmap, e4b->bd_group); | |
814 | } | |
815 | ||
c9de560d AT |
816 | /* The buddy information is attached the buddy cache inode |
817 | * for convenience. The information regarding each group | |
818 | * is loaded via ext4_mb_load_buddy. The information involve | |
819 | * block bitmap and buddy information. The information are | |
820 | * stored in the inode as | |
821 | * | |
822 | * { page } | |
c3a326a6 | 823 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
824 | * |
825 | * | |
826 | * one block each for bitmap and buddy information. | |
827 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 828 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
829 | * So it can have information regarding groups_per_page which |
830 | * is blocks_per_page/2 | |
8a57d9d6 CW |
831 | * |
832 | * Locking note: This routine takes the block group lock of all groups | |
833 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
834 | */ |
835 | ||
adb7ef60 | 836 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 837 | { |
8df9675f | 838 | ext4_group_t ngroups; |
c9de560d AT |
839 | int blocksize; |
840 | int blocks_per_page; | |
841 | int groups_per_page; | |
842 | int err = 0; | |
843 | int i; | |
813e5727 | 844 | ext4_group_t first_group, group; |
c9de560d AT |
845 | int first_block; |
846 | struct super_block *sb; | |
847 | struct buffer_head *bhs; | |
fa77dcfa | 848 | struct buffer_head **bh = NULL; |
c9de560d AT |
849 | struct inode *inode; |
850 | char *data; | |
851 | char *bitmap; | |
9b8b7d35 | 852 | struct ext4_group_info *grinfo; |
c9de560d | 853 | |
6ba495e9 | 854 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
855 | |
856 | inode = page->mapping->host; | |
857 | sb = inode->i_sb; | |
8df9675f | 858 | ngroups = ext4_get_groups_count(sb); |
93407472 | 859 | blocksize = i_blocksize(inode); |
09cbfeaf | 860 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d AT |
861 | |
862 | groups_per_page = blocks_per_page >> 1; | |
863 | if (groups_per_page == 0) | |
864 | groups_per_page = 1; | |
865 | ||
866 | /* allocate buffer_heads to read bitmaps */ | |
867 | if (groups_per_page > 1) { | |
c9de560d | 868 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 869 | bh = kzalloc(i, gfp); |
813e5727 TT |
870 | if (bh == NULL) { |
871 | err = -ENOMEM; | |
c9de560d | 872 | goto out; |
813e5727 | 873 | } |
c9de560d AT |
874 | } else |
875 | bh = &bhs; | |
876 | ||
877 | first_group = page->index * blocks_per_page / 2; | |
878 | ||
879 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
880 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
881 | if (group >= ngroups) | |
c9de560d AT |
882 | break; |
883 | ||
813e5727 | 884 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
885 | /* |
886 | * If page is uptodate then we came here after online resize | |
887 | * which added some new uninitialized group info structs, so | |
888 | * we must skip all initialized uptodate buddies on the page, | |
889 | * which may be currently in use by an allocating task. | |
890 | */ | |
891 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
892 | bh[i] = NULL; | |
893 | continue; | |
894 | } | |
9008a58e DW |
895 | bh[i] = ext4_read_block_bitmap_nowait(sb, group); |
896 | if (IS_ERR(bh[i])) { | |
897 | err = PTR_ERR(bh[i]); | |
898 | bh[i] = NULL; | |
c9de560d | 899 | goto out; |
2ccb5fb9 | 900 | } |
813e5727 | 901 | mb_debug(1, "read bitmap for group %u\n", group); |
c9de560d AT |
902 | } |
903 | ||
904 | /* wait for I/O completion */ | |
813e5727 | 905 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
906 | int err2; |
907 | ||
908 | if (!bh[i]) | |
909 | continue; | |
910 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
911 | if (!err) | |
912 | err = err2; | |
813e5727 | 913 | } |
c9de560d AT |
914 | |
915 | first_block = page->index * blocks_per_page; | |
916 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 917 | group = (first_block + i) >> 1; |
8df9675f | 918 | if (group >= ngroups) |
c9de560d AT |
919 | break; |
920 | ||
9b8b7d35 AG |
921 | if (!bh[group - first_group]) |
922 | /* skip initialized uptodate buddy */ | |
923 | continue; | |
924 | ||
bbdc322f LC |
925 | if (!buffer_verified(bh[group - first_group])) |
926 | /* Skip faulty bitmaps */ | |
927 | continue; | |
928 | err = 0; | |
929 | ||
c9de560d AT |
930 | /* |
931 | * data carry information regarding this | |
932 | * particular group in the format specified | |
933 | * above | |
934 | * | |
935 | */ | |
936 | data = page_address(page) + (i * blocksize); | |
937 | bitmap = bh[group - first_group]->b_data; | |
938 | ||
939 | /* | |
940 | * We place the buddy block and bitmap block | |
941 | * close together | |
942 | */ | |
943 | if ((first_block + i) & 1) { | |
944 | /* this is block of buddy */ | |
945 | BUG_ON(incore == NULL); | |
6ba495e9 | 946 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 947 | group, page->index, i * blocksize); |
f307333e | 948 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
949 | grinfo = ext4_get_group_info(sb, group); |
950 | grinfo->bb_fragments = 0; | |
951 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
952 | sizeof(*grinfo->bb_counters) * |
953 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
954 | /* |
955 | * incore got set to the group block bitmap below | |
956 | */ | |
7a2fcbf7 | 957 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
958 | /* init the buddy */ |
959 | memset(data, 0xff, blocksize); | |
c9de560d | 960 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 961 | ext4_unlock_group(sb, group); |
c9de560d AT |
962 | incore = NULL; |
963 | } else { | |
964 | /* this is block of bitmap */ | |
965 | BUG_ON(incore != NULL); | |
6ba495e9 | 966 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 967 | group, page->index, i * blocksize); |
f307333e | 968 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
969 | |
970 | /* see comments in ext4_mb_put_pa() */ | |
971 | ext4_lock_group(sb, group); | |
972 | memcpy(data, bitmap, blocksize); | |
973 | ||
974 | /* mark all preallocated blks used in in-core bitmap */ | |
975 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 976 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
977 | ext4_unlock_group(sb, group); |
978 | ||
979 | /* set incore so that the buddy information can be | |
980 | * generated using this | |
981 | */ | |
982 | incore = data; | |
983 | } | |
984 | } | |
985 | SetPageUptodate(page); | |
986 | ||
987 | out: | |
988 | if (bh) { | |
9b8b7d35 | 989 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
990 | brelse(bh[i]); |
991 | if (bh != &bhs) | |
992 | kfree(bh); | |
993 | } | |
994 | return err; | |
995 | } | |
996 | ||
eee4adc7 | 997 | /* |
2de8807b AG |
998 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
999 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1000 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
1001 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 1002 | */ |
2de8807b | 1003 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1004 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1005 | { |
2de8807b AG |
1006 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1007 | int block, pnum, poff; | |
eee4adc7 | 1008 | int blocks_per_page; |
2de8807b AG |
1009 | struct page *page; |
1010 | ||
1011 | e4b->bd_buddy_page = NULL; | |
1012 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 1013 | |
09cbfeaf | 1014 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1015 | /* |
1016 | * the buddy cache inode stores the block bitmap | |
1017 | * and buddy information in consecutive blocks. | |
1018 | * So for each group we need two blocks. | |
1019 | */ | |
1020 | block = group * 2; | |
1021 | pnum = block / blocks_per_page; | |
2de8807b | 1022 | poff = block % blocks_per_page; |
adb7ef60 | 1023 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1024 | if (!page) |
c57ab39b | 1025 | return -ENOMEM; |
2de8807b AG |
1026 | BUG_ON(page->mapping != inode->i_mapping); |
1027 | e4b->bd_bitmap_page = page; | |
1028 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1029 | ||
1030 | if (blocks_per_page >= 2) { | |
1031 | /* buddy and bitmap are on the same page */ | |
1032 | return 0; | |
eee4adc7 | 1033 | } |
2de8807b AG |
1034 | |
1035 | block++; | |
1036 | pnum = block / blocks_per_page; | |
adb7ef60 | 1037 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1038 | if (!page) |
c57ab39b | 1039 | return -ENOMEM; |
2de8807b AG |
1040 | BUG_ON(page->mapping != inode->i_mapping); |
1041 | e4b->bd_buddy_page = page; | |
1042 | return 0; | |
eee4adc7 ES |
1043 | } |
1044 | ||
2de8807b | 1045 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1046 | { |
2de8807b AG |
1047 | if (e4b->bd_bitmap_page) { |
1048 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1049 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1050 | } |
1051 | if (e4b->bd_buddy_page) { | |
1052 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1053 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1054 | } |
eee4adc7 ES |
1055 | } |
1056 | ||
8a57d9d6 CW |
1057 | /* |
1058 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1059 | * block group lock of all groups for this page; do not hold the BG lock when | |
1060 | * calling this routine! | |
1061 | */ | |
b6a758ec | 1062 | static noinline_for_stack |
adb7ef60 | 1063 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1064 | { |
1065 | ||
b6a758ec | 1066 | struct ext4_group_info *this_grp; |
2de8807b AG |
1067 | struct ext4_buddy e4b; |
1068 | struct page *page; | |
1069 | int ret = 0; | |
b6a758ec | 1070 | |
b10a44c3 | 1071 | might_sleep(); |
b6a758ec | 1072 | mb_debug(1, "init group %u\n", group); |
b6a758ec AK |
1073 | this_grp = ext4_get_group_info(sb, group); |
1074 | /* | |
08c3a813 AK |
1075 | * This ensures that we don't reinit the buddy cache |
1076 | * page which map to the group from which we are already | |
1077 | * allocating. If we are looking at the buddy cache we would | |
1078 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1079 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1080 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1081 | * page accessed. | |
b6a758ec | 1082 | */ |
adb7ef60 | 1083 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1084 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1085 | /* |
1086 | * somebody initialized the group | |
1087 | * return without doing anything | |
1088 | */ | |
b6a758ec AK |
1089 | goto err; |
1090 | } | |
2de8807b AG |
1091 | |
1092 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1093 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1094 | if (ret) |
1095 | goto err; | |
1096 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1097 | ret = -EIO; |
1098 | goto err; | |
1099 | } | |
b6a758ec | 1100 | |
2de8807b | 1101 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1102 | /* |
1103 | * If both the bitmap and buddy are in | |
1104 | * the same page we don't need to force | |
1105 | * init the buddy | |
1106 | */ | |
2de8807b AG |
1107 | ret = 0; |
1108 | goto err; | |
b6a758ec | 1109 | } |
2de8807b AG |
1110 | /* init buddy cache */ |
1111 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1112 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1113 | if (ret) |
1114 | goto err; | |
1115 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1116 | ret = -EIO; |
1117 | goto err; | |
1118 | } | |
b6a758ec | 1119 | err: |
2de8807b | 1120 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1121 | return ret; |
1122 | } | |
1123 | ||
8a57d9d6 CW |
1124 | /* |
1125 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1126 | * block group lock of all groups for this page; do not hold the BG lock when | |
1127 | * calling this routine! | |
1128 | */ | |
4ddfef7b | 1129 | static noinline_for_stack int |
adb7ef60 KK |
1130 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1131 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1132 | { |
c9de560d AT |
1133 | int blocks_per_page; |
1134 | int block; | |
1135 | int pnum; | |
1136 | int poff; | |
1137 | struct page *page; | |
fdf6c7a7 | 1138 | int ret; |
920313a7 AK |
1139 | struct ext4_group_info *grp; |
1140 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1141 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1142 | |
b10a44c3 | 1143 | might_sleep(); |
6ba495e9 | 1144 | mb_debug(1, "load group %u\n", group); |
c9de560d | 1145 | |
09cbfeaf | 1146 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1147 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1148 | |
1149 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1150 | e4b->bd_info = grp; |
c9de560d AT |
1151 | e4b->bd_sb = sb; |
1152 | e4b->bd_group = group; | |
1153 | e4b->bd_buddy_page = NULL; | |
1154 | e4b->bd_bitmap_page = NULL; | |
1155 | ||
f41c0750 | 1156 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1157 | /* |
1158 | * we need full data about the group | |
1159 | * to make a good selection | |
1160 | */ | |
adb7ef60 | 1161 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1162 | if (ret) |
1163 | return ret; | |
f41c0750 AK |
1164 | } |
1165 | ||
c9de560d AT |
1166 | /* |
1167 | * the buddy cache inode stores the block bitmap | |
1168 | * and buddy information in consecutive blocks. | |
1169 | * So for each group we need two blocks. | |
1170 | */ | |
1171 | block = group * 2; | |
1172 | pnum = block / blocks_per_page; | |
1173 | poff = block % blocks_per_page; | |
1174 | ||
1175 | /* we could use find_or_create_page(), but it locks page | |
1176 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1177 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1178 | if (page == NULL || !PageUptodate(page)) { |
1179 | if (page) | |
920313a7 AK |
1180 | /* |
1181 | * drop the page reference and try | |
1182 | * to get the page with lock. If we | |
1183 | * are not uptodate that implies | |
1184 | * somebody just created the page but | |
1185 | * is yet to initialize the same. So | |
1186 | * wait for it to initialize. | |
1187 | */ | |
09cbfeaf | 1188 | put_page(page); |
adb7ef60 | 1189 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1190 | if (page) { |
1191 | BUG_ON(page->mapping != inode->i_mapping); | |
1192 | if (!PageUptodate(page)) { | |
adb7ef60 | 1193 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1194 | if (ret) { |
1195 | unlock_page(page); | |
1196 | goto err; | |
1197 | } | |
c9de560d AT |
1198 | mb_cmp_bitmaps(e4b, page_address(page) + |
1199 | (poff * sb->s_blocksize)); | |
1200 | } | |
1201 | unlock_page(page); | |
1202 | } | |
1203 | } | |
c57ab39b YL |
1204 | if (page == NULL) { |
1205 | ret = -ENOMEM; | |
1206 | goto err; | |
1207 | } | |
1208 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1209 | ret = -EIO; |
c9de560d | 1210 | goto err; |
fdf6c7a7 | 1211 | } |
2457aec6 MG |
1212 | |
1213 | /* Pages marked accessed already */ | |
c9de560d AT |
1214 | e4b->bd_bitmap_page = page; |
1215 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1216 | |
1217 | block++; | |
1218 | pnum = block / blocks_per_page; | |
1219 | poff = block % blocks_per_page; | |
1220 | ||
2457aec6 | 1221 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1222 | if (page == NULL || !PageUptodate(page)) { |
1223 | if (page) | |
09cbfeaf | 1224 | put_page(page); |
adb7ef60 | 1225 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1226 | if (page) { |
1227 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1228 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1229 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1230 | gfp); | |
fdf6c7a7 SF |
1231 | if (ret) { |
1232 | unlock_page(page); | |
1233 | goto err; | |
1234 | } | |
1235 | } | |
c9de560d AT |
1236 | unlock_page(page); |
1237 | } | |
1238 | } | |
c57ab39b YL |
1239 | if (page == NULL) { |
1240 | ret = -ENOMEM; | |
1241 | goto err; | |
1242 | } | |
1243 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1244 | ret = -EIO; |
c9de560d | 1245 | goto err; |
fdf6c7a7 | 1246 | } |
2457aec6 MG |
1247 | |
1248 | /* Pages marked accessed already */ | |
c9de560d AT |
1249 | e4b->bd_buddy_page = page; |
1250 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1251 | |
1252 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1253 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1254 | ||
1255 | return 0; | |
1256 | ||
1257 | err: | |
26626f11 | 1258 | if (page) |
09cbfeaf | 1259 | put_page(page); |
c9de560d | 1260 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1261 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1262 | if (e4b->bd_buddy_page) |
09cbfeaf | 1263 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1264 | e4b->bd_buddy = NULL; |
1265 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1266 | return ret; |
c9de560d AT |
1267 | } |
1268 | ||
adb7ef60 KK |
1269 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1270 | struct ext4_buddy *e4b) | |
1271 | { | |
1272 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1273 | } | |
1274 | ||
e39e07fd | 1275 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1276 | { |
1277 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1278 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1279 | if (e4b->bd_buddy_page) |
09cbfeaf | 1280 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1281 | } |
1282 | ||
1283 | ||
1284 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1285 | { | |
1286 | int order = 1; | |
b5cb316c | 1287 | int bb_incr = 1 << (e4b->bd_blkbits - 1); |
c9de560d AT |
1288 | void *bb; |
1289 | ||
c5e8f3f3 | 1290 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1291 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1292 | ||
c5e8f3f3 | 1293 | bb = e4b->bd_buddy; |
c9de560d AT |
1294 | while (order <= e4b->bd_blkbits + 1) { |
1295 | block = block >> 1; | |
1296 | if (!mb_test_bit(block, bb)) { | |
1297 | /* this block is part of buddy of order 'order' */ | |
1298 | return order; | |
1299 | } | |
b5cb316c NS |
1300 | bb += bb_incr; |
1301 | bb_incr >>= 1; | |
c9de560d AT |
1302 | order++; |
1303 | } | |
1304 | return 0; | |
1305 | } | |
1306 | ||
955ce5f5 | 1307 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1308 | { |
1309 | __u32 *addr; | |
1310 | ||
1311 | len = cur + len; | |
1312 | while (cur < len) { | |
1313 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1314 | /* fast path: clear whole word at once */ | |
1315 | addr = bm + (cur >> 3); | |
1316 | *addr = 0; | |
1317 | cur += 32; | |
1318 | continue; | |
1319 | } | |
955ce5f5 | 1320 | mb_clear_bit(cur, bm); |
c9de560d AT |
1321 | cur++; |
1322 | } | |
1323 | } | |
1324 | ||
eabe0444 AS |
1325 | /* clear bits in given range |
1326 | * will return first found zero bit if any, -1 otherwise | |
1327 | */ | |
1328 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1329 | { | |
1330 | __u32 *addr; | |
1331 | int zero_bit = -1; | |
1332 | ||
1333 | len = cur + len; | |
1334 | while (cur < len) { | |
1335 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1336 | /* fast path: clear whole word at once */ | |
1337 | addr = bm + (cur >> 3); | |
1338 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1339 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1340 | *addr = 0; | |
1341 | cur += 32; | |
1342 | continue; | |
1343 | } | |
1344 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1345 | zero_bit = cur; | |
1346 | cur++; | |
1347 | } | |
1348 | ||
1349 | return zero_bit; | |
1350 | } | |
1351 | ||
c3e94d1d | 1352 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1353 | { |
1354 | __u32 *addr; | |
1355 | ||
1356 | len = cur + len; | |
1357 | while (cur < len) { | |
1358 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1359 | /* fast path: set whole word at once */ | |
1360 | addr = bm + (cur >> 3); | |
1361 | *addr = 0xffffffff; | |
1362 | cur += 32; | |
1363 | continue; | |
1364 | } | |
955ce5f5 | 1365 | mb_set_bit(cur, bm); |
c9de560d AT |
1366 | cur++; |
1367 | } | |
1368 | } | |
1369 | ||
eabe0444 AS |
1370 | /* |
1371 | * _________________________________________________________________ */ | |
1372 | ||
1373 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) | |
1374 | { | |
1375 | if (mb_test_bit(*bit + side, bitmap)) { | |
1376 | mb_clear_bit(*bit, bitmap); | |
1377 | (*bit) -= side; | |
1378 | return 1; | |
1379 | } | |
1380 | else { | |
1381 | (*bit) += side; | |
1382 | mb_set_bit(*bit, bitmap); | |
1383 | return -1; | |
1384 | } | |
1385 | } | |
1386 | ||
1387 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1388 | { | |
1389 | int max; | |
1390 | int order = 1; | |
1391 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1392 | ||
1393 | while (buddy) { | |
1394 | void *buddy2; | |
1395 | ||
1396 | /* Bits in range [first; last] are known to be set since | |
1397 | * corresponding blocks were allocated. Bits in range | |
1398 | * (first; last) will stay set because they form buddies on | |
1399 | * upper layer. We just deal with borders if they don't | |
1400 | * align with upper layer and then go up. | |
1401 | * Releasing entire group is all about clearing | |
1402 | * single bit of highest order buddy. | |
1403 | */ | |
1404 | ||
1405 | /* Example: | |
1406 | * --------------------------------- | |
1407 | * | 1 | 1 | 1 | 1 | | |
1408 | * --------------------------------- | |
1409 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1410 | * --------------------------------- | |
1411 | * 0 1 2 3 4 5 6 7 | |
1412 | * \_____________________/ | |
1413 | * | |
1414 | * Neither [1] nor [6] is aligned to above layer. | |
1415 | * Left neighbour [0] is free, so mark it busy, | |
1416 | * decrease bb_counters and extend range to | |
1417 | * [0; 6] | |
1418 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1419 | * mark [6] free, increase bb_counters and shrink range to | |
1420 | * [0; 5]. | |
1421 | * Then shift range to [0; 2], go up and do the same. | |
1422 | */ | |
1423 | ||
1424 | ||
1425 | if (first & 1) | |
1426 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1427 | if (!(last & 1)) | |
1428 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1429 | if (first > last) | |
1430 | break; | |
1431 | order++; | |
1432 | ||
