]>
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 | ||
c9de560d AT |
21 | /* |
22 | * MUSTDO: | |
23 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
24 | * - search for metadata in few groups | |
25 | * | |
26 | * TODO v4: | |
27 | * - normalization should take into account whether file is still open | |
28 | * - discard preallocations if no free space left (policy?) | |
29 | * - don't normalize tails | |
30 | * - quota | |
31 | * - reservation for superuser | |
32 | * | |
33 | * TODO v3: | |
34 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
35 | * - track min/max extents in each group for better group selection | |
36 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
37 | * - tree of groups sorted by number of free blocks | |
38 | * - error handling | |
39 | */ | |
40 | ||
41 | /* | |
42 | * The allocation request involve request for multiple number of blocks | |
43 | * near to the goal(block) value specified. | |
44 | * | |
b713a5ec TT |
45 | * During initialization phase of the allocator we decide to use the |
46 | * group preallocation or inode preallocation depending on the size of | |
47 | * the file. The size of the file could be the resulting file size we | |
48 | * would have after allocation, or the current file size, which ever | |
49 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
50 | * select to use the group preallocation. The default value of | |
51 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
52 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
53 | * terms of number of blocks. | |
c9de560d AT |
54 | * |
55 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 56 | * ensure that we have small files closer together on the disk. |
c9de560d | 57 | * |
b713a5ec TT |
58 | * First stage the allocator looks at the inode prealloc list, |
59 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
60 | * spaces for this particular inode. The inode prealloc space is | |
61 | * represented as: | |
c9de560d AT |
62 | * |
63 | * pa_lstart -> the logical start block for this prealloc space | |
64 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
65 | * pa_len -> length for this prealloc space (in clusters) |
66 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
67 | * |
68 | * The inode preallocation space is used looking at the _logical_ start | |
69 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
70 | * space we will consume the particular prealloc space. This makes sure that |
71 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
72 | * |
73 | * The important thing to be noted in case of inode prealloc space is that | |
74 | * we don't modify the values associated to inode prealloc space except | |
75 | * pa_free. | |
76 | * | |
77 | * If we are not able to find blocks in the inode prealloc space and if we | |
78 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 79 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
80 | * |
81 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
82 | * | |
83 | * The reason for having a per cpu locality group is to reduce the contention | |
84 | * between CPUs. It is possible to get scheduled at this point. | |
85 | * | |
86 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 87 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
88 | * |
89 | * If we can't allocate blocks via inode prealloc or/and locality group | |
90 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
91 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
92 | * mapped to the buddy and bitmap information regarding different | |
93 | * groups. The buddy information is attached to buddy cache inode so that | |
94 | * we can access them through the page cache. The information regarding | |
95 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
96 | * block bitmap and buddy information. The information are stored in the | |
97 | * inode as: | |
98 | * | |
99 | * { page } | |
c3a326a6 | 100 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
101 | * |
102 | * | |
103 | * one block each for bitmap and buddy information. So for each group we | |
ea1754a0 | 104 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / |
c9de560d AT |
105 | * blocksize) blocks. So it can have information regarding groups_per_page |
106 | * which is blocks_per_page/2 | |
107 | * | |
108 | * The buddy cache inode is not stored on disk. The inode is thrown | |
109 | * away when the filesystem is unmounted. | |
110 | * | |
111 | * We look for count number of blocks in the buddy cache. If we were able | |
112 | * to locate that many free blocks we return with additional information | |