1433 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1434 | mb_clear_bits(buddy, first, last - first + 1); | |
1435 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1436 | break; | |
1437 | } | |
1438 | first >>= 1; | |
1439 | last >>= 1; | |
1440 | buddy = buddy2; | |
1441 | } | |
1442 | } | |
1443 | ||
7e5a8cdd | 1444 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1445 | int first, int count) |
c9de560d | 1446 | { |
eabe0444 AS |
1447 | int left_is_free = 0; |
1448 | int right_is_free = 0; | |
1449 | int block; | |
1450 | int last = first + count - 1; | |
c9de560d AT |
1451 | struct super_block *sb = e4b->bd_sb; |
1452 | ||
c99d1e6e TT |
1453 | if (WARN_ON(count == 0)) |
1454 | return; | |
eabe0444 | 1455 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1456 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1457 | /* Don't bother if the block group is corrupt. */ |
1458 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1459 | return; | |
1460 | ||
c9de560d AT |
1461 | mb_check_buddy(e4b); |
1462 | mb_free_blocks_double(inode, e4b, first, count); | |
1463 | ||
1464 | e4b->bd_info->bb_free += count; | |
1465 | if (first < e4b->bd_info->bb_first_free) | |
1466 | e4b->bd_info->bb_first_free = first; | |
1467 | ||
eabe0444 AS |
1468 | /* access memory sequentially: check left neighbour, |
1469 | * clear range and then check right neighbour | |
1470 | */ | |
c9de560d | 1471 | if (first != 0) |
eabe0444 AS |
1472 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1473 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1474 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1475 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1476 | ||
1477 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1478 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1479 | ext4_fsblk_t blocknr; |
1480 | ||
1481 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1482 | blocknr += EXT4_C2B(sbi, block); |
eabe0444 AS |
1483 | ext4_grp_locked_error(sb, e4b->bd_group, |
1484 | inode ? inode->i_ino : 0, | |
1485 | blocknr, | |
1486 | "freeing already freed block " | |
163a203d DW |
1487 | "(bit %u); block bitmap corrupt.", |
1488 | block); | |
db79e6d1 WS |
1489 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1490 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
eabe0444 AS |
1491 | mb_regenerate_buddy(e4b); |
1492 | goto done; | |
1493 | } | |
1494 | ||
1495 | /* let's maintain fragments counter */ | |
1496 | if (left_is_free && right_is_free) | |
c9de560d | 1497 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1498 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1499 | e4b->bd_info->bb_fragments++; |
1500 | ||
eabe0444 AS |
1501 | /* buddy[0] == bd_bitmap is a special case, so handle |
1502 | * it right away and let mb_buddy_mark_free stay free of | |
1503 | * zero order checks. | |
1504 | * Check if neighbours are to be coaleasced, | |
1505 | * adjust bitmap bb_counters and borders appropriately. | |
1506 | */ | |
1507 | if (first & 1) { | |
1508 | first += !left_is_free; | |
1509 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1510 | } | |
1511 | if (!(last & 1)) { | |
1512 | last -= !right_is_free; | |
1513 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1514 | } | |
c9de560d | 1515 | |
eabe0444 AS |
1516 | if (first <= last) |
1517 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1518 | |
eabe0444 | 1519 | done: |
8a57d9d6 | 1520 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1521 | mb_check_buddy(e4b); |
c9de560d AT |
1522 | } |
1523 | ||
15c006a2 | 1524 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1525 | int needed, struct ext4_free_extent *ex) |
1526 | { | |
1527 | int next = block; | |
15c006a2 | 1528 | int max, order; |
c9de560d AT |
1529 | void *buddy; |
1530 | ||
bc8e6740 | 1531 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1532 | BUG_ON(ex == NULL); |
1533 | ||
15c006a2 | 1534 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1535 | BUG_ON(buddy == NULL); |
1536 | BUG_ON(block >= max); | |
1537 | if (mb_test_bit(block, buddy)) { | |
1538 | ex->fe_len = 0; | |
1539 | ex->fe_start = 0; | |
1540 | ex->fe_group = 0; | |
1541 | return 0; | |
1542 | } | |
1543 | ||
15c006a2 RD |
1544 | /* find actual order */ |
1545 | order = mb_find_order_for_block(e4b, block); | |
1546 | block = block >> order; | |
c9de560d AT |
1547 | |
1548 | ex->fe_len = 1 << order; | |
1549 | ex->fe_start = block << order; | |
1550 | ex->fe_group = e4b->bd_group; | |
1551 | ||
1552 | /* calc difference from given start */ | |
1553 | next = next - ex->fe_start; | |
1554 | ex->fe_len -= next; | |
1555 | ex->fe_start += next; | |
1556 | ||
1557 | while (needed > ex->fe_len && | |
d8ec0c39 | 1558 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1559 | |
1560 | if (block + 1 >= max) | |
1561 | break; | |
1562 | ||
1563 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1564 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1565 | break; |
1566 | ||
b051d8dc | 1567 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1568 | |
c9de560d AT |
1569 | block = next >> order; |
1570 | ex->fe_len += 1 << order; | |
1571 | } | |
1572 | ||
31562b95 | 1573 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1574 | /* Should never happen! (but apparently sometimes does?!?) */ |
1575 | WARN_ON(1); | |
1576 | ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent " | |
1577 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1578 | block, order, needed, ex->fe_group, ex->fe_start, | |
1579 | ex->fe_len, ex->fe_logical); | |
1580 | ex->fe_len = 0; | |
1581 | ex->fe_start = 0; | |
1582 | ex->fe_group = 0; | |
1583 | } | |
c9de560d AT |
1584 | return ex->fe_len; |
1585 | } | |
1586 | ||
1587 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1588 | { | |
1589 | int ord; | |
1590 | int mlen = 0; | |
1591 | int max = 0; | |
1592 | int cur; | |
1593 | int start = ex->fe_start; | |
1594 | int len = ex->fe_len; | |
1595 | unsigned ret = 0; | |
1596 | int len0 = len; | |
1597 | void *buddy; | |
1598 | ||
1599 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1600 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1601 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1602 | mb_check_buddy(e4b); |
1603 | mb_mark_used_double(e4b, start, len); | |
1604 | ||
1605 | e4b->bd_info->bb_free -= len; | |
1606 | if (e4b->bd_info->bb_first_free == start) | |
1607 | e4b->bd_info->bb_first_free += len; | |
1608 | ||
1609 | /* let's maintain fragments counter */ | |
1610 | if (start != 0) | |
c5e8f3f3 | 1611 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1612 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1613 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1614 | if (mlen && max) |
1615 | e4b->bd_info->bb_fragments++; | |
1616 | else if (!mlen && !max) | |
1617 | e4b->bd_info->bb_fragments--; | |
1618 | ||
1619 | /* let's maintain buddy itself */ | |
1620 | while (len) { | |
1621 | ord = mb_find_order_for_block(e4b, start); | |
1622 | ||
1623 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1624 | /* the whole chunk may be allocated at once! */ | |
1625 | mlen = 1 << ord; | |
1626 | buddy = mb_find_buddy(e4b, ord, &max); | |
1627 | BUG_ON((start >> ord) >= max); | |
1628 | mb_set_bit(start >> ord, buddy); | |
1629 | e4b->bd_info->bb_counters[ord]--; | |
1630 | start += mlen; | |
1631 | len -= mlen; | |
1632 | BUG_ON(len < 0); | |
1633 | continue; | |
1634 | } | |
1635 | ||
1636 | /* store for history */ | |
1637 | if (ret == 0) | |
1638 | ret = len | (ord << 16); | |
1639 | ||
1640 | /* we have to split large buddy */ | |
1641 | BUG_ON(ord <= 0); | |
1642 | buddy = mb_find_buddy(e4b, ord, &max); | |
1643 | mb_set_bit(start >> ord, buddy); | |
1644 | e4b->bd_info->bb_counters[ord]--; | |
1645 | ||
1646 | ord--; | |
1647 | cur = (start >> ord) & ~1U; | |
1648 | buddy = mb_find_buddy(e4b, ord, &max); | |
1649 | mb_clear_bit(cur, buddy); | |
1650 | mb_clear_bit(cur + 1, buddy); | |
1651 | e4b->bd_info->bb_counters[ord]++; | |
1652 | e4b->bd_info->bb_counters[ord]++; | |
1653 | } | |
8a57d9d6 | 1654 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1655 | |
c5e8f3f3 | 1656 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1657 | mb_check_buddy(e4b); |
1658 | ||
1659 | return ret; | |
1660 | } | |
1661 | ||
1662 | /* | |
1663 | * Must be called under group lock! | |
1664 | */ | |
1665 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1666 | struct ext4_buddy *e4b) | |
1667 | { | |
1668 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1669 | int ret; | |
1670 | ||
1671 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1672 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1673 | ||
1674 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1675 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1676 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1677 | ||
1678 | /* preallocation can change ac_b_ex, thus we store actually | |
1679 | * allocated blocks for history */ | |
1680 | ac->ac_f_ex = ac->ac_b_ex; | |
1681 | ||
1682 | ac->ac_status = AC_STATUS_FOUND; | |
1683 | ac->ac_tail = ret & 0xffff; | |
1684 | ac->ac_buddy = ret >> 16; | |
1685 | ||
c3a326a6 AK |
1686 | /* |
1687 | * take the page reference. We want the page to be pinned | |
1688 | * so that we don't get a ext4_mb_init_cache_call for this | |
1689 | * group until we update the bitmap. That would mean we | |
1690 | * double allocate blocks. The reference is dropped | |
1691 | * in ext4_mb_release_context | |
1692 | */ | |
c9de560d AT |
1693 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1694 | get_page(ac->ac_bitmap_page); | |
1695 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1696 | get_page(ac->ac_buddy_page); | |
c9de560d | 1697 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1698 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1699 | spin_lock(&sbi->s_md_lock); |
1700 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1701 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1702 | spin_unlock(&sbi->s_md_lock); | |
1703 | } | |
1704 | } | |
1705 | ||
1706 | /* | |
1707 | * regular allocator, for general purposes allocation | |
1708 | */ | |
1709 | ||
1710 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1711 | struct ext4_buddy *e4b, | |
1712 | int finish_group) | |
1713 | { | |
1714 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1715 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1716 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1717 | struct ext4_free_extent ex; | |
1718 | int max; | |
1719 | ||
032115fc AK |
1720 | if (ac->ac_status == AC_STATUS_FOUND) |
1721 | return; | |
c9de560d AT |
1722 | /* |
1723 | * We don't want to scan for a whole year | |
1724 | */ | |
1725 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1726 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1727 | ac->ac_status = AC_STATUS_BREAK; | |
1728 | return; | |
1729 | } | |
1730 | ||
1731 | /* | |
1732 | * Haven't found good chunk so far, let's continue | |
1733 | */ | |
1734 | if (bex->fe_len < gex->fe_len) | |
1735 | return; | |
1736 | ||
1737 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1738 | && bex->fe_group == e4b->bd_group) { | |
1739 | /* recheck chunk's availability - we don't know | |
1740 | * when it was found (within this lock-unlock | |
1741 | * period or not) */ | |
15c006a2 | 1742 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1743 | if (max >= gex->fe_len) { |
1744 | ext4_mb_use_best_found(ac, e4b); | |
1745 | return; | |
1746 | } | |
1747 | } | |
1748 | } | |
1749 | ||
1750 | /* | |
1751 | * The routine checks whether found extent is good enough. If it is, | |
1752 | * then the extent gets marked used and flag is set to the context | |
1753 | * to stop scanning. Otherwise, the extent is compared with the | |
1754 | * previous found extent and if new one is better, then it's stored | |
1755 | * in the context. Later, the best found extent will be used, if | |
1756 | * mballoc can't find good enough extent. | |
1757 | * | |
1758 | * FIXME: real allocation policy is to be designed yet! | |
1759 | */ | |
1760 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1761 | struct ext4_free_extent *ex, | |
1762 | struct ext4_buddy *e4b) | |
1763 | { | |
1764 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1765 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1766 | ||
1767 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1768 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1769 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1770 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1771 | ||
1772 | ac->ac_found++; | |
1773 | ||
1774 | /* | |
1775 | * The special case - take what you catch first | |
1776 | */ | |
1777 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1778 | *bex = *ex; | |
1779 | ext4_mb_use_best_found(ac, e4b); | |
1780 | return; | |
1781 | } | |
1782 | ||
1783 | /* | |
1784 | * Let's check whether the chuck is good enough | |
1785 | */ | |
1786 | if (ex->fe_len == gex->fe_len) { | |
1787 | *bex = *ex; | |
1788 | ext4_mb_use_best_found(ac, e4b); | |
1789 | return; | |
1790 | } | |
1791 | ||
1792 | /* | |
1793 | * If this is first found extent, just store it in the context | |
1794 | */ | |
1795 | if (bex->fe_len == 0) { | |
1796 | *bex = *ex; | |
1797 | return; | |
1798 | } | |
1799 | ||
1800 | /* | |
1801 | * If new found extent is better, store it in the context | |
1802 | */ | |
1803 | if (bex->fe_len < gex->fe_len) { | |
1804 | /* if the request isn't satisfied, any found extent | |
1805 | * larger than previous best one is better */ | |
1806 | if (ex->fe_len > bex->fe_len) | |
1807 | *bex = *ex; | |
1808 | } else if (ex->fe_len > gex->fe_len) { | |
1809 | /* if the request is satisfied, then we try to find | |
1810 | * an extent that still satisfy the request, but is | |
1811 | * smaller than previous one */ | |
1812 | if (ex->fe_len < bex->fe_len) | |
1813 | *bex = *ex; | |
1814 | } | |
1815 | ||
1816 | ext4_mb_check_limits(ac, e4b, 0); | |
1817 | } | |
1818 | ||
089ceecc ES |
1819 | static noinline_for_stack |
1820 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1821 | struct ext4_buddy *e4b) |
1822 | { | |
1823 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1824 | ext4_group_t group = ex.fe_group; | |
1825 | int max; | |
1826 | int err; | |
1827 | ||
1828 | BUG_ON(ex.fe_len <= 0); | |
1829 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1830 | if (err) | |
1831 | return err; | |
1832 | ||
1833 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1834 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1835 | |
1836 | if (max > 0) { | |
1837 | ac->ac_b_ex = ex; | |
1838 | ext4_mb_use_best_found(ac, e4b); | |
1839 | } | |
1840 | ||
1841 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1842 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1843 | |
1844 | return 0; | |
1845 | } | |
1846 | ||
089ceecc ES |
1847 | static noinline_for_stack |
1848 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1849 | struct ext4_buddy *e4b) |
1850 | { | |
1851 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1852 | int max; | |
1853 | int err; | |
1854 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1855 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1856 | struct ext4_free_extent ex; |
1857 | ||
1858 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1859 | return 0; | |
838cd0cf YY |
1860 | if (grp->bb_free == 0) |
1861 | return 0; | |
c9de560d AT |
1862 | |
1863 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1864 | if (err) | |
1865 | return err; | |
1866 | ||
163a203d DW |
1867 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
1868 | ext4_mb_unload_buddy(e4b); | |
1869 | return 0; | |
1870 | } | |
1871 | ||
c9de560d | 1872 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 1873 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 1874 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 1875 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
1876 | |
1877 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1878 | ext4_fsblk_t start; | |
1879 | ||
5661bd68 AM |
1880 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1881 | ex.fe_start; | |
c9de560d AT |
1882 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1883 | if (do_div(start, sbi->s_stripe) == 0) { | |
1884 | ac->ac_found++; | |
1885 | ac->ac_b_ex = ex; | |
1886 | ext4_mb_use_best_found(ac, e4b); | |
1887 | } | |
1888 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1889 | BUG_ON(ex.fe_len <= 0); | |
1890 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1891 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1892 | ac->ac_found++; | |
1893 | ac->ac_b_ex = ex; | |
1894 | ext4_mb_use_best_found(ac, e4b); | |
1895 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1896 | /* Sometimes, caller may want to merge even small | |
1897 | * number of blocks to an existing extent */ | |
1898 | BUG_ON(ex.fe_len <= 0); | |
1899 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1900 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1901 | ac->ac_found++; | |
1902 | ac->ac_b_ex = ex; | |
1903 | ext4_mb_use_best_found(ac, e4b); | |
1904 | } | |
1905 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1906 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1907 | |
1908 | return 0; | |
1909 | } | |
1910 | ||
1911 | /* | |
1912 | * The routine scans buddy structures (not bitmap!) from given order | |
1913 | * to max order and tries to find big enough chunk to satisfy the req | |
1914 | */ | |
089ceecc ES |
1915 | static noinline_for_stack |
1916 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1917 | struct ext4_buddy *e4b) |
1918 | { | |
1919 | struct super_block *sb = ac->ac_sb; | |
1920 | struct ext4_group_info *grp = e4b->bd_info; | |
1921 | void *buddy; | |
1922 | int i; | |
1923 | int k; | |
1924 | int max; | |
1925 | ||
1926 | BUG_ON(ac->ac_2order <= 0); | |
1927 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1928 | if (grp->bb_counters[i] == 0) | |
1929 | continue; | |
1930 | ||
1931 | buddy = mb_find_buddy(e4b, i, &max); | |
1932 | BUG_ON(buddy == NULL); | |
1933 | ||
ffad0a44 | 1934 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
1935 | if (k >= max) { |
1936 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
1937 | "%d free clusters of order %d. But found 0", | |
1938 | grp->bb_counters[i], i); | |
1939 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
1940 | e4b->bd_group, | |
1941 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
1942 | break; | |
1943 | } | |
c9de560d AT |
1944 | ac->ac_found++; |
1945 | ||
1946 | ac->ac_b_ex.fe_len = 1 << i; | |
1947 | ac->ac_b_ex.fe_start = k << i; | |
1948 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1949 | ||
1950 | ext4_mb_use_best_found(ac, e4b); | |
1951 | ||
1952 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1953 | ||
1954 | if (EXT4_SB(sb)->s_mb_stats) | |
1955 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1956 | ||
1957 | break; | |
1958 | } | |
1959 | } | |
1960 | ||
1961 | /* | |
1962 | * The routine scans the group and measures all found extents. | |
1963 | * In order to optimize scanning, caller must pass number of | |
1964 | * free blocks in the group, so the routine can know upper limit. | |
1965 | */ | |
089ceecc ES |
1966 | static noinline_for_stack |
1967 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1968 | struct ext4_buddy *e4b) |
1969 | { | |
1970 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 1971 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1972 | struct ext4_free_extent ex; |
1973 | int i; | |
1974 | int free; | |
1975 | ||
1976 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
1977 | if (WARN_ON(free <= 0)) |
1978 | return; | |
c9de560d AT |
1979 | |
1980 | i = e4b->bd_info->bb_first_free; | |
1981 | ||
1982 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1983 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1984 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1985 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1986 | /* |
e56eb659 | 1987 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1988 | * free blocks even though group info says we |
1989 | * we have free blocks | |
1990 | */ | |
e29136f8 | 1991 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1992 | "%d free clusters as per " |
fde4d95a | 1993 | "group info. But bitmap says 0", |
26346ff6 | 1994 | free); |
736dedbb WS |
1995 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1996 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
1997 | break; |
1998 | } | |
1999 | ||
15c006a2 | 2000 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2001 | if (WARN_ON(ex.fe_len <= 0)) |
2002 | break; | |
26346ff6 | 2003 | if (free < ex.fe_len) { |
e29136f8 | 2004 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2005 | "%d free clusters as per " |
fde4d95a | 2006 | "group info. But got %d blocks", |
26346ff6 | 2007 | free, ex.fe_len); |
736dedbb WS |
2008 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2009 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
2010 | /* |
2011 | * The number of free blocks differs. This mostly | |
2012 | * indicate that the bitmap is corrupt. So exit | |
2013 | * without claiming the space. | |
2014 | */ | |
2015 | break; | |
26346ff6 | 2016 | } |
ab0c00fc | 2017 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2018 | ext4_mb_measure_extent(ac, &ex, e4b); |
2019 | ||
2020 | i += ex.fe_len; | |
2021 | free -= ex.fe_len; | |
2022 | } | |
2023 | ||
2024 | ext4_mb_check_limits(ac, e4b, 1); | |
2025 | } | |
2026 | ||
2027 | /* | |
2028 | * This is a special case for storages like raid5 | |
506bf2d8 | 2029 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2030 | */ |
089ceecc ES |
2031 | static noinline_for_stack |
2032 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2033 | struct ext4_buddy *e4b) |
2034 | { | |
2035 | struct super_block *sb = ac->ac_sb; | |
2036 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2037 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2038 | struct ext4_free_extent ex; |
2039 | ext4_fsblk_t first_group_block; | |
2040 | ext4_fsblk_t a; | |