113 | * regarding rest of the contiguous physical block available | |
114 | * | |
115 | * Before allocating blocks via buddy cache we normalize the request | |
116 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
117 | * blocks that we get after allocation is added to the respective prealloc | |
118 | * list. In case of inode preallocation we follow a list of heuristics | |
119 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
120 | * we are doing a group prealloc we try to normalize the request to | |
27baebb8 TT |
121 | * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is |
122 | * dependent on the cluster size; for non-bigalloc file systems, it is | |
c9de560d | 123 | * 512 blocks. This can be tuned via |
d7a1fee1 | 124 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
125 | * terms of number of blocks. If we have mounted the file system with -O |
126 | * stripe=<value> option the group prealloc request is normalized to the | |
b483bb77 | 127 | * smallest multiple of the stripe value (sbi->s_stripe) which is |
d7a1fee1 | 128 | * greater than the default mb_group_prealloc. |
c9de560d | 129 | * |
d7a1fee1 | 130 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 131 | * |
b713a5ec TT |
132 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
133 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
134 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 135 | * |
b713a5ec | 136 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
137 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
138 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 139 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 140 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
141 | * stripe size. This should result in better allocation on RAID setups. If |
142 | * not, we search in the specific group using bitmap for best extents. The | |
143 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 144 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 145 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
146 | * best extent in the found extents. Searching for the blocks starts with |
147 | * the group specified as the goal value in allocation context via | |
148 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 149 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
150 | * checked. |
151 | * | |
152 | * Both the prealloc space are getting populated as above. So for the first | |
153 | * request we will hit the buddy cache which will result in this prealloc | |
154 | * space getting filled. The prealloc space is then later used for the | |
155 | * subsequent request. | |
156 | */ | |
157 | ||
158 | /* | |
159 | * mballoc operates on the following data: | |
160 | * - on-disk bitmap | |
161 | * - in-core buddy (actually includes buddy and bitmap) | |
162 | * - preallocation descriptors (PAs) | |
163 | * | |
164 | * there are two types of preallocations: | |
165 | * - inode | |
166 | * assiged to specific inode and can be used for this inode only. | |
167 | * it describes part of inode's space preallocated to specific | |
168 | * physical blocks. any block from that preallocated can be used | |
169 | * independent. the descriptor just tracks number of blocks left | |
170 | * unused. so, before taking some block from descriptor, one must | |
171 | * make sure corresponded logical block isn't allocated yet. this | |
172 | * also means that freeing any block within descriptor's range | |
173 | * must discard all preallocated blocks. | |
174 | * - locality group | |
175 | * assigned to specific locality group which does not translate to | |
176 | * permanent set of inodes: inode can join and leave group. space | |
177 | * from this type of preallocation can be used for any inode. thus | |
178 | * it's consumed from the beginning to the end. | |
179 | * | |
180 | * relation between them can be expressed as: | |
181 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
182 | * | |
183 | * this mean blocks mballoc considers used are: | |
184 | * - allocated blocks (persistent) | |
185 | * - preallocated blocks (non-persistent) | |
186 | * | |
187 | * consistency in mballoc world means that at any time a block is either | |
188 | * free or used in ALL structures. notice: "any time" should not be read | |
189 | * literally -- time is discrete and delimited by locks. | |
190 | * | |
191 | * to keep it simple, we don't use block numbers, instead we count number of | |
192 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
193 | * | |
194 | * all operations can be expressed as: | |
195 | * - init buddy: buddy = on-disk + PAs | |
196 | * - new PA: buddy += N; PA = N | |
197 | * - use inode PA: on-disk += N; PA -= N | |
198 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
199 | * - use locality group PA on-disk += N; PA -= N | |
200 | * - discard locality group PA buddy -= PA; PA = 0 | |
201 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
202 | * is used in real operation because we can't know actual used | |
203 | * bits from PA, only from on-disk bitmap | |
204 | * | |
205 | * if we follow this strict logic, then all operations above should be atomic. | |
206 | * given some of them can block, we'd have to use something like semaphores | |
207 | * killing performance on high-end SMP hardware. let's try to relax it using | |
208 | * the following knowledge: | |
209 | * 1) if buddy is referenced, it's already initialized | |
210 | * 2) while block is used in buddy and the buddy is referenced, | |
211 | * nobody can re-allocate that block | |
212 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
213 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
214 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
215 | * block | |
216 | * | |
217 | * so, now we're building a concurrency table: | |
218 | * - init buddy vs. | |
219 | * - new PA | |
220 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
221 | * until PA is linked to allocation group to avoid concurrent buddy init | |
222 | * - use inode PA | |
223 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
224 | * given (3) we care that PA-=N operation doesn't interfere with init | |
225 | * - discard inode PA | |
226 | * the simplest way would be to have buddy initialized by the discard | |
227 | * - use locality group PA | |
228 | * again PA-=N must be serialized with init | |
229 | * - discard locality group PA | |
230 | * the simplest way would be to have buddy initialized by the discard | |
231 | * - new PA vs. | |
232 | * - use inode PA | |
233 | * i_data_sem serializes them | |
234 | * - discard inode PA | |
235 | * discard process must wait until PA isn't used by another process | |
236 | * - use locality group PA | |
237 | * some mutex should serialize them | |
238 | * - discard locality group PA | |
239 | * discard process must wait until PA isn't used by another process | |
240 | * - use inode PA | |
241 | * - use inode PA | |
242 | * i_data_sem or another mutex should serializes them | |
243 | * - discard inode PA | |
244 | * discard process must wait until PA isn't used by another process | |
245 | * - use locality group PA | |
246 | * nothing wrong here -- they're different PAs covering different blocks | |
247 | * - discard locality group PA | |
248 | * discard process must wait until PA isn't used by another process | |
249 | * | |
250 | * now we're ready to make few consequences: | |
251 | * - PA is referenced and while it is no discard is possible | |
252 | * - PA is referenced until block isn't marked in on-disk bitmap | |
253 | * - PA changes only after on-disk bitmap | |
254 | * - discard must not compete with init. either init is done before | |
255 | * any discard or they're serialized somehow | |
256 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
257 | * | |
258 | * a special case when we've used PA to emptiness. no need to modify buddy | |
259 | * in this case, but we should care about concurrent init | |
260 | * | |
261 | */ | |
262 | ||
263 | /* | |
264 | * Logic in few words: | |
265 | * | |
266 | * - allocation: | |
267 | * load group | |
268 | * find blocks | |
269 | * mark bits in on-disk bitmap | |
270 | * release group | |
271 | * | |
272 | * - use preallocation: | |
273 | * find proper PA (per-inode or group) | |
274 | * load group | |
275 | * mark bits in on-disk bitmap | |
276 | * release group | |
277 | * release PA | |
278 | * | |
279 | * - free: | |
280 | * load group | |
281 | * mark bits in on-disk bitmap | |
282 | * release group | |
283 | * | |
284 | * - discard preallocations in group: | |
285 | * mark PAs deleted | |
286 | * move them onto local list | |
287 | * load on-disk bitmap | |
288 | * load group | |
289 | * remove PA from object (inode or locality group) | |
290 | * mark free blocks in-core | |
291 | * | |
292 | * - discard inode's preallocations: | |
293 | */ | |
294 | ||
295 | /* | |
296 | * Locking rules | |
297 | * | |
298 | * Locks: | |
299 | * - bitlock on a group (group) | |
300 | * - object (inode/locality) (object) | |
301 | * - per-pa lock (pa) | |
302 | * | |
303 | * Paths: | |
304 | * - new pa | |
305 | * object | |