2041 | ext4_grpblk_t i; | |
2042 | int max; | |
2043 | ||
2044 | BUG_ON(sbi->s_stripe == 0); | |
2045 | ||
2046 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2047 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2048 | ||
c9de560d AT |
2049 | a = first_group_block + sbi->s_stripe - 1; |
2050 | do_div(a, sbi->s_stripe); | |
2051 | i = (a * sbi->s_stripe) - first_group_block; | |
2052 | ||
7137d7a4 | 2053 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2054 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2055 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2056 | if (max >= sbi->s_stripe) { |
2057 | ac->ac_found++; | |
ab0c00fc | 2058 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2059 | ac->ac_b_ex = ex; |
2060 | ext4_mb_use_best_found(ac, e4b); | |
2061 | break; | |
2062 | } | |
2063 | } | |
2064 | i += sbi->s_stripe; | |
2065 | } | |
2066 | } | |
2067 | ||
42ac1848 LC |
2068 | /* |
2069 | * This is now called BEFORE we load the buddy bitmap. | |
2070 | * Returns either 1 or 0 indicating that the group is either suitable | |
2071 | * for the allocation or not. In addition it can also return negative | |
2072 | * error code when something goes wrong. | |
2073 | */ | |
c9de560d AT |
2074 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
2075 | ext4_group_t group, int cr) | |
2076 | { | |
2077 | unsigned free, fragments; | |
a4912123 | 2078 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2079 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2080 | ||
2081 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2082 | |
01fc48e8 TT |
2083 | free = grp->bb_free; |
2084 | if (free == 0) | |
2085 | return 0; | |
2086 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2087 | return 0; | |
2088 | ||
163a203d DW |
2089 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
2090 | return 0; | |
2091 | ||
8a57d9d6 CW |
2092 | /* We only do this if the grp has never been initialized */ |
2093 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 2094 | int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS); |
8a57d9d6 | 2095 | if (ret) |
42ac1848 | 2096 | return ret; |
8a57d9d6 | 2097 | } |
c9de560d | 2098 | |
c9de560d | 2099 | fragments = grp->bb_fragments; |
c9de560d AT |
2100 | if (fragments == 0) |
2101 | return 0; | |
2102 | ||
2103 | switch (cr) { | |
2104 | case 0: | |
2105 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2106 | |
a4912123 TT |
2107 | /* Avoid using the first bg of a flexgroup for data files */ |
2108 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2109 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2110 | ((group % flex_size) == 0)) | |
2111 | return 0; | |
2112 | ||
40ae3487 TT |
2113 | if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || |
2114 | (free / fragments) >= ac->ac_g_ex.fe_len) | |
2115 | return 1; | |
2116 | ||
2117 | if (grp->bb_largest_free_order < ac->ac_2order) | |
2118 | return 0; | |
2119 | ||
8a57d9d6 | 2120 | return 1; |
c9de560d AT |
2121 | case 1: |
2122 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
2123 | return 1; | |
2124 | break; | |
2125 | case 2: | |
2126 | if (free >= ac->ac_g_ex.fe_len) | |
2127 | return 1; | |
2128 | break; | |
2129 | case 3: | |
2130 | return 1; | |
2131 | default: | |
2132 | BUG(); | |
2133 | } | |
2134 | ||
2135 | return 0; | |
2136 | } | |
2137 | ||
4ddfef7b ES |
2138 | static noinline_for_stack int |
2139 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2140 | { |
8df9675f | 2141 | ext4_group_t ngroups, group, i; |
bbc4ec77 | 2142 | int cr = -1; |
42ac1848 | 2143 | int err = 0, first_err = 0; |
c9de560d AT |
2144 | struct ext4_sb_info *sbi; |
2145 | struct super_block *sb; | |
2146 | struct ext4_buddy e4b; | |
c9de560d AT |
2147 | |
2148 | sb = ac->ac_sb; | |
2149 | sbi = EXT4_SB(sb); | |
8df9675f | 2150 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2151 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2152 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2153 | ngroups = sbi->s_blockfile_groups; |
2154 | ||
c9de560d AT |
2155 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2156 | ||
2157 | /* first, try the goal */ | |
2158 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2159 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2160 | goto out; | |
2161 | ||
2162 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2163 | goto out; | |
2164 | ||
2165 | /* | |
2166 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
2167 | * if ac2_order is set we also set criteria to 0 so that we | |
2168 | * try exact allocation using buddy. | |
2169 | */ | |
2170 | i = fls(ac->ac_g_ex.fe_len); | |
2171 | ac->ac_2order = 0; | |
2172 | /* | |
2173 | * We search using buddy data only if the order of the request | |
2174 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2175 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2176 | * We also support searching for power-of-two requests only for |
2177 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2178 | */ |
d9b22cf9 | 2179 | if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) { |
c9de560d AT |
2180 | /* |
2181 | * This should tell if fe_len is exactly power of 2 | |
2182 | */ | |
2183 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d JC |
2184 | ac->ac_2order = array_index_nospec(i - 1, |
2185 | sb->s_blocksize_bits + 2); | |
c9de560d AT |
2186 | } |
2187 | ||
4ba74d00 TT |
2188 | /* if stream allocation is enabled, use global goal */ |
2189 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2190 | /* TBD: may be hot point */ |
2191 | spin_lock(&sbi->s_md_lock); | |
2192 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2193 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2194 | spin_unlock(&sbi->s_md_lock); | |
2195 | } | |
4ba74d00 | 2196 | |
c9de560d AT |
2197 | /* Let's just scan groups to find more-less suitable blocks */ |
2198 | cr = ac->ac_2order ? 0 : 1; | |
2199 | /* | |
2200 | * cr == 0 try to get exact allocation, | |
2201 | * cr == 3 try to get anything | |
2202 | */ | |
2203 | repeat: | |
2204 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2205 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2206 | /* |
2207 | * searching for the right group start | |
2208 | * from the goal value specified | |
2209 | */ | |
2210 | group = ac->ac_g_ex.fe_group; | |
2211 | ||
8df9675f | 2212 | for (i = 0; i < ngroups; group++, i++) { |
42ac1848 | 2213 | int ret = 0; |
2ed5724d | 2214 | cond_resched(); |
e6155736 LM |
2215 | /* |
2216 | * Artificially restricted ngroups for non-extent | |
2217 | * files makes group > ngroups possible on first loop. | |
2218 | */ | |
2219 | if (group >= ngroups) | |
c9de560d AT |
2220 | group = 0; |
2221 | ||
8a57d9d6 | 2222 | /* This now checks without needing the buddy page */ |
42ac1848 LC |
2223 | ret = ext4_mb_good_group(ac, group, cr); |
2224 | if (ret <= 0) { | |
2225 | if (!first_err) | |
2226 | first_err = ret; | |
c9de560d | 2227 | continue; |
42ac1848 | 2228 | } |
c9de560d | 2229 | |
c9de560d AT |
2230 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2231 | if (err) | |
2232 | goto out; | |
2233 | ||
2234 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2235 | |
2236 | /* | |
2237 | * We need to check again after locking the | |
2238 | * block group | |
2239 | */ | |
42ac1848 LC |
2240 | ret = ext4_mb_good_group(ac, group, cr); |
2241 | if (ret <= 0) { | |
c9de560d | 2242 | ext4_unlock_group(sb, group); |
e39e07fd | 2243 | ext4_mb_unload_buddy(&e4b); |
42ac1848 LC |
2244 | if (!first_err) |
2245 | first_err = ret; | |
c9de560d AT |
2246 | continue; |
2247 | } | |
2248 | ||
2249 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2250 | if (cr == 0) |
c9de560d | 2251 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2252 | else if (cr == 1 && sbi->s_stripe && |
2253 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2254 | ext4_mb_scan_aligned(ac, &e4b); |
2255 | else | |
2256 | ext4_mb_complex_scan_group(ac, &e4b); | |
2257 | ||
2258 | ext4_unlock_group(sb, group); | |
e39e07fd | 2259 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2260 | |
2261 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2262 | break; | |
2263 | } | |
2264 | } | |
2265 | ||
2266 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2267 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2268 | /* | |
2269 | * We've been searching too long. Let's try to allocate | |
2270 | * the best chunk we've found so far | |
2271 | */ | |
2272 | ||
2273 | ext4_mb_try_best_found(ac, &e4b); | |
2274 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2275 | /* | |
2276 | * Someone more lucky has already allocated it. | |
2277 | * The only thing we can do is just take first | |
2278 | * found block(s) | |
2279 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2280 | */ | |
2281 | ac->ac_b_ex.fe_group = 0; | |
2282 | ac->ac_b_ex.fe_start = 0; | |
2283 | ac->ac_b_ex.fe_len = 0; | |
2284 | ac->ac_status = AC_STATUS_CONTINUE; | |
2285 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2286 | cr = 3; | |
2287 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2288 | goto repeat; | |
2289 | } | |
2290 | } | |
2291 | out: | |
42ac1848 LC |
2292 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2293 | err = first_err; | |
bbc4ec77 RH |
2294 | |
2295 | mb_debug(1, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", | |
2296 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, | |
2297 | ac->ac_flags, cr, err); | |
c9de560d AT |
2298 | return err; |
2299 | } | |
2300 | ||
c9de560d AT |
2301 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2302 | { | |
247dbed8 | 2303 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2304 | ext4_group_t group; |
2305 | ||
8df9675f | 2306 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2307 | return NULL; |
c9de560d | 2308 | group = *pos + 1; |
a9df9a49 | 2309 | return (void *) ((unsigned long) group); |
c9de560d AT |
2310 | } |
2311 | ||
2312 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2313 | { | |
247dbed8 | 2314 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2315 | ext4_group_t group; |
2316 | ||
2317 | ++*pos; | |
8df9675f | 2318 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2319 | return NULL; |
2320 | group = *pos + 1; | |
a9df9a49 | 2321 | return (void *) ((unsigned long) group); |
c9de560d AT |
2322 | } |
2323 | ||
2324 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2325 | { | |
247dbed8 | 2326 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
a9df9a49 | 2327 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2328 | int i; |
1c8457ca | 2329 | int err, buddy_loaded = 0; |
c9de560d | 2330 | struct ext4_buddy e4b; |
1c8457ca | 2331 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2332 | unsigned char blocksize_bits = min_t(unsigned char, |
2333 | sb->s_blocksize_bits, | |
2334 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2335 | struct sg { |
2336 | struct ext4_group_info info; | |
b80b32b6 | 2337 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2338 | } sg; |
2339 | ||
2340 | group--; | |
2341 | if (group == 0) | |
97b4af2f RV |
2342 | seq_puts(seq, "#group: free frags first [" |
2343 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2344 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2345 | |
b80b32b6 TT |
2346 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2347 | sizeof(struct ext4_group_info); | |
2348 | ||
1c8457ca AK |
2349 | grinfo = ext4_get_group_info(sb, group); |
2350 | /* Load the group info in memory only if not already loaded. */ | |
2351 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2352 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2353 | if (err) { | |
2354 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2355 | return 0; | |
2356 | } | |
2357 | buddy_loaded = 1; | |
c9de560d | 2358 | } |
1c8457ca | 2359 | |
b80b32b6 | 2360 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2361 | |
2362 | if (buddy_loaded) | |
2363 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2364 | |
a9df9a49 | 2365 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2366 | sg.info.bb_fragments, sg.info.bb_first_free); |
2367 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2368 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d AT |
2369 | sg.info.bb_counters[i] : 0); |
2370 | seq_printf(seq, " ]\n"); | |
2371 | ||
2372 | return 0; | |
2373 | } | |
2374 | ||
2375 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2376 | { | |
2377 | } | |
2378 | ||
247dbed8 | 2379 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2380 | .start = ext4_mb_seq_groups_start, |
2381 | .next = ext4_mb_seq_groups_next, | |
2382 | .stop = ext4_mb_seq_groups_stop, | |
2383 | .show = ext4_mb_seq_groups_show, | |
2384 | }; | |
2385 | ||
fb1813f4 CW |
2386 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2387 | { | |
2388 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2389 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2390 | ||
2391 | BUG_ON(!cachep); | |
2392 | return cachep; | |
2393 | } | |
5f21b0e6 | 2394 | |
28623c2f TT |
2395 | /* |
2396 | * Allocate the top-level s_group_info array for the specified number | |
2397 | * of groups | |
2398 | */ | |
2399 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2400 | { | |
2401 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2402 | unsigned size; | |
df3da4ea | 2403 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
2404 | |
2405 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2406 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2407 | if (size <= sbi->s_group_info_size) | |
2408 | return 0; | |
2409 | ||
2410 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 2411 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
2412 | if (!new_groupinfo) { |
2413 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2414 | return -ENOMEM; | |
2415 | } | |
df3da4ea SJS |
2416 | rcu_read_lock(); |
2417 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
2418 | if (old_groupinfo) | |
2419 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 2420 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
2421 | rcu_read_unlock(); |
2422 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 2423 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
2424 | if (old_groupinfo) |
2425 | ext4_kvfree_array_rcu(old_groupinfo); | |
28623c2f TT |
2426 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
2427 | sbi->s_group_info_size); | |
2428 | return 0; | |
2429 | } | |
2430 | ||
5f21b0e6 | 2431 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2432 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2433 | struct ext4_group_desc *desc) |
2434 | { | |
fb1813f4 | 2435 | int i; |
5f21b0e6 | 2436 | int metalen = 0; |
df3da4ea | 2437 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
2438 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2439 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2440 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2441 | |
2442 | /* | |
2443 | * First check if this group is the first of a reserved block. | |
2444 | * If it's true, we have to allocate a new table of pointers | |
2445 | * to ext4_group_info structures | |
2446 | */ | |
2447 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2448 | metalen = sizeof(*meta_group_info) << | |
2449 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 2450 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 2451 | if (meta_group_info == NULL) { |
7f6a11e7 | 2452 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2453 | "for a buddy group"); |
5f21b0e6 FB |
2454 | goto exit_meta_group_info; |
2455 | } | |
df3da4ea SJS |
2456 | rcu_read_lock(); |
2457 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
2458 | rcu_read_unlock(); | |
5f21b0e6 FB |
2459 | } |
2460 | ||
df3da4ea | 2461 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
2462 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
2463 | ||
4fdb5543 | 2464 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 2465 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2466 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2467 | goto exit_group_info; |
2468 | } | |
2469 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2470 | &(meta_group_info[i]->bb_state)); | |
2471 | ||
2472 | /* | |
2473 | * initialize bb_free to be able to skip | |
2474 | * empty groups without initialization | |
2475 | */ | |
8844618d TT |
2476 | if (ext4_has_group_desc_csum(sb) && |
2477 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 2478 | meta_group_info[i]->bb_free = |
cff1dfd7 | 2479 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2480 | } else { |
2481 | meta_group_info[i]->bb_free = | |
021b65bb | 2482 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2483 | } |
2484 | ||
2485 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2486 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2487 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2488 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 | 2489 | |
a3450215 | 2490 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
2491 | return 0; |
2492 | ||
2493 | exit_group_info: | |
2494 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2495 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
2496 | struct ext4_group_info ***group_info; |
2497 | ||
2498 | rcu_read_lock(); | |
2499 | group_info = rcu_dereference(sbi->s_group_info); | |
2500 | kfree(group_info[idx]); | |
2501 | group_info[idx] = NULL; | |
2502 | rcu_read_unlock(); | |
caaf7a29 | 2503 | } |
5f21b0e6 FB |
2504 | exit_meta_group_info: |
2505 | return -ENOMEM; | |
2506 | } /* ext4_mb_add_groupinfo */ | |
2507 | ||
c9de560d AT |
2508 | static int ext4_mb_init_backend(struct super_block *sb) |
2509 | { | |
8df9675f | 2510 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2511 | ext4_group_t i; |
c9de560d | 2512 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2513 | int err; |
5f21b0e6 | 2514 | struct ext4_group_desc *desc; |
df3da4ea | 2515 | struct ext4_group_info ***group_info; |
fb1813f4 | 2516 | struct kmem_cache *cachep; |
5f21b0e6 | 2517 | |
28623c2f TT |
2518 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2519 | if (err) | |
2520 | return err; | |
c9de560d | 2521 | |
c9de560d AT |
2522 | sbi->s_buddy_cache = new_inode(sb); |
2523 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2524 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2525 | goto err_freesgi; |
2526 | } | |
48e6061b YJ |
2527 | /* To avoid potentially colliding with an valid on-disk inode number, |
2528 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2529 | * not in the inode hash, so it should never be found by iget(), but | |
2530 | * this will avoid confusion if it ever shows up during debugging. */ | |
2531 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2532 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2533 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 2534 | cond_resched(); |
c9de560d AT |
2535 | desc = ext4_get_group_desc(sb, i, NULL); |
2536 | if (desc == NULL) { | |
9d8b9ec4 | 2537 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2538 | goto err_freebuddy; |
2539 | } | |
5f21b0e6 FB |
2540 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2541 | goto err_freebuddy; | |
c9de560d AT |
2542 | } |
2543 | ||
2544 | return 0; | |
2545 | ||
2546 | err_freebuddy: | |
fb1813f4 | 2547 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2548 | while (i-- > 0) |
fb1813f4 | 2549 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2550 | i = sbi->s_group_info_size; |
df3da4ea SJS |
2551 | rcu_read_lock(); |
2552 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 2553 | while (i-- > 0) |
df3da4ea SJS |
2554 | kfree(group_info[i]); |
2555 | rcu_read_unlock(); | |
c9de560d AT |
2556 | iput(sbi->s_buddy_cache); |
2557 | err_freesgi: | |
df3da4ea SJS |
2558 | rcu_read_lock(); |
2559 | kvfree(rcu_dereference(sbi->s_group_info)); | |
2560 | rcu_read_unlock(); | |
c9de560d AT |
2561 | return -ENOMEM; |
2562 | } | |
2563 | ||
2892c15d ES |
2564 | static void ext4_groupinfo_destroy_slabs(void) |
2565 | { | |
2566 | int i; | |
2567 | ||
2568 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 2569 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
2570 | ext4_groupinfo_caches[i] = NULL; |
2571 | } | |
2572 | } | |
2573 | ||
2574 | static int ext4_groupinfo_create_slab(size_t size) | |
2575 | { | |
2576 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2577 | int slab_size; | |
2578 | int blocksize_bits = order_base_2(size); | |
2579 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2580 | struct kmem_cache *cachep; | |
2581 | ||
2582 | if (cache_index >= NR_GRPINFO_CACHES) | |
2583 | return -EINVAL; | |
2584 | ||
2585 | if (unlikely(cache_index < 0)) | |
2586 | cache_index = 0; | |
2587 | ||
2588 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2589 | if (ext4_groupinfo_caches[cache_index]) { | |
2590 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2591 | return 0; /* Already created */ | |
2592 | } | |
2593 | ||
2594 | slab_size = offsetof(struct ext4_group_info, | |
2595 | bb_counters[blocksize_bits + 2]); | |
2596 | ||
2597 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2598 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2599 | NULL); | |
2600 | ||
823ba01f TM |
2601 | ext4_groupinfo_caches[cache_index] = cachep; |
2602 | ||
2892c15d ES |
2603 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2604 | if (!cachep) { | |
9d8b9ec4 TT |
2605 | printk(KERN_EMERG |
2606 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2607 | return -ENOMEM; |
2608 | } | |
2609 | ||
2892c15d ES |
2610 | return 0; |
2611 | } | |
2612 | ||
9d99012f | 2613 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2614 | { |
2615 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2616 | unsigned i, j; |
935244cd | 2617 | unsigned offset, offset_incr; |
c9de560d | 2618 | unsigned max; |
74767c5a | 2619 | int ret; |
c9de560d | 2620 | |
1927805e | 2621 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2622 | |
2623 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2624 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2625 | ret = -ENOMEM; |