306 | * group | |
307 | * | |
308 | * - find and use pa: | |
309 | * pa | |
310 | * | |
311 | * - release consumed pa: | |
312 | * pa | |
313 | * group | |
314 | * object | |
315 | * | |
316 | * - generate in-core bitmap: | |
317 | * group | |
318 | * pa | |
319 | * | |
320 | * - discard all for given object (inode, locality group): | |
321 | * object | |
322 | * pa | |
323 | * group | |
324 | * | |
325 | * - discard all for given group: | |
326 | * group | |
327 | * pa | |
328 | * group | |
329 | * object | |
330 | * | |
331 | */ | |
c3a326a6 AK |
332 | static struct kmem_cache *ext4_pspace_cachep; |
333 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 334 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
335 | |
336 | /* We create slab caches for groupinfo data structures based on the | |
337 | * superblock block size. There will be one per mounted filesystem for | |
338 | * each unique s_blocksize_bits */ | |
2892c15d | 339 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
340 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
341 | ||
d6006186 | 342 | static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
2892c15d ES |
343 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", |
344 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
345 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
346 | }; | |
347 | ||
c3a326a6 AK |
348 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
349 | ext4_group_t group); | |
7a2fcbf7 AK |
350 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
351 | ext4_group_t group); | |
53f86b17 | 352 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); |
c3a326a6 | 353 | |
07b5b8e1 RH |
354 | /* |
355 | * The algorithm using this percpu seq counter goes below: | |
356 | * 1. We sample the percpu discard_pa_seq counter before trying for block | |
357 | * allocation in ext4_mb_new_blocks(). | |
358 | * 2. We increment this percpu discard_pa_seq counter when we either allocate | |
359 | * or free these blocks i.e. while marking those blocks as used/free in | |
360 | * mb_mark_used()/mb_free_blocks(). | |
361 | * 3. We also increment this percpu seq counter when we successfully identify | |
362 | * that the bb_prealloc_list is not empty and hence proceed for discarding | |
363 | * of those PAs inside ext4_mb_discard_group_preallocations(). | |
364 | * | |
365 | * Now to make sure that the regular fast path of block allocation is not | |
366 | * affected, as a small optimization we only sample the percpu seq counter | |
367 | * on that cpu. Only when the block allocation fails and when freed blocks | |
368 | * found were 0, that is when we sample percpu seq counter for all cpus using | |
369 | * below function ext4_get_discard_pa_seq_sum(). This happens after making | |
370 | * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. | |
371 | */ | |
372 | static DEFINE_PER_CPU(u64, discard_pa_seq); | |
373 | static inline u64 ext4_get_discard_pa_seq_sum(void) | |
374 | { | |
375 | int __cpu; | |
376 | u64 __seq = 0; | |
377 | ||
378 | for_each_possible_cpu(__cpu) | |
379 | __seq += per_cpu(discard_pa_seq, __cpu); | |
380 | return __seq; | |
381 | } | |
382 | ||
ffad0a44 AK |
383 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
384 | { | |
c9de560d | 385 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
386 | *bit += ((unsigned long) addr & 7UL) << 3; |
387 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 388 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
389 | *bit += ((unsigned long) addr & 3UL) << 3; |
390 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
391 | #else |
392 | #error "how many bits you are?!" | |
393 | #endif | |
ffad0a44 AK |
394 | return addr; |
395 | } | |
c9de560d AT |
396 | |
397 | static inline int mb_test_bit(int bit, void *addr) | |
398 | { | |
399 | /* | |
400 | * ext4_test_bit on architecture like powerpc | |
401 | * needs unsigned long aligned address | |
402 | */ | |
ffad0a44 | 403 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
404 | return ext4_test_bit(bit, addr); |
405 | } | |
406 | ||
407 | static inline void mb_set_bit(int bit, void *addr) | |
408 | { | |
ffad0a44 | 409 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
410 | ext4_set_bit(bit, addr); |
411 | } | |
412 | ||
c9de560d AT |
413 | static inline void mb_clear_bit(int bit, void *addr) |
414 | { | |
ffad0a44 | 415 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
416 | ext4_clear_bit(bit, addr); |
417 | } | |
418 | ||
eabe0444 AS |
419 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
420 | { | |
421 | addr = mb_correct_addr_and_bit(&bit, addr); | |
422 | return ext4_test_and_clear_bit(bit, addr); | |
423 | } | |
424 | ||
ffad0a44 AK |
425 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
426 | { | |
e7dfb246 | 427 | int fix = 0, ret, tmpmax; |
ffad0a44 | 428 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 429 | tmpmax = max + fix; |
ffad0a44 AK |
430 | start += fix; |
431 | ||
e7dfb246 AK |
432 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
433 | if (ret > max) | |
434 | return max; | |
435 | return ret; | |
ffad0a44 AK |
436 | } |
437 | ||
438 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
439 | { | |
e7dfb246 | 440 | int fix = 0, ret, tmpmax; |
ffad0a44 | 441 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 442 | tmpmax = max + fix; |
ffad0a44 AK |
443 | start += fix; |
444 | ||
e7dfb246 AK |
445 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
446 | if (ret > max) | |
447 | return max; | |
448 | return ret; | |
ffad0a44 AK |
449 | } |
450 | ||
c9de560d AT |
451 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
452 | { | |
453 | char *bb; | |
454 | ||
c5e8f3f3 | 455 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
456 | BUG_ON(max == NULL); |
457 | ||
458 | if (order > e4b->bd_blkbits + 1) { | |
459 | *max = 0; | |
460 | return NULL; | |
461 | } | |
462 | ||
463 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
464 | if (order == 0) { |
465 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 466 | return e4b->bd_bitmap; |
84b775a3 | 467 | } |
c9de560d | 468 | |
c5e8f3f3 | 469 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
470 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
471 | ||
472 | return bb; | |
473 | } | |
474 | ||
475 | #ifdef DOUBLE_CHECK | |
476 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
477 | int first, int count) | |
478 | { | |
479 | int i; | |
480 | struct super_block *sb = e4b->bd_sb; | |
481 | ||
482 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
483 | return; | |
bc8e6740 | 484 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
485 | for (i = 0; i < count; i++) { |
486 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
487 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
488 | |
489 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 490 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 491 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
492 | inode ? inode->i_ino : 0, |
493 | blocknr, | |
494 | "freeing block already freed " | |
495 | "(bit %u)", | |
496 | first + i); | |
736dedbb WS |
497 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
498 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
499 | } |
500 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
501 | } | |
502 | } | |
503 | ||
504 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
505 | { | |
506 | int i; | |
507 | ||
508 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
509 | return; | |
bc8e6740 | 510 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
511 | for (i = 0; i < count; i++) { |
512 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
513 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
514 | } | |
515 | } | |
516 | ||
517 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
518 | { | |
eb2b8ebb RH |
519 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
520 | return; | |
c9de560d AT |
521 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { |
522 | unsigned char *b1, *b2; | |
523 | int i; | |
524 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
525 | b2 = (unsigned char *) bitmap; | |
526 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
527 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
528 | ext4_msg(e4b->bd_sb, KERN_ERR, |
529 | "corruption in group %u " | |
530 | "at byte %u(%u): %x in copy != %x " | |
531 | "on disk/prealloc", | |
532 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
533 | BUG(); |
534 | } | |
535 | } | |
536 | } | |
537 | } | |
538 | ||
a3450215 RH |
539 | static void mb_group_bb_bitmap_alloc(struct super_block *sb, |
540 | struct ext4_group_info *grp, ext4_group_t group) | |
541 | { | |
542 | struct buffer_head *bh; | |
543 | ||
544 | grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); | |
eb2b8ebb RH |
545 | if (!grp->bb_bitmap) |
546 | return; | |
a3450215 RH |
547 | |
548 | bh = ext4_read_block_bitmap(sb, group); | |
eb2b8ebb RH |
549 | if (IS_ERR_OR_NULL(bh)) { |
550 | kfree(grp->bb_bitmap); | |
551 | grp->bb_bitmap = NULL; | |
552 | return; | |
553 | } | |
a3450215 RH |
554 | |