2626 | goto out; | |
c9de560d | 2627 | } |
ff7ef329 | 2628 | |
1927805e | 2629 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2630 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2631 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2632 | ret = -ENOMEM; |
2633 | goto out; | |
2634 | } | |
2635 | ||
2892c15d ES |
2636 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2637 | if (ret < 0) | |
2638 | goto out; | |
c9de560d AT |
2639 | |
2640 | /* order 0 is regular bitmap */ | |
2641 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2642 | sbi->s_mb_offsets[0] = 0; | |
2643 | ||
2644 | i = 1; | |
2645 | offset = 0; | |
935244cd | 2646 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
2647 | max = sb->s_blocksize << 2; |
2648 | do { | |
2649 | sbi->s_mb_offsets[i] = offset; | |
2650 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
2651 | offset += offset_incr; |
2652 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
2653 | max = max >> 1; |
2654 | i++; | |
2655 | } while (i <= sb->s_blocksize_bits + 1); | |
2656 | ||
c9de560d | 2657 | spin_lock_init(&sbi->s_md_lock); |
c9de560d | 2658 | spin_lock_init(&sbi->s_bal_lock); |
d08854f5 | 2659 | sbi->s_mb_free_pending = 0; |
a0154344 | 2660 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
c9de560d AT |
2661 | |
2662 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2663 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2664 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2665 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2666 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
2667 | /* |
2668 | * The default group preallocation is 512, which for 4k block | |
2669 | * sizes translates to 2 megabytes. However for bigalloc file | |
2670 | * systems, this is probably too big (i.e, if the cluster size | |
2671 | * is 1 megabyte, then group preallocation size becomes half a | |
2672 | * gigabyte!). As a default, we will keep a two megabyte | |
2673 | * group pralloc size for cluster sizes up to 64k, and after | |
2674 | * that, we will force a minimum group preallocation size of | |
2675 | * 32 clusters. This translates to 8 megs when the cluster | |
2676 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2677 | * which seems reasonable as a default. | |
2678 | */ | |
2679 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2680 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2681 | /* |
2682 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2683 | * to the lowest multiple of s_stripe which is bigger than | |
2684 | * the s_mb_group_prealloc as determined above. We want | |
2685 | * the preallocation size to be an exact multiple of the | |
2686 | * RAID stripe size so that preallocations don't fragment | |
2687 | * the stripes. | |
2688 | */ | |
2689 | if (sbi->s_stripe > 1) { | |
2690 | sbi->s_mb_group_prealloc = roundup( | |
2691 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2692 | } | |
c9de560d | 2693 | |
730c213c | 2694 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2695 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2696 | ret = -ENOMEM; |
029b10c5 | 2697 | goto out; |
c9de560d | 2698 | } |
730c213c | 2699 | for_each_possible_cpu(i) { |
c9de560d | 2700 | struct ext4_locality_group *lg; |
730c213c | 2701 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2702 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2703 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2704 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2705 | spin_lock_init(&lg->lg_prealloc_lock); |
2706 | } | |
2707 | ||
79a77c5a YJ |
2708 | /* init file for buddy data */ |
2709 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2710 | if (ret != 0) |
2711 | goto out_free_locality_groups; | |
79a77c5a | 2712 | |
7aa0baea TM |
2713 | return 0; |
2714 | ||
2715 | out_free_locality_groups: | |
2716 | free_percpu(sbi->s_locality_groups); | |
2717 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 2718 | out: |
7aa0baea TM |
2719 | kfree(sbi->s_mb_offsets); |
2720 | sbi->s_mb_offsets = NULL; | |
2721 | kfree(sbi->s_mb_maxs); | |
2722 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2723 | return ret; |
c9de560d AT |
2724 | } |
2725 | ||
955ce5f5 | 2726 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2727 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2728 | { | |
2729 | struct ext4_prealloc_space *pa; | |
2730 | struct list_head *cur, *tmp; | |
2731 | int count = 0; | |
2732 | ||
2733 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2734 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2735 | list_del(&pa->pa_group_list); | |
2736 | count++; | |
688f05a0 | 2737 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2738 | } |
2739 | if (count) | |
6ba495e9 | 2740 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2741 | |
2742 | } | |
2743 | ||
2744 | int ext4_mb_release(struct super_block *sb) | |
2745 | { | |
8df9675f | 2746 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2747 | ext4_group_t i; |
2748 | int num_meta_group_infos; | |
df3da4ea | 2749 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 2750 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 2751 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2752 | |
c9de560d | 2753 | if (sbi->s_group_info) { |
8df9675f | 2754 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 2755 | cond_resched(); |
c9de560d | 2756 | grinfo = ext4_get_group_info(sb, i); |
a3450215 | 2757 | mb_group_bb_bitmap_free(grinfo); |
c9de560d AT |
2758 | ext4_lock_group(sb, i); |
2759 | ext4_mb_cleanup_pa(grinfo); | |
2760 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2761 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2762 | } |
8df9675f | 2763 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2764 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2765 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
2766 | rcu_read_lock(); |
2767 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 2768 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
2769 | kfree(group_info[i]); |
2770 | kvfree(group_info); | |
2771 | rcu_read_unlock(); | |
c9de560d AT |
2772 | } |
2773 | kfree(sbi->s_mb_offsets); | |
2774 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 2775 | iput(sbi->s_buddy_cache); |
c9de560d | 2776 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
2777 | ext4_msg(sb, KERN_INFO, |
2778 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2779 | atomic_read(&sbi->s_bal_allocated), |
2780 | atomic_read(&sbi->s_bal_reqs), | |
2781 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2782 | ext4_msg(sb, KERN_INFO, |
2783 | "mballoc: %u extents scanned, %u goal hits, " | |
2784 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2785 | atomic_read(&sbi->s_bal_ex_scanned), |
2786 | atomic_read(&sbi->s_bal_goals), | |
2787 | atomic_read(&sbi->s_bal_2orders), | |
2788 | atomic_read(&sbi->s_bal_breaks), | |
2789 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2790 | ext4_msg(sb, KERN_INFO, |
2791 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2792 | sbi->s_mb_buddies_generated, |
c9de560d | 2793 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2794 | ext4_msg(sb, KERN_INFO, |
2795 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2796 | atomic_read(&sbi->s_mb_preallocated), |
2797 | atomic_read(&sbi->s_mb_discarded)); | |
2798 | } | |
2799 | ||
730c213c | 2800 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2801 | |
2802 | return 0; | |
2803 | } | |
2804 | ||
77ca6cdf | 2805 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
2806 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
2807 | struct bio **biop) | |
5c521830 | 2808 | { |
5c521830 JZ |
2809 | ext4_fsblk_t discard_block; |
2810 | ||
84130193 TT |
2811 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
2812 | ext4_group_first_block_no(sb, block_group)); | |
2813 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
2814 | trace_ext4_discard_blocks(sb, |
2815 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
2816 | if (biop) { |
2817 | return __blkdev_issue_discard(sb->s_bdev, | |
2818 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
2819 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
2820 | GFP_NOFS, 0, biop); | |
2821 | } else | |
2822 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
2823 | } |
2824 | ||
a0154344 DJ |
2825 | static void ext4_free_data_in_buddy(struct super_block *sb, |
2826 | struct ext4_free_data *entry) | |
c9de560d | 2827 | { |
c9de560d | 2828 | struct ext4_buddy e4b; |
c894058d | 2829 | struct ext4_group_info *db; |
d9f34504 | 2830 | int err, count = 0, count2 = 0; |
c9de560d | 2831 | |
18aadd47 BJ |
2832 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
2833 | entry->efd_count, entry->efd_group, entry); | |
c9de560d | 2834 | |
18aadd47 BJ |
2835 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
2836 | /* we expect to find existing buddy because it's pinned */ | |
2837 | BUG_ON(err != 0); | |
b90f6870 | 2838 | |
d08854f5 TT |
2839 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
2840 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
2841 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 2842 | |
18aadd47 BJ |
2843 | db = e4b.bd_info; |
2844 | /* there are blocks to put in buddy to make them really free */ | |
2845 | count += entry->efd_count; | |
2846 | count2++; | |
2847 | ext4_lock_group(sb, entry->efd_group); | |
2848 | /* Take it out of per group rb tree */ | |
2849 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
2850 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 2851 | |
18aadd47 BJ |
2852 | /* |
2853 | * Clear the trimmed flag for the group so that the next | |
2854 | * ext4_trim_fs can trim it. | |
2855 | * If the volume is mounted with -o discard, online discard | |
2856 | * is supported and the free blocks will be trimmed online. | |
2857 | */ | |
2858 | if (!test_opt(sb, DISCARD)) | |
2859 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 2860 | |
18aadd47 BJ |
2861 | if (!db->bb_free_root.rb_node) { |
2862 | /* No more items in the per group rb tree | |
2863 | * balance refcounts from ext4_mb_free_metadata() | |
2864 | */ | |
09cbfeaf KS |
2865 | put_page(e4b.bd_buddy_page); |
2866 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 2867 | } |
18aadd47 BJ |
2868 | ext4_unlock_group(sb, entry->efd_group); |
2869 | kmem_cache_free(ext4_free_data_cachep, entry); | |
2870 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2871 | |
6ba495e9 | 2872 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2873 | } |
2874 | ||
a0154344 DJ |
2875 | /* |
2876 | * This function is called by the jbd2 layer once the commit has finished, | |
2877 | * so we know we can free the blocks that were released with that commit. | |
2878 | */ | |
2879 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
2880 | { | |
2881 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2882 | struct ext4_free_data *entry, *tmp; | |
2883 | struct bio *discard_bio = NULL; | |
2884 | struct list_head freed_data_list; | |
2885 | struct list_head *cut_pos = NULL; | |
2886 | int err; | |
2887 | ||
2888 | INIT_LIST_HEAD(&freed_data_list); | |
2889 | ||
2890 | spin_lock(&sbi->s_md_lock); | |
2891 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
2892 | if (entry->efd_tid != commit_tid) | |
2893 | break; | |
2894 | cut_pos = &entry->efd_list; | |
2895 | } | |
2896 | if (cut_pos) | |
2897 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
2898 | cut_pos); | |
2899 | spin_unlock(&sbi->s_md_lock); | |
2900 | ||
2901 | if (test_opt(sb, DISCARD)) { | |
2902 | list_for_each_entry(entry, &freed_data_list, efd_list) { | |
2903 | err = ext4_issue_discard(sb, entry->efd_group, | |
2904 | entry->efd_start_cluster, | |
2905 | entry->efd_count, | |
2906 | &discard_bio); | |
2907 | if (err && err != -EOPNOTSUPP) { | |
2908 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
2909 | " group:%d block:%d count:%d failed" | |
2910 | " with %d", entry->efd_group, | |
2911 | entry->efd_start_cluster, | |
2912 | entry->efd_count, err); | |
2913 | } else if (err == -EOPNOTSUPP) | |
2914 | break; | |
2915 | } | |
2916 | ||
e4510577 | 2917 | if (discard_bio) { |
a0154344 | 2918 | submit_bio_wait(discard_bio); |
e4510577 DJ |
2919 | bio_put(discard_bio); |
2920 | } | |
a0154344 DJ |
2921 | } |
2922 | ||
2923 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
2924 | ext4_free_data_in_buddy(sb, entry); | |
2925 | } | |
2926 | ||
5dabfc78 | 2927 | int __init ext4_init_mballoc(void) |
c9de560d | 2928 | { |
16828088 TT |
2929 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2930 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 2931 | if (ext4_pspace_cachep == NULL) |
f283529a | 2932 | goto out; |
c9de560d | 2933 | |
16828088 TT |
2934 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2935 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
2936 | if (ext4_ac_cachep == NULL) |
2937 | goto out_pa_free; | |
c894058d | 2938 | |
18aadd47 BJ |
2939 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
2940 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
2941 | if (ext4_free_data_cachep == NULL) |
2942 | goto out_ac_free; | |
2943 | ||
c9de560d | 2944 | return 0; |
f283529a RH |
2945 | |
2946 | out_ac_free: | |
2947 | kmem_cache_destroy(ext4_ac_cachep); | |
2948 | out_pa_free: | |
2949 | kmem_cache_destroy(ext4_pspace_cachep); | |
2950 | out: | |
2951 | return -ENOMEM; | |
c9de560d AT |
2952 | } |
2953 | ||
5dabfc78 | 2954 | void ext4_exit_mballoc(void) |
c9de560d | 2955 | { |
60e6679e | 2956 | /* |
3e03f9ca JDB |
2957 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2958 | * before destroying the slab cache. | |
2959 | */ | |
2960 | rcu_barrier(); | |
c9de560d | 2961 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2962 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 2963 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 2964 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
2965 | } |
2966 | ||
2967 | ||
2968 | /* | |
73b2c716 | 2969 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
2970 | * Returns 0 if success or error code |
2971 | */ | |
4ddfef7b ES |
2972 | static noinline_for_stack int |
2973 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 2974 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
2975 | { |
2976 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
2977 | struct ext4_group_desc *gdp; |
2978 | struct buffer_head *gdp_bh; | |
2979 | struct ext4_sb_info *sbi; | |
2980 | struct super_block *sb; | |
2981 | ext4_fsblk_t block; | |
519deca0 | 2982 | int err, len; |
c9de560d AT |
2983 | |
2984 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2985 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2986 | ||
2987 | sb = ac->ac_sb; | |
2988 | sbi = EXT4_SB(sb); | |
c9de560d | 2989 | |
574ca174 | 2990 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
2991 | if (IS_ERR(bitmap_bh)) { |
2992 | err = PTR_ERR(bitmap_bh); | |
2993 | bitmap_bh = NULL; | |
c9de560d | 2994 | goto out_err; |
9008a58e | 2995 | } |
c9de560d | 2996 | |
5d601255 | 2997 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
c9de560d AT |
2998 | err = ext4_journal_get_write_access(handle, bitmap_bh); |
2999 | if (err) | |
3000 | goto out_err; | |
3001 | ||
3002 | err = -EIO; | |
3003 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3004 | if (!gdp) | |
3005 | goto out_err; | |
3006 | ||
a9df9a49 | 3007 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3008 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3009 | |
5d601255 | 3010 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
c9de560d AT |
3011 | err = ext4_journal_get_write_access(handle, gdp_bh); |
3012 | if (err) | |
3013 | goto out_err; | |
3014 | ||
bda00de7 | 3015 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3016 | |
53accfa9 | 3017 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 3018 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 3019 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3020 | "fs metadata", block, block+len); |
519deca0 | 3021 | /* File system mounted not to panic on error |
554a5ccc | 3022 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3023 | * We leak some of the blocks here. |
3024 | */ | |
955ce5f5 | 3025 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
3026 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3027 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 3028 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3029 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3030 | if (!err) |
554a5ccc | 3031 | err = -EFSCORRUPTED; |
519deca0 | 3032 | goto out_err; |
c9de560d | 3033 | } |
955ce5f5 AK |
3034 | |
3035 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3036 | #ifdef AGGRESSIVE_CHECK |
3037 | { | |
3038 | int i; | |
3039 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3040 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3041 | bitmap_bh->b_data)); | |
3042 | } | |
3043 | } | |
3044 | #endif | |
c3e94d1d YY |
3045 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3046 | ac->ac_b_ex.fe_len); | |
8844618d TT |
3047 | if (ext4_has_group_desc_csum(sb) && |
3048 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3049 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3050 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3051 | ext4_free_clusters_after_init(sb, |
021b65bb | 3052 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3053 | } |
021b65bb TT |
3054 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3055 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3056 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3057 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3058 | |
3059 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3060 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3061 | /* |
6bc6e63f | 3062 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3063 | */ |
6bc6e63f AK |
3064 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3065 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3066 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3067 | reserv_clstrs); | |
c9de560d | 3068 | |
772cb7c8 JS |
3069 | if (sbi->s_log_groups_per_flex) { |
3070 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3071 | ac->ac_b_ex.fe_group); | |
90ba983f | 3072 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3073 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3074 | flex_group)->free_clusters); | |
772cb7c8 JS |
3075 | } |
3076 | ||
0390131b | 3077 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3078 | if (err) |
3079 | goto out_err; | |
0390131b | 3080 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3081 | |
3082 | out_err: | |
42a10add | 3083 | brelse(bitmap_bh); |
c9de560d AT |
3084 | return err; |
3085 | } | |
3086 | ||
3087 | /* | |
3088 | * here we normalize request for locality group | |
d7a1fee1 DE |
3089 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3090 | * s_strip if we set the same via mount option. | |
3091 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3092 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3093 | * |
3094 | * XXX: should we try to preallocate more than the group has now? | |
3095 | */ | |
3096 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3097 | { | |
3098 | struct super_block *sb = ac->ac_sb; | |
3099 | struct ext4_locality_group *lg = ac->ac_lg; | |
3100 | ||
3101 | BUG_ON(lg == NULL); | |
d7a1fee1 | 3102 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 3103 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
3104 | current->pid, ac->ac_g_ex.fe_len); |
3105 | } | |
3106 | ||
3107 | /* | |
3108 | * Normalization means making request better in terms of | |
3109 | * size and alignment | |
3110 | */ | |
4ddfef7b ES |
3111 | static noinline_for_stack void |
3112 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3113 | struct ext4_allocation_request *ar) |
3114 | { | |
53accfa9 | 3115 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3116 | int bsbits, max; |
3117 | ext4_lblk_t end; | |
1592d2c5 CW |
3118 | loff_t size, start_off; |
3119 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3120 | ext4_lblk_t start; |
c9de560d | 3121 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 3122 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3123 | |
3124 | /* do normalize only data requests, metadata requests | |
3125 | do not need preallocation */ | |
3126 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3127 | return; | |
3128 | ||
3129 | /* sometime caller may want exact blocks */ | |
3130 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3131 | return; | |
3132 | ||
3133 | /* caller may indicate that preallocation isn't | |
3134 | * required (it's a tail, for example) */ | |
3135 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3136 | return; | |
3137 | ||
3138 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3139 | ext4_mb_normalize_group_request(ac); | |
3140 | return ; | |
3141 | } | |
3142 | ||
3143 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3144 | ||
3145 | /* first, let's learn actual file size | |
3146 | * given current request is allocated */ | |
53accfa9 | 3147 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
3148 | size = size << bsbits; |
3149 | if (size < i_size_read(ac->ac_inode)) | |
3150 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 3151 | orig_size = size; |
c9de560d | 3152 | |
1930479c VC |
3153 | /* max size of free chunks */ |
3154 | max = 2 << bsbits; | |
c9de560d | 3155 | |
1930479c VC |
3156 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3157 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3158 | |
3159 | /* first, try to predict filesize */ | |
3160 | /* XXX: should this table be tunable? */ | |
3161 | start_off = 0; | |
3162 | if (size <= 16 * 1024) { | |
3163 | size = 16 * 1024; | |
3164 | } else if (size <= 32 * 1024) { | |
3165 | size = 32 * 1024; | |
3166 | } else if (size <= 64 * 1024) { | |
3167 | size = 64 * 1024; | |
3168 | } else if (size <= 128 * 1024) { | |
3169 | size = 128 * 1024; | |