555 | memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); | |
556 | put_bh(bh); | |
557 | } | |
558 | ||
559 | static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
560 | { | |
561 | kfree(grp->bb_bitmap); | |
562 | } | |
563 | ||
c9de560d AT |
564 | #else |
565 | static inline void mb_free_blocks_double(struct inode *inode, | |
566 | struct ext4_buddy *e4b, int first, int count) | |
567 | { | |
568 | return; | |
569 | } | |
570 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
571 | int first, int count) | |
572 | { | |
573 | return; | |
574 | } | |
575 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
576 | { | |
577 | return; | |
578 | } | |
a3450215 RH |
579 | |
580 | static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, | |
581 | struct ext4_group_info *grp, ext4_group_t group) | |
582 | { | |
583 | return; | |
584 | } | |
585 | ||
586 | static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
587 | { | |
588 | return; | |
589 | } | |
c9de560d AT |
590 | #endif |
591 | ||
592 | #ifdef AGGRESSIVE_CHECK | |
593 | ||
594 | #define MB_CHECK_ASSERT(assert) \ | |
595 | do { \ | |
596 | if (!(assert)) { \ | |
597 | printk(KERN_EMERG \ | |
598 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
599 | function, file, line, # assert); \ | |
600 | BUG(); \ | |
601 | } \ | |
602 | } while (0) | |
603 | ||
604 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
605 | const char *function, int line) | |
606 | { | |
607 | struct super_block *sb = e4b->bd_sb; | |
608 | int order = e4b->bd_blkbits + 1; | |
609 | int max; | |
610 | int max2; | |
611 | int i; | |
612 | int j; | |
613 | int k; | |
614 | int count; | |
615 | struct ext4_group_info *grp; | |
616 | int fragments = 0; | |
617 | int fstart; | |
618 | struct list_head *cur; | |
619 | void *buddy; | |
620 | void *buddy2; | |
621 | ||
c9de560d AT |
622 | { |
623 | static int mb_check_counter; | |
624 | if (mb_check_counter++ % 100 != 0) | |
625 | return 0; | |
626 | } | |
627 | ||
628 | while (order > 1) { | |
629 | buddy = mb_find_buddy(e4b, order, &max); | |
630 | MB_CHECK_ASSERT(buddy); | |
631 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
632 | MB_CHECK_ASSERT(buddy2); | |
633 | MB_CHECK_ASSERT(buddy != buddy2); | |
634 | MB_CHECK_ASSERT(max * 2 == max2); | |
635 | ||
636 | count = 0; | |
637 | for (i = 0; i < max; i++) { | |
638 | ||
639 | if (mb_test_bit(i, buddy)) { | |
640 | /* only single bit in buddy2 may be 1 */ | |
641 | if (!mb_test_bit(i << 1, buddy2)) { | |
642 | MB_CHECK_ASSERT( | |
643 | mb_test_bit((i<<1)+1, buddy2)); | |
644 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
645 | MB_CHECK_ASSERT( | |
646 | mb_test_bit(i << 1, buddy2)); | |
647 | } | |
648 | continue; | |
649 | } | |
650 | ||
0a10da73 | 651 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
652 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
653 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
654 | ||
655 | for (j = 0; j < (1 << order); j++) { | |
656 | k = (i * (1 << order)) + j; | |
657 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 658 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
659 | } |
660 | count++; | |
661 | } | |
662 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
663 | order--; | |
664 | } | |
665 | ||
666 | fstart = -1; | |
667 | buddy = mb_find_buddy(e4b, 0, &max); | |
668 | for (i = 0; i < max; i++) { | |
669 | if (!mb_test_bit(i, buddy)) { | |
670 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
671 | if (fstart == -1) { | |
672 | fragments++; | |
673 | fstart = i; | |
674 | } | |
675 | continue; | |
676 | } | |
677 | fstart = -1; | |
678 | /* check used bits only */ | |
679 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
680 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
681 | k = i >> j; | |
682 | MB_CHECK_ASSERT(k < max2); | |
683 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
684 | } | |
685 | } | |
686 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
687 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
688 | ||
689 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
690 | list_for_each(cur, &grp->bb_prealloc_list) { |
691 | ext4_group_t groupnr; | |
692 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
693 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
694 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 695 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 696 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
697 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
698 | } | |
699 | return 0; | |
700 | } | |
701 | #undef MB_CHECK_ASSERT | |
702 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 703 | __FILE__, __func__, __LINE__) |
c9de560d AT |
704 | #else |
705 | #define mb_check_buddy(e4b) | |
706 | #endif | |
707 | ||
7c786059 CL |
708 | /* |
709 | * Divide blocks started from @first with length @len into | |
710 | * smaller chunks with power of 2 blocks. | |
711 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
712 | * then increase bb_counters[] for corresponded chunk size. | |
713 | */ | |
c9de560d | 714 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 715 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
716 | struct ext4_group_info *grp) |
717 | { | |
718 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
719 | ext4_grpblk_t min; |
720 | ext4_grpblk_t max; | |
721 | ext4_grpblk_t chunk; | |
69e43e8c | 722 | unsigned int border; |
c9de560d | 723 | |
7137d7a4 | 724 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
725 | |
726 | border = 2 << sb->s_blocksize_bits; | |
727 | ||
728 | while (len > 0) { | |
729 | /* find how many blocks can be covered since this position */ | |
730 | max = ffs(first | border) - 1; | |
731 | ||
732 | /* find how many blocks of power 2 we need to mark */ | |
733 | min = fls(len) - 1; | |
734 | ||
735 | if (max < min) | |
736 | min = max; | |
737 | chunk = 1 << min; | |
738 | ||
739 | /* mark multiblock chunks only */ | |
740 | grp->bb_counters[min]++; | |
741 | if (min > 0) | |
742 | mb_clear_bit(first >> min, | |
743 | buddy + sbi->s_mb_offsets[min]); | |
744 | ||
745 | len -= chunk; | |
746 | first += chunk; | |
747 | } | |
748 | } | |
749 | ||
8a57d9d6 CW |
750 | /* |
751 | * Cache the order of the largest free extent we have available in this block | |
752 | * group. | |
753 | */ | |
754 | static void | |
755 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
756 | { | |
757 | int i; | |
758 | int bits; | |
759 | ||
760 | grp->bb_largest_free_order = -1; /* uninit */ | |
761 | ||
762 | bits = sb->s_blocksize_bits + 1; | |
763 | for (i = bits; i >= 0; i--) { | |
764 | if (grp->bb_counters[i] > 0) { | |
765 | grp->bb_largest_free_order = i; | |
766 | break; | |
767 | } | |
768 | } | |
769 | } | |
770 | ||
089ceecc ES |
771 | static noinline_for_stack |
772 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
773 | void *buddy, void *bitmap, ext4_group_t group) |
774 | { | |
775 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 776 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 777 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
778 | ext4_grpblk_t i = 0; |
779 | ext4_grpblk_t first; | |
780 | ext4_grpblk_t len; | |
c9de560d AT |
781 | unsigned free = 0; |
782 | unsigned fragments = 0; | |
783 | unsigned long long period = get_cycles(); | |
784 | ||
785 | /* initialize buddy from bitmap which is aggregation | |
786 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 787 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
788 | grp->bb_first_free = i; |
789 | while (i < max) { | |
790 | fragments++; | |
791 | first = i; | |
ffad0a44 | 792 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
793 | len = i - first; |
794 | free += len; | |
795 | if (len > 1) | |
796 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
797 | else | |
798 | grp->bb_counters[0]++; | |
799 | if (i < max) | |
ffad0a44 | 800 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
801 | } |
802 | grp->bb_fragments = fragments; | |
803 | ||
804 | if (free != grp->bb_free) { | |
e29136f8 | 805 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
806 | "block bitmap and bg descriptor " |
807 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 808 | free, grp->bb_free); |
e56eb659 | 809 | /* |
163a203d | 810 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
811 | * corrupt and update bb_free using bitmap value |
812 | */ | |
c9de560d | 813 | grp->bb_free = free; |
db79e6d1 WS |
814 | ext4_mark_group_bitmap_corrupted(sb, group, |
815 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 816 | } |
8a57d9d6 | 817 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
818 | |
819 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
820 | ||
821 | period = get_cycles() - period; | |
49598e04 JP |
822 | spin_lock(&sbi->s_bal_lock); |
823 | sbi->s_mb_buddies_generated++; | |
824 | sbi->s_mb_generation_time += period; | |
825 | spin_unlock(&sbi->s_bal_lock); | |
c9de560d AT |
826 | } |
827 | ||
eabe0444 AS |
828 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
829 | { | |
830 | int count; | |
831 | int order = 1; | |
832 | void *buddy; | |
833 | ||