3170 | } else if (size <= 256 * 1024) { | |
3171 | size = 256 * 1024; | |
3172 | } else if (size <= 512 * 1024) { | |
3173 | size = 512 * 1024; | |
3174 | } else if (size <= 1024 * 1024) { | |
3175 | size = 1024 * 1024; | |
1930479c | 3176 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3177 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3178 | (21 - bsbits)) << 21; |
3179 | size = 2 * 1024 * 1024; | |
3180 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3181 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3182 | (22 - bsbits)) << 22; | |
3183 | size = 4 * 1024 * 1024; | |
3184 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3185 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3186 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3187 | (23 - bsbits)) << 23; | |
3188 | size = 8 * 1024 * 1024; | |
3189 | } else { | |
b27b1535 XW |
3190 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
3191 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
3192 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 3193 | } |
5a0790c2 AK |
3194 | size = size >> bsbits; |
3195 | start = start_off >> bsbits; | |
c9de560d AT |
3196 | |
3197 | /* don't cover already allocated blocks in selected range */ | |
3198 | if (ar->pleft && start <= ar->lleft) { | |
3199 | size -= ar->lleft + 1 - start; | |
3200 | start = ar->lleft + 1; | |
3201 | } | |
3202 | if (ar->pright && start + size - 1 >= ar->lright) | |
3203 | size -= start + size - ar->lright; | |
3204 | ||
cd648b8a JK |
3205 | /* |
3206 | * Trim allocation request for filesystems with artificially small | |
3207 | * groups. | |
3208 | */ | |
3209 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
3210 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
3211 | ||
c9de560d AT |
3212 | end = start + size; |
3213 | ||
3214 | /* check we don't cross already preallocated blocks */ | |
3215 | rcu_read_lock(); | |
9a0762c5 | 3216 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3217 | ext4_lblk_t pa_end; |
c9de560d | 3218 | |
c9de560d AT |
3219 | if (pa->pa_deleted) |
3220 | continue; | |
3221 | spin_lock(&pa->pa_lock); | |
3222 | if (pa->pa_deleted) { | |
3223 | spin_unlock(&pa->pa_lock); | |
3224 | continue; | |
3225 | } | |
3226 | ||
53accfa9 TT |
3227 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3228 | pa->pa_len); | |
c9de560d AT |
3229 | |
3230 | /* PA must not overlap original request */ | |
3231 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3232 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3233 | ||
38877f4e ES |
3234 | /* skip PAs this normalized request doesn't overlap with */ |
3235 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3236 | spin_unlock(&pa->pa_lock); |
3237 | continue; | |
3238 | } | |
3239 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3240 | ||
38877f4e | 3241 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3242 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3243 | BUG_ON(pa_end < start); | |
3244 | start = pa_end; | |
38877f4e | 3245 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3246 | BUG_ON(pa->pa_lstart > end); |
3247 | end = pa->pa_lstart; | |
3248 | } | |
3249 | spin_unlock(&pa->pa_lock); | |
3250 | } | |
3251 | rcu_read_unlock(); | |
3252 | size = end - start; | |
3253 | ||
3254 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3255 | rcu_read_lock(); | |
9a0762c5 | 3256 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3257 | ext4_lblk_t pa_end; |
53accfa9 | 3258 | |
c9de560d AT |
3259 | spin_lock(&pa->pa_lock); |
3260 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3261 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3262 | pa->pa_len); | |
c9de560d AT |
3263 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3264 | } | |
3265 | spin_unlock(&pa->pa_lock); | |
3266 | } | |
3267 | rcu_read_unlock(); | |
3268 | ||
3269 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3270 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3271 | ext4_msg(ac->ac_sb, KERN_ERR, |
3272 | "start %lu, size %lu, fe_logical %lu", | |
3273 | (unsigned long) start, (unsigned long) size, | |
3274 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 3275 | BUG(); |
c9de560d | 3276 | } |
b5b60778 | 3277 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3278 | |
3279 | /* now prepare goal request */ | |
3280 | ||
3281 | /* XXX: is it better to align blocks WRT to logical | |
3282 | * placement or satisfy big request as is */ | |
3283 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3284 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3285 | |
3286 | /* define goal start in order to merge */ | |
3287 | if (ar->pright && (ar->lright == (start + size))) { | |
3288 | /* merge to the right */ | |
3289 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3290 | &ac->ac_f_ex.fe_group, | |
3291 | &ac->ac_f_ex.fe_start); | |
3292 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3293 | } | |
3294 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3295 | /* merge to the left */ | |
3296 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3297 | &ac->ac_f_ex.fe_group, | |
3298 | &ac->ac_f_ex.fe_start); | |
3299 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3300 | } | |
3301 | ||
004379d0 RH |
3302 | mb_debug(1, "goal: %lld(was %lld) blocks at %u\n", size, orig_size, |
3303 | start); | |
c9de560d AT |
3304 | } |
3305 | ||
3306 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3307 | { | |
3308 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3309 | ||
3310 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3311 | atomic_inc(&sbi->s_bal_reqs); | |
3312 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3313 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3314 | atomic_inc(&sbi->s_bal_success); |
3315 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3316 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3317 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3318 | atomic_inc(&sbi->s_bal_goals); | |
3319 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3320 | atomic_inc(&sbi->s_bal_breaks); | |
3321 | } | |
3322 | ||
296c355c TT |
3323 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3324 | trace_ext4_mballoc_alloc(ac); | |
3325 | else | |
3326 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3327 | } |
3328 | ||
b844167e CW |
3329 | /* |
3330 | * Called on failure; free up any blocks from the inode PA for this | |
3331 | * context. We don't need this for MB_GROUP_PA because we only change | |
3332 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3333 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3334 | */ | |
3335 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3336 | { | |
3337 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
3338 | struct ext4_buddy e4b; |
3339 | int err; | |
b844167e | 3340 | |
86f0afd4 | 3341 | if (pa == NULL) { |
c99d1e6e TT |
3342 | if (ac->ac_f_ex.fe_len == 0) |
3343 | return; | |
86f0afd4 TT |
3344 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
3345 | if (err) { | |
3346 | /* | |
3347 | * This should never happen since we pin the | |
3348 | * pages in the ext4_allocation_context so | |
3349 | * ext4_mb_load_buddy() should never fail. | |
3350 | */ | |
3351 | WARN(1, "mb_load_buddy failed (%d)", err); | |
3352 | return; | |
3353 | } | |
3354 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
3355 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
3356 | ac->ac_f_ex.fe_len); | |
3357 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 3358 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
3359 | return; |
3360 | } | |
3361 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 3362 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
3363 | } |
3364 | ||
c9de560d AT |
3365 | /* |
3366 | * use blocks preallocated to inode | |
3367 | */ | |
3368 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3369 | struct ext4_prealloc_space *pa) | |
3370 | { | |
53accfa9 | 3371 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3372 | ext4_fsblk_t start; |
3373 | ext4_fsblk_t end; | |
3374 | int len; | |
3375 | ||
3376 | /* found preallocated blocks, use them */ | |
3377 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3378 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3379 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3380 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3381 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3382 | &ac->ac_b_ex.fe_start); | |
3383 | ac->ac_b_ex.fe_len = len; | |
3384 | ac->ac_status = AC_STATUS_FOUND; | |
3385 | ac->ac_pa = pa; | |
3386 | ||
3387 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3388 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3389 | BUG_ON(pa->pa_free < len); |
3390 | pa->pa_free -= len; | |
3391 | ||
004379d0 | 3392 | mb_debug(1, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
3393 | } |
3394 | ||
3395 | /* | |
3396 | * use blocks preallocated to locality group | |
3397 | */ | |
3398 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3399 | struct ext4_prealloc_space *pa) | |
3400 | { | |
03cddb80 | 3401 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3402 | |
c9de560d AT |
3403 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3404 | &ac->ac_b_ex.fe_group, | |
3405 | &ac->ac_b_ex.fe_start); | |
3406 | ac->ac_b_ex.fe_len = len; | |
3407 | ac->ac_status = AC_STATUS_FOUND; | |
3408 | ac->ac_pa = pa; | |
3409 | ||
3410 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3411 | * possible race when the group is being loaded concurrently |
c9de560d | 3412 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3413 | * in on-disk bitmap -- see ext4_mb_release_context() |
3414 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3415 | */ |
6ba495e9 | 3416 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3417 | } |
3418 | ||
5e745b04 AK |
3419 | /* |
3420 | * Return the prealloc space that have minimal distance | |
3421 | * from the goal block. @cpa is the prealloc | |
3422 | * space that is having currently known minimal distance | |
3423 | * from the goal block. | |
3424 | */ | |
3425 | static struct ext4_prealloc_space * | |
3426 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3427 | struct ext4_prealloc_space *pa, | |
3428 | struct ext4_prealloc_space *cpa) | |
3429 | { | |
3430 | ext4_fsblk_t cur_distance, new_distance; | |
3431 | ||
3432 | if (cpa == NULL) { | |
3433 | atomic_inc(&pa->pa_count); | |
3434 | return pa; | |
3435 | } | |
79211c8e AM |
3436 | cur_distance = abs(goal_block - cpa->pa_pstart); |
3437 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 3438 | |
5a54b2f1 | 3439 | if (cur_distance <= new_distance) |
5e745b04 AK |
3440 | return cpa; |
3441 | ||
3442 | /* drop the previous reference */ | |
3443 | atomic_dec(&cpa->pa_count); | |
3444 | atomic_inc(&pa->pa_count); | |
3445 | return pa; | |
3446 | } | |
3447 | ||
c9de560d AT |
3448 | /* |
3449 | * search goal blocks in preallocated space | |
3450 | */ | |
4fca8f07 | 3451 | static noinline_for_stack bool |
4ddfef7b | 3452 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 3453 | { |
53accfa9 | 3454 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3455 | int order, i; |
c9de560d AT |
3456 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3457 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3458 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3459 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3460 | |
3461 | /* only data can be preallocated */ | |
3462 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 3463 | return false; |
c9de560d AT |
3464 | |
3465 | /* first, try per-file preallocation */ | |
3466 | rcu_read_lock(); | |
9a0762c5 | 3467 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3468 | |
3469 | /* all fields in this condition don't change, | |
3470 | * so we can skip locking for them */ | |
3471 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3472 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3473 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3474 | continue; |
3475 | ||
fb0a387d | 3476 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3477 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3478 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3479 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3480 | continue; |
3481 | ||
c9de560d AT |
3482 | /* found preallocated blocks, use them */ |
3483 | spin_lock(&pa->pa_lock); | |
3484 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3485 | atomic_inc(&pa->pa_count); | |
3486 | ext4_mb_use_inode_pa(ac, pa); | |
3487 | spin_unlock(&pa->pa_lock); | |
3488 | ac->ac_criteria = 10; | |
3489 | rcu_read_unlock(); | |
4fca8f07 | 3490 | return true; |
c9de560d AT |
3491 | } |
3492 | spin_unlock(&pa->pa_lock); | |
3493 | } | |
3494 | rcu_read_unlock(); | |
3495 | ||
3496 | /* can we use group allocation? */ | |
3497 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 3498 | return false; |
c9de560d AT |
3499 | |
3500 | /* inode may have no locality group for some reason */ | |
3501 | lg = ac->ac_lg; | |
3502 | if (lg == NULL) | |
4fca8f07 | 3503 | return false; |
6be2ded1 AK |
3504 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3505 | if (order > PREALLOC_TB_SIZE - 1) | |
3506 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3507 | order = PREALLOC_TB_SIZE - 1; | |
3508 | ||
bda00de7 | 3509 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3510 | /* |
3511 | * search for the prealloc space that is having | |
3512 | * minimal distance from the goal block. | |
3513 | */ | |
6be2ded1 AK |
3514 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3515 | rcu_read_lock(); | |
3516 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3517 | pa_inode_list) { | |
3518 | spin_lock(&pa->pa_lock); | |
3519 | if (pa->pa_deleted == 0 && | |
3520 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3521 | |
3522 | cpa = ext4_mb_check_group_pa(goal_block, | |
3523 | pa, cpa); | |
6be2ded1 | 3524 | } |
c9de560d | 3525 | spin_unlock(&pa->pa_lock); |
c9de560d | 3526 | } |
6be2ded1 | 3527 | rcu_read_unlock(); |
c9de560d | 3528 | } |
5e745b04 AK |
3529 | if (cpa) { |
3530 | ext4_mb_use_group_pa(ac, cpa); | |
3531 | ac->ac_criteria = 20; | |
4fca8f07 | 3532 | return true; |
5e745b04 | 3533 | } |
4fca8f07 | 3534 | return false; |
c9de560d AT |
3535 | } |
3536 | ||
7a2fcbf7 AK |
3537 | /* |
3538 | * the function goes through all block freed in the group | |
3539 | * but not yet committed and marks them used in in-core bitmap. | |
3540 | * buddy must be generated from this bitmap | |
955ce5f5 | 3541 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3542 | */ |
3543 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3544 | ext4_group_t group) | |
3545 | { | |
3546 | struct rb_node *n; | |
3547 | struct ext4_group_info *grp; | |
3548 | struct ext4_free_data *entry; | |
3549 | ||
3550 | grp = ext4_get_group_info(sb, group); | |
3551 | n = rb_first(&(grp->bb_free_root)); | |
3552 | ||
3553 | while (n) { | |
18aadd47 BJ |
3554 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3555 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3556 | n = rb_next(n); |
3557 | } | |
3558 | return; | |
3559 | } | |
3560 | ||
c9de560d AT |
3561 | /* |
3562 | * the function goes through all preallocation in this group and marks them | |
3563 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3564 | * Need to be called with ext4 group lock held |
c9de560d | 3565 | */ |
089ceecc ES |
3566 | static noinline_for_stack |
3567 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3568 | ext4_group_t group) |
3569 | { | |
3570 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3571 | struct ext4_prealloc_space *pa; | |
3572 | struct list_head *cur; | |
3573 | ext4_group_t groupnr; | |
3574 | ext4_grpblk_t start; | |
3575 | int preallocated = 0; | |
c9de560d AT |
3576 | int len; |
3577 | ||
3578 | /* all form of preallocation discards first load group, | |
3579 | * so the only competing code is preallocation use. | |
3580 | * we don't need any locking here | |
3581 | * notice we do NOT ignore preallocations with pa_deleted | |
3582 | * otherwise we could leave used blocks available for | |
3583 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3584 | * is dropping preallocation | |
3585 | */ | |
3586 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3587 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3588 | spin_lock(&pa->pa_lock); | |
3589 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3590 | &groupnr, &start); | |
3591 | len = pa->pa_len; | |
3592 | spin_unlock(&pa->pa_lock); | |
3593 | if (unlikely(len == 0)) | |
3594 | continue; | |
3595 | BUG_ON(groupnr != group); | |
c3e94d1d | 3596 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3597 | preallocated += len; |
c9de560d | 3598 | } |
004379d0 | 3599 | mb_debug(1, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
3600 | } |
3601 | ||
3602 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3603 | { | |
3604 | struct ext4_prealloc_space *pa; | |
3605 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
3606 | |
3607 | BUG_ON(atomic_read(&pa->pa_count)); | |
3608 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
3609 | kmem_cache_free(ext4_pspace_cachep, pa); |
3610 | } | |
3611 | ||
3612 | /* | |
3613 | * drops a reference to preallocated space descriptor | |
3614 | * if this was the last reference and the space is consumed | |
3615 | */ | |
3616 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3617 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3618 | { | |
a9df9a49 | 3619 | ext4_group_t grp; |
d33a1976 | 3620 | ext4_fsblk_t grp_blk; |
c9de560d | 3621 | |
c9de560d AT |
3622 | /* in this short window concurrent discard can set pa_deleted */ |
3623 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
3624 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
3625 | spin_unlock(&pa->pa_lock); | |
3626 | return; | |
3627 | } | |
3628 | ||
c9de560d AT |
3629 | if (pa->pa_deleted == 1) { |
3630 | spin_unlock(&pa->pa_lock); | |
3631 | return; | |
3632 | } | |
3633 | ||
3634 | pa->pa_deleted = 1; | |
3635 | spin_unlock(&pa->pa_lock); | |
3636 | ||
d33a1976 | 3637 | grp_blk = pa->pa_pstart; |
60e6679e | 3638 | /* |
cc0fb9ad AK |
3639 | * If doing group-based preallocation, pa_pstart may be in the |
3640 | * next group when pa is used up | |
3641 | */ | |
3642 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3643 | grp_blk--; |
3644 | ||
bd86298e | 3645 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
3646 | |
3647 | /* | |
3648 | * possible race: | |
3649 | * | |
3650 | * P1 (buddy init) P2 (regular allocation) | |
3651 | * find block B in PA | |
3652 | * copy on-disk bitmap to buddy | |
3653 | * mark B in on-disk bitmap | |
3654 | * drop PA from group | |
3655 | * mark all PAs in buddy | |
3656 | * | |
3657 | * thus, P1 initializes buddy with B available. to prevent this | |
3658 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3659 | * against that pair | |
3660 | */ | |
3661 | ext4_lock_group(sb, grp); | |
3662 | list_del(&pa->pa_group_list); | |
3663 | ext4_unlock_group(sb, grp); | |
3664 | ||
3665 | spin_lock(pa->pa_obj_lock); | |
3666 | list_del_rcu(&pa->pa_inode_list); | |
3667 | spin_unlock(pa->pa_obj_lock); | |
3668 | ||
3669 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3670 | } | |
3671 | ||
3672 | /* | |
3673 | * creates new preallocated space for given inode | |
3674 | */ | |
4ddfef7b ES |
3675 | static noinline_for_stack int |
3676 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3677 | { |
3678 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3679 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3680 | struct ext4_prealloc_space *pa; |
3681 | struct ext4_group_info *grp; | |
3682 | struct ext4_inode_info *ei; | |
3683 | ||
3684 | /* preallocate only when found space is larger then requested */ | |
3685 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3686 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3687 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3688 | ||
3689 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3690 | if (pa == NULL) | |
3691 | return -ENOMEM; | |
3692 | ||
3693 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3694 | int winl; | |
3695 | int wins; | |
3696 | int win; | |
3697 | int offs; | |
3698 | ||
3699 | /* we can't allocate as much as normalizer wants. | |
3700 | * so, found space must get proper lstart | |
3701 | * to cover original request */ | |
3702 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3703 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3704 | ||
3705 | /* we're limited by original request in that | |
3706 | * logical block must be covered any way | |
3707 | * winl is window we can move our chunk within */ | |
3708 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3709 | ||
3710 | /* also, we should cover whole original request */ | |
53accfa9 | 3711 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3712 | |
3713 | /* the smallest one defines real window */ | |
3714 | win = min(winl, wins); | |
3715 | ||
53accfa9 TT |
3716 | offs = ac->ac_o_ex.fe_logical % |
3717 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3718 | if (offs && offs < win) |
3719 | win = offs; | |
3720 | ||
53accfa9 | 3721 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3722 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3723 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3724 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3725 | } | |
3726 | ||
3727 | /* preallocation can change ac_b_ex, thus we store actually | |
3728 | * allocated blocks for history */ | |
3729 | ac->ac_f_ex = ac->ac_b_ex; | |
3730 | ||
3731 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3732 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3733 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3734 | pa->pa_free = pa->pa_len; | |
3735 | atomic_set(&pa->pa_count, 1); | |
3736 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3737 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3738 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3739 | pa->pa_deleted = 0; |
cc0fb9ad | 3740 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3741 | |
36bad423 | 3742 | mb_debug(1, "new inode pa %p: %llu/%d for %u\n", pa, |
c9de560d | 3743 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3744 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3745 | |
3746 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3747 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3748 | |
3749 | ei = EXT4_I(ac->ac_inode); | |
3750 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3751 | ||
3752 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3753 | pa->pa_inode = ac->ac_inode; | |
3754 | ||
3755 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3756 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3757 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3758 | ||
3759 | spin_lock(pa->pa_obj_lock); | |
3760 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3761 | spin_unlock(pa->pa_obj_lock); | |
3762 | ||
3763 | return 0; | |
3764 | } | |
3765 | ||
3766 | /* | |
3767 | * creates new preallocated space for locality group inodes belongs to | |
3768 | */ | |
4ddfef7b ES |
3769 | static noinline_for_stack int |
3770 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3771 | { |
3772 | struct super_block *sb = ac->ac_sb; | |
3773 | struct ext4_locality_group *lg; | |
3774 | struct ext4_prealloc_space *pa; | |
3775 | struct ext4_group_info *grp; | |
3776 | ||
3777 | /* preallocate only when found space is larger then requested */ | |
3778 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3779 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3780 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3781 | ||
3782 | BUG_ON(ext4_pspace_cachep == NULL); | |
3783 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3784 | if (pa == NULL) | |
3785 | return -ENOMEM; | |
3786 | ||
3787 | /* preallocation can change ac_b_ex, thus we store actually | |
3788 | * allocated blocks for history */ | |
3789 | ac->ac_f_ex = ac->ac_b_ex; | |
3790 | ||
3791 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3792 | pa->pa_lstart = pa->pa_pstart; | |
3793 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3794 | pa->pa_free = pa->pa_len; | |
3795 | atomic_set(&pa->pa_count, 1); | |
3796 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3797 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3798 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3799 | pa->pa_deleted = 0; |
cc0fb9ad | 3800 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3801 | |
36bad423 | 3802 | mb_debug(1, "new group pa %p: %llu/%d for %u\n", pa, |
9bffad1e TT |
3803 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3804 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3805 | |
3806 | ext4_mb_use_group_pa(ac, pa); | |
3807 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3808 | ||
3809 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3810 | lg = ac->ac_lg; | |
3811 | BUG_ON(lg == NULL); | |
3812 | ||
3813 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3814 | pa->pa_inode = NULL; | |
3815 | ||
3816 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3817 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3818 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3819 | ||
6be2ded1 AK |
3820 | /* |
3821 | * We will later add the new pa to the right bucket | |
3822 | * after updating the pa_free in ext4_mb_release_context | |
3823 | */ | |
c9de560d AT |
3824 | return 0; |
3825 | } | |
3826 | ||
3827 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3828 | { | |
3829 | int err; | |
3830 | ||
3831 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3832 | err = ext4_mb_new_group_pa(ac); | |
3833 | else | |
3834 | err = ext4_mb_new_inode_pa(ac); | |
3835 | return err; | |
3836 | } | |
3837 | ||
3838 | /* | |
3839 | * finds all unused blocks in on-disk bitmap, frees them in | |
3840 | * in-core bitmap and buddy. | |
3841 | * @pa must be unlinked from inode and group lists, so that | |
3842 | * nobody else can find/use it. | |
3843 | * the caller MUST hold group/inode locks. | |
3844 | * TODO: optimize the case when there are no in-core structures yet | |
3845 | */ | |
4ddfef7b ES |
3846 | static noinline_for_stack int |
3847 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3848 | struct ext4_prealloc_space *pa) |
c9de560d | 3849 | { |
c9de560d AT |
3850 | struct super_block *sb = e4b->bd_sb; |
3851 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3852 | unsigned int end; |
3853 | unsigned int next; | |
c9de560d AT |
3854 | ext4_group_t group; |
3855 | ext4_grpblk_t bit; | |
ba80b101 | 3856 | unsigned long long grp_blk_start; |
c9de560d AT |
3857 | int free = 0; |
3858 | ||
3859 | BUG_ON(pa->pa_deleted == 0); | |
3860 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3861 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3862 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3863 | end = bit + pa->pa_len; | |
3864 | ||
c9de560d | 3865 | while (bit < end) { |
ffad0a44 | 3866 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3867 | if (bit >= end) |
3868 | break; | |
ffad0a44 | 3869 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3870 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3871 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3872 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3873 | free += next - bit; |
3874 | ||
3e1e5f50 | 3875 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3876 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3877 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3878 | next - bit); |
c9de560d AT |
3879 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3880 | bit = next + 1; | |
3881 | } | |
3882 | if (free != pa->pa_free) { | |
9d8b9ec4 | 3883 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 3884 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
3885 | pa, (unsigned long) pa->pa_lstart, |
3886 | (unsigned long) pa->pa_pstart, | |
36bad423 | 3887 | pa->pa_len); |
e29136f8 | 3888 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3889 | free, pa->pa_free); |
e56eb659 AK |
3890 | /* |
3891 | * pa is already deleted so we use the value obtained | |
3892 | * from the bitmap and continue. | |
3893 | */ | |
c9de560d | 3894 | } |
c9de560d AT |
3895 | atomic_add(free, &sbi->s_mb_discarded); |
3896 | ||
863c37fc | 3897 | return 0; |
c9de560d AT |
3898 | } |
3899 | ||
4ddfef7b ES |
3900 | static noinline_for_stack int |
3901 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3902 | struct ext4_prealloc_space *pa) |
c9de560d | 3903 | { |
c9de560d AT |
3904 | struct super_block *sb = e4b->bd_sb; |
3905 | ext4_group_t group; | |
3906 | ext4_grpblk_t bit; | |
3907 | ||
60e07cf5 | 3908 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
3909 | BUG_ON(pa->pa_deleted == 0); |
3910 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3911 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3912 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3913 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3914 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3915 | |
3916 | return 0; | |
3917 | } | |
3918 | ||
3919 | /* | |
3920 | * releases all preallocations in given group | |
3921 | * | |
3922 | * first, we need to decide discard policy: | |
3923 | * - when do we discard | |
3924 | * 1) ENOSPC | |
3925 | * - how many do we discard | |
3926 | * 1) how many requested | |
3927 | */ | |
4ddfef7b ES |
3928 | static noinline_for_stack int |
3929 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3930 | ext4_group_t group, int needed) |
3931 | { | |
3932 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3933 | struct buffer_head *bitmap_bh = NULL; | |
3934 | struct ext4_prealloc_space *pa, *tmp; | |
3935 | struct list_head list; | |
3936 | struct ext4_buddy e4b; | |
3937 | int err; | |
3938 | int busy = 0; | |
3939 | int free = 0; | |
3940 | ||
6ba495e9 | 3941 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3942 | |
3943 | if (list_empty(&grp->bb_prealloc_list)) | |
bbc4ec77 | 3944 | goto out_dbg; |
c9de560d | 3945 | |
574ca174 | 3946 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
3947 | if (IS_ERR(bitmap_bh)) { |
3948 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
3949 | ext4_error_err(sb, -err, |
3950 | "Error %d reading block bitmap for %u", | |
3951 | err, group); | |
bbc4ec77 | 3952 | goto out_dbg; |
c9de560d AT |
3953 | } |
3954 | ||
3955 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3956 | if (err) { |
9651e6b2 KK |
3957 | ext4_warning(sb, "Error %d loading buddy information for %u", |
3958 | err, group); | |
ce89f46c | 3959 | put_bh(bitmap_bh); |
bbc4ec77 | 3960 | goto out_dbg; |
ce89f46c | 3961 | } |
c9de560d AT |
3962 | |
3963 | if (needed == 0) | |
7137d7a4 | 3964 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3965 | |
c9de560d | 3966 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3967 | repeat: |
3968 | ext4_lock_group(sb, group); | |
3969 | list_for_each_entry_safe(pa, tmp, | |
3970 | &grp->bb_prealloc_list, pa_group_list) { | |
3971 | spin_lock(&pa->pa_lock); | |
3972 | if (atomic_read(&pa->pa_count)) { | |
3973 | spin_unlock(&pa->pa_lock); | |
3974 | busy = 1; | |
3975 | continue; | |
3976 | } | |
3977 | if (pa->pa_deleted) { | |
3978 | spin_unlock(&pa->pa_lock); | |
3979 | continue; | |
3980 | } | |
3981 | ||
3982 | /* seems this one can be freed ... */ | |
3983 | pa->pa_deleted = 1; | |
3984 | ||
3985 | /* we can trust pa_free ... */ | |
3986 | free += pa->pa_free; | |
3987 | ||
3988 | spin_unlock(&pa->pa_lock); | |
3989 | ||
3990 | list_del(&pa->pa_group_list); | |
3991 | list_add(&pa->u.pa_tmp_list, &list); | |
3992 | } | |
3993 | ||
3994 | /* if we still need more blocks and some PAs were used, try again */ | |
3995 | if (free < needed && busy) { | |
3996 | busy = 0; | |
3997 | ext4_unlock_group(sb, group); | |
bb8b20ed | 3998 | cond_resched(); |
c9de560d AT |
3999 | goto repeat; |
4000 | } | |
4001 | ||
4002 | /* found anything to free? */ | |
4003 | if (list_empty(&list)) { | |
4004 | BUG_ON(free != 0); | |
bbc4ec77 RH |
4005 | mb_debug(1, "Someone else may have freed PA for this group %u\n", |
4006 | group); | |
c9de560d AT |
4007 | goto out; |
4008 | } | |
4009 | ||
4010 | /* now free all selected PAs */ | |
4011 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
4012 | ||
4013 | /* remove from object (inode or locality group) */ | |
4014 | spin_lock(pa->pa_obj_lock); | |
4015 | list_del_rcu(&pa->pa_inode_list); | |
4016 | spin_unlock(pa->pa_obj_lock); | |
4017 | ||
cc0fb9ad | 4018 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 4019 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 4020 | else |
3e1e5f50 | 4021 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4022 | |
4023 | list_del(&pa->u.pa_tmp_list); | |
4024 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4025 | } | |
4026 | ||
4027 | out: | |
4028 | ext4_unlock_group(sb, group); | |
e39e07fd | 4029 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4030 | put_bh(bitmap_bh); |
bbc4ec77 RH |
4031 | out_dbg: |
4032 | mb_debug(1, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", | |
4033 | free, group, grp->bb_free); | |
c9de560d AT |
4034 | return free; |
4035 | } | |
4036 | ||
4037 | /* | |
4038 | * releases all non-used preallocated blocks for given inode | |
4039 | * | |
4040 | * It's important to discard preallocations under i_data_sem | |
4041 | * We don't want another block to be served from the prealloc | |
4042 | * space when we are discarding the inode prealloc space. | |
4043 | * | |
4044 | * FIXME!! Make sure it is valid at all the call sites | |
4045 | */ | |
c2ea3fde | 4046 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
4047 | { |
4048 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4049 | struct super_block *sb = inode->i_sb; | |
4050 | struct buffer_head *bitmap_bh = NULL; | |
4051 | struct ext4_prealloc_space *pa, *tmp; | |
4052 | ext4_group_t group = 0; | |
4053 | struct list_head list; | |
4054 | struct ext4_buddy e4b; | |
4055 | int err; | |
4056 | ||
c2ea3fde | 4057 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4058 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4059 | return; | |
4060 | } | |
4061 | ||
6ba495e9 | 4062 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 4063 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
4064 | |
4065 | INIT_LIST_HEAD(&list); | |
4066 | ||
4067 | repeat: | |
4068 | /* first, collect all pa's in the inode */ | |
4069 | spin_lock(&ei->i_prealloc_lock); | |
4070 | while (!list_empty(&ei->i_prealloc_list)) { | |
4071 | pa = list_entry(ei->i_prealloc_list.next, | |
4072 | struct ext4_prealloc_space, pa_inode_list); | |
4073 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4074 | spin_lock(&pa->pa_lock); | |
4075 | if (atomic_read(&pa->pa_count)) { | |
4076 | /* this shouldn't happen often - nobody should | |
4077 | * use preallocation while we're discarding it */ | |
4078 | spin_unlock(&pa->pa_lock); | |
4079 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
4080 | ext4_msg(sb, KERN_ERR, |
4081 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
4082 | WARN_ON(1); |
4083 | schedule_timeout_uninterruptible(HZ); | |
4084 | goto repeat; | |
4085 | ||
4086 | } | |
4087 | if (pa->pa_deleted == 0) { | |
4088 | pa->pa_deleted = 1; | |
4089 | spin_unlock(&pa->pa_lock); | |
4090 | list_del_rcu(&pa->pa_inode_list); | |
4091 | list_add(&pa->u.pa_tmp_list, &list); | |
4092 | continue; | |
4093 | } | |
4094 | ||
4095 | /* someone is deleting pa right now */ | |
4096 | spin_unlock(&pa->pa_lock); | |
4097 | spin_unlock(&ei->i_prealloc_lock); | |
4098 | ||
4099 | /* we have to wait here because pa_deleted | |
4100 | * doesn't mean pa is already unlinked from | |
4101 | * the list. as we might be called from | |
4102 | * ->clear_inode() the inode will get freed | |
4103 | * and concurrent thread which is unlinking | |
4104 | * pa from inode's list may access already | |
4105 | * freed memory, bad-bad-bad */ | |
4106 | ||
4107 | /* XXX: if this happens too often, we can | |
4108 | * add a flag to force wait only in case | |
4109 | * of ->clear_inode(), but not in case of | |
4110 | * regular truncate */ | |
4111 | schedule_timeout_uninterruptible(HZ); | |
4112 | goto repeat; | |
4113 | } | |
4114 | spin_unlock(&ei->i_prealloc_lock); | |
4115 | ||
4116 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4117 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 4118 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 4119 | |
9651e6b2 KK |
4120 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4121 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 4122 | if (err) { |
54d3adbc TT |
4123 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
4124 | err, group); | |
ce89f46c AK |
4125 | continue; |
4126 | } | |
c9de560d | 4127 | |
574ca174 | 4128 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4129 | if (IS_ERR(bitmap_bh)) { |
4130 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4131 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
4132 | err, group); | |
e39e07fd | 4133 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 4134 | continue; |
c9de560d AT |
4135 | } |
4136 | ||
4137 | ext4_lock_group(sb, group); | |
4138 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4139 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4140 | ext4_unlock_group(sb, group); |
4141 | ||
e39e07fd | 4142 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4143 | put_bh(bitmap_bh); |
4144 | ||
4145 | list_del(&pa->u.pa_tmp_list); | |
4146 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4147 | } | |
4148 | } | |
4149 | ||
6ba495e9 | 4150 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 4151 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 4152 | { |
e68cf40c | 4153 | ext4_group_t i, ngroups; |
c9de560d | 4154 | |
a0b30c12 | 4155 | if (!ext4_mballoc_debug || |
4dd89fc6 | 4156 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) |
e3570639 ES |
4157 | return; |
4158 | ||
8df9675f | 4159 | ngroups = ext4_get_groups_count(sb); |
e68cf40c | 4160 | ext4_msg(sb, KERN_ERR, "groups: "); |
8df9675f | 4161 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
4162 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4163 | struct ext4_prealloc_space *pa; | |
4164 | ext4_grpblk_t start; | |
4165 | struct list_head *cur; | |
4166 | ext4_lock_group(sb, i); | |
4167 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4168 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4169 | pa_group_list); | |
4170 | spin_lock(&pa->pa_lock); | |
4171 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4172 | NULL, &start); | |
4173 | spin_unlock(&pa->pa_lock); | |
36bad423 | 4174 | printk(KERN_ERR "PA:%u:%d:%d \n", i, |
1c718505 | 4175 | start, pa->pa_len); |
c9de560d | 4176 | } |
60bd63d1 | 4177 | ext4_unlock_group(sb, i); |
c9de560d | 4178 | |
1c718505 | 4179 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
4180 | i, grp->bb_free, grp->bb_fragments); |
4181 | } | |
4182 | printk(KERN_ERR "\n"); | |
4183 | } | |
e68cf40c RH |
4184 | |
4185 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4186 | { | |
4187 | struct super_block *sb = ac->ac_sb; | |
4188 | ||
4189 | if (!ext4_mballoc_debug || | |
4190 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) | |
4191 | return; | |
4192 | ||
4193 | ext4_msg(sb, KERN_ERR, "Can't allocate:" | |
4194 | " Allocation context details:"); | |
004379d0 | 4195 | ext4_msg(sb, KERN_ERR, "status %u flags 0x%x", |
e68cf40c RH |
4196 | ac->ac_status, ac->ac_flags); |
4197 | ext4_msg(sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " | |
4198 | "goal %lu/%lu/%lu@%lu, " | |
4199 | "best %lu/%lu/%lu@%lu cr %d", | |
4200 | (unsigned long)ac->ac_o_ex.fe_group, | |
4201 | (unsigned long)ac->ac_o_ex.fe_start, | |
4202 | (unsigned long)ac->ac_o_ex.fe_len, | |
4203 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4204 | (unsigned long)ac->ac_g_ex.fe_group, | |
4205 | (unsigned long)ac->ac_g_ex.fe_start, | |
4206 | (unsigned long)ac->ac_g_ex.fe_len, | |
4207 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4208 | (unsigned long)ac->ac_b_ex.fe_group, | |
4209 | (unsigned long)ac->ac_b_ex.fe_start, | |
4210 | (unsigned long)ac->ac_b_ex.fe_len, | |
4211 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4212 | (int)ac->ac_criteria); | |
004379d0 | 4213 | ext4_msg(sb, KERN_ERR, "%u found", ac->ac_found); |
e68cf40c RH |
4214 | ext4_mb_show_pa(sb); |
4215 | } | |
c9de560d | 4216 | #else |
e68cf40c RH |
4217 | static inline void ext4_mb_show_pa(struct super_block *sb) |
4218 | { | |
4219 | return; | |
4220 | } | |
c9de560d AT |
4221 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
4222 | { | |
e68cf40c | 4223 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
4224 | return; |
4225 | } | |
4226 | #endif | |
4227 | ||
4228 | /* | |
4229 | * We use locality group preallocation for small size file. The size of the | |
4230 | * file is determined by the current size or the resulting size after | |
4231 | * allocation which ever is larger | |
4232 | * | |
b713a5ec | 4233 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4234 | */ |
4235 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4236 | { | |
4237 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4238 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4239 | loff_t size, isize; | |
4240 | ||
4241 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4242 | return; | |
4243 | ||
4ba74d00 TT |
4244 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4245 | return; | |
4246 | ||
53accfa9 | 4247 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
4248 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4249 | >> bsbits; | |
c9de560d | 4250 | |
82dd124c NB |
4251 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
4252 | !inode_is_open_for_write(ac->ac_inode)) { | |
50797481 TT |
4253 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; |
4254 | return; | |
4255 | } | |
4256 | ||
ebbe0277 RD |
4257 | if (sbi->s_mb_group_prealloc <= 0) { |
4258 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
4259 | return; | |
4260 | } | |
4261 | ||
c9de560d | 4262 | /* don't use group allocation for large files */ |
71780577 | 4263 | size = max(size, isize); |
cc483f10 | 4264 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 4265 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 4266 | return; |
4ba74d00 | 4267 | } |
c9de560d AT |
4268 | |
4269 | BUG_ON(ac->ac_lg != NULL); | |
4270 | /* | |
4271 | * locality group prealloc space are per cpu. The reason for having | |
4272 | * per cpu locality group is to reduce the contention between block | |
4273 | * request from multiple CPUs. | |
4274 | */ | |
a0b6bc63 | 4275 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
4276 | |
4277 | /* we're going to use group allocation */ | |
4278 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4279 | ||
4280 | /* serialize all allocations in the group */ | |
4281 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4282 | } | |
4283 | ||
4ddfef7b ES |
4284 | static noinline_for_stack int |
4285 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4286 | struct ext4_allocation_request *ar) |
4287 | { | |
4288 | struct super_block *sb = ar->inode->i_sb; | |
4289 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4290 | struct ext4_super_block *es = sbi->s_es; | |
4291 | ext4_group_t group; | |
498e5f24 TT |
4292 | unsigned int len; |
4293 | ext4_fsblk_t goal; | |
c9de560d AT |
4294 | ext4_grpblk_t block; |
4295 | ||
4296 | /* we can't allocate > group size */ | |
4297 | len = ar->len; | |
4298 | ||
4299 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
4300 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
4301 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4302 | |
4303 | /* start searching from the goal */ | |
4304 | goal = ar->goal; | |
4305 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4306 | goal >= ext4_blocks_count(es)) | |
4307 | goal = le32_to_cpu(es->s_first_data_block); | |
4308 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4309 | ||
4310 | /* set up allocation goals */ | |
f5a44db5 | 4311 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 4312 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4313 | ac->ac_sb = sb; |
4314 | ac->ac_inode = ar->inode; | |
53accfa9 | 4315 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4316 | ac->ac_o_ex.fe_group = group; |
4317 | ac->ac_o_ex.fe_start = block; | |
4318 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4319 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4320 | ac->ac_flags = ar->flags; |
c9de560d AT |
4321 | |
4322 | /* we have to define context: we'll we work with a file or | |
4323 | * locality group. this is a policy, actually */ | |
4324 | ext4_mb_group_or_file(ac); | |
4325 | ||
6ba495e9 | 4326 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4327 | "left: %u/%u, right %u/%u to %swritable\n", |
4328 | (unsigned) ar->len, (unsigned) ar->logical, | |
4329 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4330 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4331 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 4332 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
4333 | return 0; |
4334 | ||
4335 | } | |
4336 | ||
6be2ded1 AK |
4337 | static noinline_for_stack void |
4338 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4339 | struct ext4_locality_group *lg, | |
4340 | int order, int total_entries) | |
4341 | { | |
4342 | ext4_group_t group = 0; | |
4343 | struct ext4_buddy e4b; | |
4344 | struct list_head discard_list; | |
4345 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4346 | |
6ba495e9 | 4347 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4348 | |
4349 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4350 | |
4351 | spin_lock(&lg->lg_prealloc_lock); | |
4352 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
92e9c58c MB |
4353 | pa_inode_list, |
4354 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
4355 | spin_lock(&pa->pa_lock); |
4356 | if (atomic_read(&pa->pa_count)) { | |
4357 | /* | |
4358 | * This is the pa that we just used | |
4359 | * for block allocation. So don't | |
4360 | * free that | |
4361 | */ | |
4362 | spin_unlock(&pa->pa_lock); | |
4363 | continue; | |
4364 | } | |
4365 | if (pa->pa_deleted) { | |
4366 | spin_unlock(&pa->pa_lock); | |
4367 | continue; | |
4368 | } | |
4369 | /* only lg prealloc space */ | |
cc0fb9ad | 4370 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4371 | |
4372 | /* seems this one can be freed ... */ | |
4373 | pa->pa_deleted = 1; | |
4374 | spin_unlock(&pa->pa_lock); | |
4375 | ||
4376 | list_del_rcu(&pa->pa_inode_list); | |
4377 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4378 | ||
4379 | total_entries--; | |
4380 | if (total_entries <= 5) { | |
4381 | /* | |
4382 | * we want to keep only 5 entries | |
4383 | * allowing it to grow to 8. This | |
4384 | * mak sure we don't call discard | |
4385 | * soon for this list. | |
4386 | */ | |
4387 | break; | |
4388 | } | |
4389 | } | |
4390 | spin_unlock(&lg->lg_prealloc_lock); | |
4391 | ||
4392 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 4393 | int err; |
6be2ded1 | 4394 | |
bd86298e | 4395 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
4396 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4397 | GFP_NOFS|__GFP_NOFAIL); | |
4398 | if (err) { | |
54d3adbc TT |
4399 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
4400 | err, group); | |
6be2ded1 AK |
4401 | continue; |
4402 | } | |
4403 | ext4_lock_group(sb, group); | |
4404 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4405 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4406 | ext4_unlock_group(sb, group); |
4407 | ||
e39e07fd | 4408 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4409 | list_del(&pa->u.pa_tmp_list); |
4410 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4411 | } | |
6be2ded1 AK |
4412 | } |
4413 | ||
4414 | /* | |
4415 | * We have incremented pa_count. So it cannot be freed at this | |
4416 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4417 | * possible from this lg. That means pa_free cannot be updated. | |
4418 | * | |
4419 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4420 | * which can cause the lg_prealloc_list to be updated. | |
4421 | */ | |
4422 | ||
4423 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4424 | { | |
4425 | int order, added = 0, lg_prealloc_count = 1; | |
4426 | struct super_block *sb = ac->ac_sb; | |
4427 | struct ext4_locality_group *lg = ac->ac_lg; | |
4428 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4429 | ||
4430 | order = fls(pa->pa_free) - 1; | |
4431 | if (order > PREALLOC_TB_SIZE - 1) | |
4432 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4433 | order = PREALLOC_TB_SIZE - 1; | |
4434 | /* Add the prealloc space to lg */ | |
f1167009 | 4435 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 4436 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
92e9c58c MB |
4437 | pa_inode_list, |
4438 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
4439 | spin_lock(&tmp_pa->pa_lock); |
4440 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4441 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4442 | continue; |
4443 | } | |
4444 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4445 | /* Add to the tail of the previous entry */ | |
4446 | list_add_tail_rcu(&pa->pa_inode_list, | |
4447 | &tmp_pa->pa_inode_list); | |
4448 | added = 1; | |
4449 | /* | |
4450 | * we want to count the total | |
4451 | * number of entries in the list | |
4452 | */ | |
4453 | } | |
4454 | spin_unlock(&tmp_pa->pa_lock); | |
4455 | lg_prealloc_count++; | |
4456 | } | |
4457 | if (!added) | |
4458 | list_add_tail_rcu(&pa->pa_inode_list, | |
4459 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4460 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4461 | |
4462 | /* Now trim the list to be not more than 8 elements */ | |
4463 | if (lg_prealloc_count > 8) { | |
4464 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4465 | order, lg_prealloc_count); |
6be2ded1 AK |
4466 | return; |
4467 | } | |
4468 | return ; | |
4469 | } | |
4470 | ||
c9de560d AT |
4471 | /* |
4472 | * release all resource we used in allocation | |
4473 | */ | |
4474 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4475 | { | |
53accfa9 | 4476 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4477 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4478 | if (pa) { | |
cc0fb9ad | 4479 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4480 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4481 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4482 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4483 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4484 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4485 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4486 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4487 | } |
c9de560d | 4488 | } |
ba443916 AK |
4489 | if (pa) { |
4490 | /* | |
4491 | * We want to add the pa to the right bucket. | |
4492 | * Remove it from the list and while adding | |
4493 | * make sure the list to which we are adding | |
44183d42 | 4494 | * doesn't grow big. |
ba443916 | 4495 | */ |
cc0fb9ad | 4496 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4497 | spin_lock(pa->pa_obj_lock); |
4498 | list_del_rcu(&pa->pa_inode_list); | |
4499 | spin_unlock(pa->pa_obj_lock); | |
4500 | ext4_mb_add_n_trim(ac); | |
4501 | } | |
4502 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4503 | } | |
c9de560d | 4504 | if (ac->ac_bitmap_page) |
09cbfeaf | 4505 | put_page(ac->ac_bitmap_page); |
c9de560d | 4506 | if (ac->ac_buddy_page) |
09cbfeaf | 4507 | put_page(ac->ac_buddy_page); |
c9de560d AT |
4508 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
4509 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4510 | ext4_mb_collect_stats(ac); | |
4511 | return 0; | |
4512 | } | |
4513 | ||
4514 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4515 | { | |
8df9675f | 4516 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4517 | int ret; |
4518 | int freed = 0; | |
4519 | ||
9bffad1e | 4520 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4521 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4522 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4523 | freed += ret; | |
4524 | needed -= ret; | |
4525 | } | |
4526 | ||
4527 | return freed; | |
4528 | } | |
4529 | ||
4530 | /* | |
4531 | * Main entry point into mballoc to allocate blocks | |
4532 | * it tries to use preallocation first, then falls back | |
4533 | * to usual allocation | |
4534 | */ | |
4535 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4536 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4537 | { |
6bc6e63f | 4538 | int freed; |
256bdb49 | 4539 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4540 | struct ext4_sb_info *sbi; |
4541 | struct super_block *sb; | |
4542 | ext4_fsblk_t block = 0; | |
60e58e0f | 4543 | unsigned int inquota = 0; |
53accfa9 | 4544 | unsigned int reserv_clstrs = 0; |
c9de560d | 4545 | |
b10a44c3 | 4546 | might_sleep(); |
c9de560d AT |
4547 | sb = ar->inode->i_sb; |
4548 | sbi = EXT4_SB(sb); | |
4549 | ||
9bffad1e | 4550 | trace_ext4_request_blocks(ar); |
ba80b101 | 4551 | |
45dc63e7 | 4552 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 4553 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
4554 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
4555 | ||
e3cf5d5d | 4556 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
4557 | /* Without delayed allocation we need to verify |
4558 | * there is enough free blocks to do block allocation | |
4559 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4560 | */ |
55f020db | 4561 | while (ar->len && |
e7d5f315 | 4562 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4563 | |
030ba6bc | 4564 | /* let others to free the space */ |
bb8b20ed | 4565 | cond_resched(); |
030ba6bc AK |
4566 | ar->len = ar->len >> 1; |
4567 | } | |
4568 | if (!ar->len) { | |
bbc4ec77 | 4569 | ext4_mb_show_pa(sb); |
a30d542a AK |
4570 | *errp = -ENOSPC; |
4571 | return 0; | |
4572 | } | |
53accfa9 | 4573 | reserv_clstrs = ar->len; |
55f020db | 4574 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4575 | dquot_alloc_block_nofail(ar->inode, |
4576 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4577 | } else { |
4578 | while (ar->len && | |
53accfa9 TT |
4579 | dquot_alloc_block(ar->inode, |
4580 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4581 | |
4582 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4583 | ar->len--; | |
4584 | } | |
60e58e0f MC |
4585 | } |
4586 | inquota = ar->len; | |
4587 | if (ar->len == 0) { | |
4588 | *errp = -EDQUOT; | |
6c7a120a | 4589 | goto out; |
60e58e0f | 4590 | } |
07031431 | 4591 | } |
d2a17637 | 4592 | |
85556c9a | 4593 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4594 | if (!ac) { |
363d4251 | 4595 | ar->len = 0; |
256bdb49 | 4596 | *errp = -ENOMEM; |
6c7a120a | 4597 | goto out; |
256bdb49 ES |
4598 | } |
4599 | ||
256bdb49 | 4600 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4601 | if (*errp) { |
4602 | ar->len = 0; | |
6c7a120a | 4603 | goto out; |
c9de560d AT |
4604 | } |
4605 | ||
256bdb49 ES |
4606 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4607 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4608 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4609 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4610 | repeat: |
4611 | /* allocate space in core */ | |
6c7a120a | 4612 | *errp = ext4_mb_regular_allocator(ac); |
2c00ef3e AK |
4613 | if (*errp) |
4614 | goto discard_and_exit; | |
c9de560d AT |
4615 | |
4616 | /* as we've just preallocated more space than | |
2c00ef3e | 4617 | * user requested originally, we store allocated |
c9de560d | 4618 | * space in a special descriptor */ |
256bdb49 | 4619 | if (ac->ac_status == AC_STATUS_FOUND && |
2c00ef3e AK |
4620 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) |
4621 | *errp = ext4_mb_new_preallocation(ac); | |
4622 | if (*errp) { | |
4623 | discard_and_exit: | |
4624 | ext4_discard_allocated_blocks(ac); | |
4625 | goto errout; | |
4626 | } | |
c9de560d | 4627 | } |
256bdb49 | 4628 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4629 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 4630 | if (*errp) { |
b844167e | 4631 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4632 | goto errout; |
4633 | } else { | |
519deca0 AK |
4634 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4635 | ar->len = ac->ac_b_ex.fe_len; | |
4636 | } | |
c9de560d | 4637 | } else { |
256bdb49 | 4638 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4639 | if (freed) |
4640 | goto repeat; | |
4641 | *errp = -ENOSPC; | |
6c7a120a AK |
4642 | } |
4643 | ||
6d138ced | 4644 | errout: |
6c7a120a | 4645 | if (*errp) { |
256bdb49 | 4646 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4647 | ar->len = 0; |
256bdb49 | 4648 | ext4_mb_show_ac(ac); |
c9de560d | 4649 | } |
256bdb49 | 4650 | ext4_mb_release_context(ac); |
6c7a120a AK |
4651 | out: |
4652 | if (ac) | |
4653 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4654 | if (inquota && ar->len < inquota) |
53accfa9 | 4655 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4656 | if (!ar->len) { |
e3cf5d5d | 4657 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 4658 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4659 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4660 | reserv_clstrs); |
0087d9fb | 4661 | } |
c9de560d | 4662 | |
9bffad1e | 4663 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4664 | |
c9de560d AT |
4665 | return block; |
4666 | } | |
c9de560d | 4667 | |
c894058d AK |
4668 | /* |
4669 | * We can merge two free data extents only if the physical blocks | |
4670 | * are contiguous, AND the extents were freed by the same transaction, | |
4671 | * AND the blocks are associated with the same group. | |
4672 | */ | |
a0154344 DJ |
4673 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
4674 | struct ext4_free_data *entry, | |
4675 | struct ext4_free_data *new_entry, | |
4676 | struct rb_root *entry_rb_root) | |
c894058d | 4677 | { |
a0154344 DJ |
4678 | if ((entry->efd_tid != new_entry->efd_tid) || |
4679 | (entry->efd_group != new_entry->efd_group)) | |
4680 | return; | |
4681 | if (entry->efd_start_cluster + entry->efd_count == | |
4682 | new_entry->efd_start_cluster) { | |
4683 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
4684 | new_entry->efd_count += entry->efd_count; | |
4685 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
4686 | entry->efd_start_cluster) { | |
4687 | new_entry->efd_count += entry->efd_count; | |
4688 | } else | |
4689 | return; | |
4690 | spin_lock(&sbi->s_md_lock); | |
4691 | list_del(&entry->efd_list); | |
4692 | spin_unlock(&sbi->s_md_lock); | |
4693 | rb_erase(&entry->efd_node, entry_rb_root); | |
4694 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
4695 | } |
4696 | ||
4ddfef7b ES |
4697 | static noinline_for_stack int |
4698 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4699 | struct ext4_free_data *new_entry) |
c9de560d | 4700 | { |
e29136f8 | 4701 | ext4_group_t group = e4b->bd_group; |
84130193 | 4702 | ext4_grpblk_t cluster; |
d08854f5 | 4703 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 4704 | struct ext4_free_data *entry; |
c9de560d AT |
4705 | struct ext4_group_info *db = e4b->bd_info; |
4706 | struct super_block *sb = e4b->bd_sb; | |
4707 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4708 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4709 | struct rb_node *parent = NULL, *new_node; | |
4710 | ||
0390131b | 4711 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4712 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4713 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4714 | ||
18aadd47 BJ |
4715 | new_node = &new_entry->efd_node; |
4716 | cluster = new_entry->efd_start_cluster; | |
c894058d | 4717 | |
c894058d AK |
4718 | if (!*n) { |
4719 | /* first free block exent. We need to | |
4720 | protect buddy cache from being freed, | |
4721 | * otherwise we'll refresh it from | |
4722 | * on-disk bitmap and lose not-yet-available | |
4723 | * blocks */ | |
09cbfeaf KS |
4724 | get_page(e4b->bd_buddy_page); |
4725 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
4726 | } |
4727 | while (*n) { | |
4728 | parent = *n; | |
18aadd47 BJ |
4729 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
4730 | if (cluster < entry->efd_start_cluster) | |
c894058d | 4731 | n = &(*n)->rb_left; |
18aadd47 | 4732 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
4733 | n = &(*n)->rb_right; |
4734 | else { | |
e29136f8 | 4735 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
4736 | ext4_group_first_block_no(sb, group) + |
4737 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 4738 | "Block already on to-be-freed list"); |
c894058d | 4739 | return 0; |
c9de560d | 4740 | } |
c894058d | 4741 | } |
c9de560d | 4742 | |
c894058d AK |
4743 | rb_link_node(new_node, parent, n); |
4744 | rb_insert_color(new_node, &db->bb_free_root); | |
4745 | ||
4746 | /* Now try to see the extent can be merged to left and right */ | |
4747 | node = rb_prev(new_node); | |
4748 | if (node) { | |
18aadd47 | 4749 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
4750 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
4751 | &(db->bb_free_root)); | |
c894058d | 4752 | } |
c9de560d | 4753 | |
c894058d AK |
4754 | node = rb_next(new_node); |
4755 | if (node) { | |
18aadd47 | 4756 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
4757 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
4758 | &(db->bb_free_root)); | |
c9de560d | 4759 | } |
a0154344 | 4760 | |
d08854f5 | 4761 | spin_lock(&sbi->s_md_lock); |
a0154344 | 4762 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
4763 | sbi->s_mb_free_pending += clusters; |
4764 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
4765 | return 0; |
4766 | } | |
4767 | ||
44338711 TT |
4768 | /** |
4769 | * ext4_free_blocks() -- Free given blocks and update quota | |
4770 | * @handle: handle for this transaction | |
4771 | * @inode: inode | |
c60990b3 TT |
4772 | * @bh: optional buffer of the block to be freed |
4773 | * @block: starting physical block to be freed | |
4774 | * @count: number of blocks to be freed | |
5def1360 | 4775 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4776 | */ |
44338711 | 4777 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4778 | struct buffer_head *bh, ext4_fsblk_t block, |
4779 | unsigned long count, int flags) | |
c9de560d | 4780 | { |
26346ff6 | 4781 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4782 | struct super_block *sb = inode->i_sb; |
c9de560d | 4783 | struct ext4_group_desc *gdp; |
498e5f24 | 4784 | unsigned int overflow; |
c9de560d AT |
4785 | ext4_grpblk_t bit; |
4786 | struct buffer_head *gd_bh; | |
4787 | ext4_group_t block_group; | |
4788 | struct ext4_sb_info *sbi; | |
4789 | struct ext4_buddy e4b; | |
84130193 | 4790 | unsigned int count_clusters; |
c9de560d AT |
4791 | int err = 0; |
4792 | int ret; | |
4793 | ||
b10a44c3 | 4794 | might_sleep(); |
e6362609 TT |
4795 | if (bh) { |
4796 | if (block) | |
4797 | BUG_ON(block != bh->b_blocknr); | |
4798 | else | |
4799 | block = bh->b_blocknr; | |
4800 | } | |
c9de560d | 4801 | |
c9de560d | 4802 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4803 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4804 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4805 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4806 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4807 | goto error_return; |
4808 | } | |
4809 | ||
0610b6e9 | 4810 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4811 | trace_ext4_free_blocks(inode, block, count, flags); |
4812 | ||
9c02ac97 DJ |
4813 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
4814 | BUG_ON(count > 1); | |
e6362609 | 4815 | |
9c02ac97 DJ |
4816 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
4817 | inode, bh, block); | |
e6362609 TT |
4818 | } |
4819 | ||
84130193 TT |
4820 | /* |
4821 | * If the extent to be freed does not begin on a cluster | |
4822 | * boundary, we need to deal with partial clusters at the | |
4823 | * beginning and end of the extent. Normally we will free | |
4824 | * blocks at the beginning or the end unless we are explicitly | |
4825 | * requested to avoid doing so. | |
4826 | */ | |
f5a44db5 | 4827 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
4828 | if (overflow) { |
4829 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
4830 | overflow = sbi->s_cluster_ratio - overflow; | |
4831 | block += overflow; | |
4832 | if (count > overflow) | |
4833 | count -= overflow; | |
4834 | else | |
4835 | return; | |
4836 | } else { | |
4837 | block -= overflow; | |
4838 | count += overflow; | |
4839 | } | |
4840 | } | |
f5a44db5 | 4841 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
4842 | if (overflow) { |
4843 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
4844 | if (count > overflow) | |
4845 | count -= overflow; | |
4846 | else | |
4847 | return; | |
4848 | } else | |
4849 | count += sbi->s_cluster_ratio - overflow; | |
4850 | } | |
4851 | ||
9c02ac97 DJ |
4852 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
4853 | int i; | |
f96c450d | 4854 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; |
9c02ac97 DJ |
4855 | |
4856 | for (i = 0; i < count; i++) { | |
4857 | cond_resched(); | |
f96c450d DJ |
4858 | if (is_metadata) |
4859 | bh = sb_find_get_block(inode->i_sb, block + i); | |
4860 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
9c02ac97 DJ |
4861 | } |
4862 | } | |
4863 | ||
c9de560d AT |
4864 | do_more: |
4865 | overflow = 0; | |
4866 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4867 | ||
163a203d DW |
4868 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
4869 | ext4_get_group_info(sb, block_group)))) | |
4870 | return; | |
4871 | ||
c9de560d AT |
4872 | /* |
4873 | * Check to see if we are freeing blocks across a group | |
4874 | * boundary. | |
4875 | */ | |
84130193 TT |
4876 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4877 | overflow = EXT4_C2B(sbi, bit) + count - | |
4878 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
4879 | count -= overflow; |
4880 | } | |
810da240 | 4881 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 4882 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
4883 | if (IS_ERR(bitmap_bh)) { |
4884 | err = PTR_ERR(bitmap_bh); | |
4885 | bitmap_bh = NULL; | |
c9de560d | 4886 | goto error_return; |
ce89f46c | 4887 | } |
c9de560d | 4888 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4889 | if (!gdp) { |
4890 | err = -EIO; | |
c9de560d | 4891 | goto error_return; |
ce89f46c | 4892 | } |
c9de560d AT |
4893 | |
4894 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4895 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4896 | in_range(block, ext4_inode_table(sb, gdp), | |
49598e04 | 4897 | sbi->s_itb_per_group) || |
c9de560d | 4898 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
49598e04 | 4899 | sbi->s_itb_per_group)) { |
c9de560d | 4900 | |
12062ddd | 4901 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4902 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4903 | /* err = 0. ext4_std_error should be a no op */ |
4904 | goto error_return; | |
c9de560d AT |
4905 | } |
4906 | ||
4907 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4908 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4909 | if (err) | |
4910 | goto error_return; | |
4911 | ||
4912 | /* | |
4913 | * We are about to modify some metadata. Call the journal APIs | |
4914 | * to unshare ->b_data if a currently-committing transaction is | |
4915 | * using it | |
4916 | */ | |
4917 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4918 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4919 | if (err) | |
4920 | goto error_return; | |
c9de560d AT |
4921 | #ifdef AGGRESSIVE_CHECK |
4922 | { | |
4923 | int i; | |
84130193 | 4924 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
4925 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
4926 | } | |
4927 | #endif | |
84130193 | 4928 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 4929 | |
adb7ef60 KK |
4930 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
4931 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
4932 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
4933 | if (err) |
4934 | goto error_return; | |
e6362609 | 4935 | |
f96c450d DJ |
4936 | /* |
4937 | * We need to make sure we don't reuse the freed block until after the | |
4938 | * transaction is committed. We make an exception if the inode is to be | |
4939 | * written in writeback mode since writeback mode has weak data | |
4940 | * consistency guarantees. | |
4941 | */ | |
4942 | if (ext4_handle_valid(handle) && | |
4943 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
4944 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
4945 | struct ext4_free_data *new_entry; |
4946 | /* | |
7444a072 MH |
4947 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
4948 | * to fail. | |
7a2fcbf7 | 4949 | */ |
7444a072 MH |
4950 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
4951 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
4952 | new_entry->efd_start_cluster = bit; |
4953 | new_entry->efd_group = block_group; | |
4954 | new_entry->efd_count = count_clusters; | |
4955 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4956 | |
7a2fcbf7 | 4957 | ext4_lock_group(sb, block_group); |
84130193 | 4958 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 4959 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4960 | } else { |
7a2fcbf7 AK |
4961 | /* need to update group_info->bb_free and bitmap |
4962 | * with group lock held. generate_buddy look at | |
4963 | * them with group lock_held | |
4964 | */ | |
d71c1ae2 | 4965 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
4966 | err = ext4_issue_discard(sb, block_group, bit, count, |
4967 | NULL); | |
d71c1ae2 LC |
4968 | if (err && err != -EOPNOTSUPP) |
4969 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
4970 | " group:%d block:%d count:%lu failed" | |
4971 | " with %d", block_group, bit, count, | |
4972 | err); | |
8f9ff189 LC |
4973 | } else |
4974 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 4975 | |
955ce5f5 | 4976 | ext4_lock_group(sb, block_group); |
84130193 TT |
4977 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
4978 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
4979 | } |
4980 | ||
021b65bb TT |
4981 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
4982 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 4983 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 4984 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 4985 | ext4_unlock_group(sb, block_group); |
c9de560d | 4986 | |
772cb7c8 JS |
4987 | if (sbi->s_log_groups_per_flex) { |
4988 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 4989 | atomic64_add(count_clusters, |
7c990728 SJS |
4990 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
4991 | flex_group)->free_clusters); | |
772cb7c8 JS |
4992 | } |
4993 | ||
9fe67149 EW |
4994 | /* |
4995 | * on a bigalloc file system, defer the s_freeclusters_counter | |
4996 | * update to the caller (ext4_remove_space and friends) so they | |
4997 | * can determine if a cluster freed here should be rereserved | |
4998 | */ | |
4999 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
5000 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
5001 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
5002 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
5003 | count_clusters); | |
5004 | } | |
7d734532 JK |
5005 | |
5006 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 5007 | |
7a2fcbf7 AK |
5008 | /* We dirtied the bitmap block */ |
5009 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5010 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5011 | ||
c9de560d AT |
5012 | /* And the group descriptor block */ |
5013 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 5014 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
5015 | if (!err) |
5016 | err = ret; | |
5017 | ||
5018 | if (overflow && !err) { | |
5019 | block += count; | |
5020 | count = overflow; | |
5021 | put_bh(bitmap_bh); | |
5022 | goto do_more; | |
5023 | } | |
c9de560d AT |
5024 | error_return: |
5025 | brelse(bitmap_bh); | |
5026 | ext4_std_error(sb, err); | |
5027 | return; | |
5028 | } | |
7360d173 | 5029 | |
2846e820 | 5030 | /** |
0529155e | 5031 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
5032 | * @handle: handle to this transaction |
5033 | * @sb: super block | |
4907cb7b | 5034 | * @block: start physical block to add to the block group |
2846e820 AG |
5035 | * @count: number of blocks to free |
5036 | * | |
e73a347b | 5037 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 5038 | */ |
cc7365df | 5039 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
5040 | ext4_fsblk_t block, unsigned long count) |
5041 | { | |
5042 | struct buffer_head *bitmap_bh = NULL; | |
5043 | struct buffer_head *gd_bh; | |
5044 | ext4_group_t block_group; | |
5045 | ext4_grpblk_t bit; | |
5046 | unsigned int i; | |
5047 | struct ext4_group_desc *desc; | |
5048 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 5049 | struct ext4_buddy e4b; |
d77147ff | 5050 | int err = 0, ret, free_clusters_count; |
5051 | ext4_grpblk_t clusters_freed; | |
5052 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
5053 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
5054 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
5055 | |
5056 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
5057 | ||
4740b830 YY |
5058 | if (count == 0) |
5059 | return 0; | |
5060 | ||
2846e820 | 5061 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
5062 | /* |
5063 | * Check to see if we are freeing blocks across a group | |
5064 | * boundary. | |
5065 | */ | |
d77147ff | 5066 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
5067 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
5068 | block_group); |
5069 | err = -EINVAL; | |
2846e820 | 5070 | goto error_return; |
cc7365df | 5071 | } |
2cd05cc3 | 5072 | |
2846e820 | 5073 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5074 | if (IS_ERR(bitmap_bh)) { |
5075 | err = PTR_ERR(bitmap_bh); | |
5076 | bitmap_bh = NULL; | |
2846e820 | 5077 | goto error_return; |
cc7365df YY |
5078 | } |
5079 | ||
2846e820 | 5080 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
5081 | if (!desc) { |
5082 | err = -EIO; | |
2846e820 | 5083 | goto error_return; |
cc7365df | 5084 | } |
2846e820 AG |
5085 | |
5086 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
5087 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
5088 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
5089 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
5090 | sbi->s_itb_per_group)) { | |
5091 | ext4_error(sb, "Adding blocks in system zones - " | |
5092 | "Block = %llu, count = %lu", | |
5093 | block, count); | |
cc7365df | 5094 | err = -EINVAL; |
2846e820 AG |
5095 | goto error_return; |
5096 | } | |
5097 | ||
2cd05cc3 TT |
5098 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
5099 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
5100 | if (err) |
5101 | goto error_return; | |
5102 | ||
5103 | /* | |
5104 | * We are about to modify some metadata. Call the journal APIs | |
5105 | * to unshare ->b_data if a currently-committing transaction is | |
5106 | * using it | |
5107 | */ | |
5108 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
5109 | err = ext4_journal_get_write_access(handle, gd_bh); | |
5110 | if (err) | |
5111 | goto error_return; | |
e73a347b | 5112 | |
d77147ff | 5113 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 5114 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 5115 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
5116 | ext4_error(sb, "bit already cleared for block %llu", |
5117 | (ext4_fsblk_t)(block + i)); | |
5118 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
5119 | } else { | |
d77147ff | 5120 | clusters_freed++; |
2846e820 AG |
5121 | } |
5122 | } | |
e73a347b AG |
5123 | |
5124 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
5125 | if (err) | |
5126 | goto error_return; | |
5127 | ||
5128 | /* | |
5129 | * need to update group_info->bb_free and bitmap | |
5130 | * with group lock held. generate_buddy look at | |
5131 | * them with group lock_held | |
5132 | */ | |
2846e820 | 5133 | ext4_lock_group(sb, block_group); |
d77147ff | 5134 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
5135 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
5136 | free_clusters_count = clusters_freed + | |
5137 | ext4_free_group_clusters(sb, desc); | |
5138 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
79f1ba49 | 5139 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 5140 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 5141 | ext4_unlock_group(sb, block_group); |
57042651 | 5142 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 5143 | clusters_freed); |
2846e820 AG |
5144 | |
5145 | if (sbi->s_log_groups_per_flex) { | |
5146 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 5147 | atomic64_add(clusters_freed, |
7c990728 SJS |
5148 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
5149 | flex_group)->free_clusters); | |
2846e820 | 5150 | } |
e73a347b AG |
5151 | |
5152 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
5153 | |
5154 | /* We dirtied the bitmap block */ | |
5155 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5156 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5157 | ||
5158 | /* And the group descriptor block */ | |
5159 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
5160 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
5161 | if (!err) | |
5162 | err = ret; | |
5163 | ||
5164 | error_return: | |
5165 | brelse(bitmap_bh); | |
5166 | ext4_std_error(sb, err); | |
cc7365df | 5167 | return err; |
2846e820 AG |
5168 | } |
5169 | ||
7360d173 LC |
5170 | /** |
5171 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
5172 | * @sb: super block for the file system | |
5173 | * @start: starting block of the free extent in the alloc. group | |
5174 | * @count: number of blocks to TRIM | |
5175 | * @group: alloc. group we are working with | |
5176 | * @e4b: ext4 buddy for the group | |
5177 | * | |
5178 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
5179 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
5180 | * be called with under the group lock. | |
5181 | */ | |
d71c1ae2 | 5182 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
d9f34504 | 5183 | ext4_group_t group, struct ext4_buddy *e4b) |
e2cbd587 | 5184 | __releases(bitlock) |
5185 | __acquires(bitlock) | |
7360d173 LC |
5186 | { |
5187 | struct ext4_free_extent ex; | |
d71c1ae2 | 5188 | int ret = 0; |
7360d173 | 5189 | |
b3d4c2b1 TM |
5190 | trace_ext4_trim_extent(sb, group, start, count); |
5191 | ||
7360d173 LC |
5192 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
5193 | ||
5194 | ex.fe_start = start; | |
5195 | ex.fe_group = group; | |
5196 | ex.fe_len = count; | |
5197 | ||
5198 | /* | |
5199 | * Mark blocks used, so no one can reuse them while | |
5200 | * being trimmed. | |
5201 | */ | |
5202 | mb_mark_used(e4b, &ex); | |
5203 | ext4_unlock_group(sb, group); | |
a0154344 | 5204 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
5205 | ext4_lock_group(sb, group); |
5206 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 5207 | return ret; |
7360d173 LC |
5208 | } |
5209 | ||
5210 | /** | |
5211 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
5212 | * @sb: super block for file system | |
22612283 | 5213 | * @group: group to be trimmed |
7360d173 LC |
5214 | * @start: first group block to examine |
5215 | * @max: last group block to examine | |
5216 | * @minblocks: minimum extent block count | |
5217 | * | |
5218 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
5219 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
5220 | * the extent. | |
5221 | * | |
5222 | * | |
5223 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
5224 | * extents. When the free extent is found, mark it as used in group buddy | |
5225 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
5226 | * the group buddy bitmap. This is done until whole group is scanned. | |
5227 | */ | |
0b75a840 | 5228 | static ext4_grpblk_t |
78944086 LC |
5229 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
5230 | ext4_grpblk_t start, ext4_grpblk_t max, | |
5231 | ext4_grpblk_t minblocks) | |
7360d173 LC |
5232 | { |
5233 | void *bitmap; | |
169ddc3e | 5234 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 5235 | struct ext4_buddy e4b; |
d71c1ae2 | 5236 | int ret = 0; |
7360d173 | 5237 | |
b3d4c2b1 TM |
5238 | trace_ext4_trim_all_free(sb, group, start, max); |
5239 | ||
78944086 LC |
5240 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
5241 | if (ret) { | |
9651e6b2 KK |
5242 | ext4_warning(sb, "Error %d loading buddy information for %u", |
5243 | ret, group); | |
78944086 LC |
5244 | return ret; |
5245 | } | |
78944086 | 5246 | bitmap = e4b.bd_bitmap; |
28739eea LC |
5247 | |
5248 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
5249 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
5250 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
5251 | goto out; | |
5252 | ||
78944086 LC |
5253 | start = (e4b.bd_info->bb_first_free > start) ? |
5254 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 5255 | |
913eed83 LC |
5256 | while (start <= max) { |
5257 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
5258 | if (start > max) | |
7360d173 | 5259 | break; |
913eed83 | 5260 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
5261 | |
5262 | if ((next - start) >= minblocks) { | |
d71c1ae2 LC |
5263 | ret = ext4_trim_extent(sb, start, |
5264 | next - start, group, &e4b); | |
5265 | if (ret && ret != -EOPNOTSUPP) | |
5266 | break; | |
5267 | ret = 0; | |
7360d173 LC |
5268 | count += next - start; |
5269 | } | |
169ddc3e | 5270 | free_count += next - start; |
7360d173 LC |
5271 | start = next + 1; |
5272 | ||
5273 | if (fatal_signal_pending(current)) { | |
5274 | count = -ERESTARTSYS; | |
5275 | break; | |
5276 | } | |
5277 | ||
5278 | if (need_resched()) { | |
5279 | ext4_unlock_group(sb, group); | |
5280 | cond_resched(); | |
5281 | ext4_lock_group(sb, group); | |
5282 | } | |
5283 | ||
169ddc3e | 5284 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
5285 | break; |
5286 | } | |
3d56b8d2 | 5287 | |
d71c1ae2 LC |
5288 | if (!ret) { |
5289 | ret = count; | |
3d56b8d2 | 5290 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 5291 | } |
3d56b8d2 | 5292 | out: |
7360d173 | 5293 | ext4_unlock_group(sb, group); |
78944086 | 5294 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
5295 | |
5296 | ext4_debug("trimmed %d blocks in the group %d\n", | |
5297 | count, group); | |
5298 | ||
d71c1ae2 | 5299 | return ret; |
7360d173 LC |
5300 | } |
5301 | ||
5302 | /** | |
5303 | * ext4_trim_fs() -- trim ioctl handle function | |
5304 | * @sb: superblock for filesystem | |
5305 | * @range: fstrim_range structure | |
5306 | * | |
5307 | * start: First Byte to trim | |
5308 | * len: number of Bytes to trim from start | |
5309 | * minlen: minimum extent length in Bytes | |
5310 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
5311 | * start to start+len. For each such a group ext4_trim_all_free function | |
5312 | * is invoked to trim all free space. | |
5313 | */ | |
5314 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
5315 | { | |
78944086 | 5316 | struct ext4_group_info *grp; |
913eed83 | 5317 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 5318 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 5319 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
5320 | ext4_fsblk_t first_data_blk = |
5321 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 5322 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
5323 | int ret = 0; |
5324 | ||
5325 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 5326 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
5327 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
5328 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 5329 | |
5de35e8d LC |
5330 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
5331 | start >= max_blks || | |
5332 | range->len < sb->s_blocksize) | |
7360d173 | 5333 | return -EINVAL; |
913eed83 LC |
5334 | if (end >= max_blks) |
5335 | end = max_blks - 1; | |
5336 | if (end <= first_data_blk) | |
22f10457 | 5337 | goto out; |
913eed83 | 5338 | if (start < first_data_blk) |
0f0a25bf | 5339 | start = first_data_blk; |
7360d173 | 5340 | |
913eed83 | 5341 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5342 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5343 | &first_group, &first_cluster); |
913eed83 | 5344 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5345 | &last_group, &last_cluster); |
7360d173 | 5346 | |
913eed83 LC |
5347 | /* end now represents the last cluster to discard in this group */ |
5348 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5349 | |
5350 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5351 | grp = ext4_get_group_info(sb, group); |
5352 | /* We only do this if the grp has never been initialized */ | |
5353 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 5354 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
5355 | if (ret) |
5356 | break; | |
7360d173 LC |
5357 | } |
5358 | ||
0ba08517 | 5359 | /* |
913eed83 LC |
5360 | * For all the groups except the last one, last cluster will |
5361 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5362 | * change it for the last group, note that last_cluster is | |
5363 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5364 | */ |
913eed83 LC |
5365 | if (group == last_group) |
5366 | end = last_cluster; | |
7360d173 | 5367 | |
78944086 | 5368 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5369 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 5370 | end, minlen); |
7360d173 LC |
5371 | if (cnt < 0) { |
5372 | ret = cnt; | |
7360d173 LC |
5373 | break; |
5374 | } | |
21e7fd22 | 5375 | trimmed += cnt; |
7360d173 | 5376 | } |
913eed83 LC |
5377 | |
5378 | /* | |
5379 | * For every group except the first one, we are sure | |
5380 | * that the first cluster to discard will be cluster #0. | |
5381 | */ | |
7137d7a4 | 5382 | first_cluster = 0; |
7360d173 | 5383 | } |
7360d173 | 5384 | |
3d56b8d2 TM |
5385 | if (!ret) |
5386 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5387 | ||
22f10457 | 5388 | out: |
aaf7d73e | 5389 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5390 | return ret; |
5391 | } | |
0c9ec4be DW |
5392 | |
5393 | /* Iterate all the free extents in the group. */ | |
5394 | int | |
5395 | ext4_mballoc_query_range( | |
5396 | struct super_block *sb, | |
5397 | ext4_group_t group, | |
5398 | ext4_grpblk_t start, | |
5399 | ext4_grpblk_t end, | |
5400 | ext4_mballoc_query_range_fn formatter, | |
5401 | void *priv) | |
5402 | { | |
5403 | void *bitmap; | |
5404 | ext4_grpblk_t next; | |
5405 | struct ext4_buddy e4b; | |
5406 | int error; | |
5407 | ||
5408 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
5409 | if (error) | |
5410 | return error; | |
5411 | bitmap = e4b.bd_bitmap; | |
5412 | ||
5413 | ext4_lock_group(sb, group); | |
5414 | ||
5415 | start = (e4b.bd_info->bb_first_free > start) ? | |
5416 | e4b.bd_info->bb_first_free : start; | |
5417 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
5418 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
5419 | ||
5420 | while (start <= end) { | |
5421 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
5422 | if (start > end) | |
5423 | break; | |
5424 | next = mb_find_next_bit(bitmap, end + 1, start); | |
5425 | ||
5426 | ext4_unlock_group(sb, group); | |
5427 | error = formatter(sb, group, start, next - start, priv); | |
5428 | if (error) | |
5429 | goto out_unload; | |
5430 | ext4_lock_group(sb, group); | |
5431 | ||
5432 | start = next + 1; | |
5433 | } | |
5434 | ||
5435 | ext4_unlock_group(sb, group); | |
5436 | out_unload: | |
5437 | ext4_mb_unload_buddy(&e4b); | |
5438 | ||
5439 | return error; | |
5440 | } |