834 | while ((buddy = mb_find_buddy(e4b, order++, &count))) { | |
835 | ext4_set_bits(buddy, 0, count); | |
836 | } | |
837 | e4b->bd_info->bb_fragments = 0; | |
838 | memset(e4b->bd_info->bb_counters, 0, | |
839 | sizeof(*e4b->bd_info->bb_counters) * | |
840 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
841 | ||
842 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
843 | e4b->bd_bitmap, e4b->bd_group); | |
844 | } | |
845 | ||
c9de560d AT |
846 | /* The buddy information is attached the buddy cache inode |
847 | * for convenience. The information regarding each group | |
848 | * is loaded via ext4_mb_load_buddy. The information involve | |
849 | * block bitmap and buddy information. The information are | |
850 | * stored in the inode as | |
851 | * | |
852 | * { page } | |
c3a326a6 | 853 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
854 | * |
855 | * | |
856 | * one block each for bitmap and buddy information. | |
857 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 858 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
859 | * So it can have information regarding groups_per_page which |
860 | * is blocks_per_page/2 | |
8a57d9d6 CW |
861 | * |
862 | * Locking note: This routine takes the block group lock of all groups | |
863 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
864 | */ |
865 | ||
adb7ef60 | 866 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 867 | { |
8df9675f | 868 | ext4_group_t ngroups; |
c9de560d AT |
869 | int blocksize; |
870 | int blocks_per_page; | |
871 | int groups_per_page; | |
872 | int err = 0; | |
873 | int i; | |
813e5727 | 874 | ext4_group_t first_group, group; |
c9de560d AT |
875 | int first_block; |
876 | struct super_block *sb; | |
877 | struct buffer_head *bhs; | |
fa77dcfa | 878 | struct buffer_head **bh = NULL; |
c9de560d AT |
879 | struct inode *inode; |
880 | char *data; | |
881 | char *bitmap; | |
9b8b7d35 | 882 | struct ext4_group_info *grinfo; |
c9de560d | 883 | |
c9de560d AT |
884 | inode = page->mapping->host; |
885 | sb = inode->i_sb; | |
8df9675f | 886 | ngroups = ext4_get_groups_count(sb); |
93407472 | 887 | blocksize = i_blocksize(inode); |
09cbfeaf | 888 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d | 889 | |
d3df1453 RH |
890 | mb_debug(sb, "init page %lu\n", page->index); |
891 | ||
c9de560d AT |
892 | groups_per_page = blocks_per_page >> 1; |
893 | if (groups_per_page == 0) | |
894 | groups_per_page = 1; | |
895 | ||
896 | /* allocate buffer_heads to read bitmaps */ | |
897 | if (groups_per_page > 1) { | |
c9de560d | 898 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 899 | bh = kzalloc(i, gfp); |
813e5727 TT |
900 | if (bh == NULL) { |
901 | err = -ENOMEM; | |
c9de560d | 902 | goto out; |
813e5727 | 903 | } |
c9de560d AT |
904 | } else |
905 | bh = &bhs; | |
906 | ||
907 | first_group = page->index * blocks_per_page / 2; | |
908 | ||
909 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
910 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
911 | if (group >= ngroups) | |
c9de560d AT |
912 | break; |
913 | ||
813e5727 | 914 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
915 | /* |
916 | * If page is uptodate then we came here after online resize | |
917 | * which added some new uninitialized group info structs, so | |
918 | * we must skip all initialized uptodate buddies on the page, | |
919 | * which may be currently in use by an allocating task. | |
920 | */ | |
921 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
922 | bh[i] = NULL; | |
923 | continue; | |
924 | } | |
cfd73237 | 925 | bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
9008a58e DW |
926 | if (IS_ERR(bh[i])) { |
927 | err = PTR_ERR(bh[i]); | |
928 | bh[i] = NULL; | |
c9de560d | 929 | goto out; |
2ccb5fb9 | 930 | } |
d3df1453 | 931 | mb_debug(sb, "read bitmap for group %u\n", group); |
c9de560d AT |
932 | } |
933 | ||
934 | /* wait for I/O completion */ | |
813e5727 | 935 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
936 | int err2; |
937 | ||
938 | if (!bh[i]) | |
939 | continue; | |
940 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
941 | if (!err) | |
942 | err = err2; | |
813e5727 | 943 | } |
c9de560d AT |
944 | |
945 | first_block = page->index * blocks_per_page; | |
946 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 947 | group = (first_block + i) >> 1; |
8df9675f | 948 | if (group >= ngroups) |
c9de560d AT |
949 | break; |
950 | ||
9b8b7d35 AG |
951 | if (!bh[group - first_group]) |
952 | /* skip initialized uptodate buddy */ | |
953 | continue; | |
954 | ||
bbdc322f LC |
955 | if (!buffer_verified(bh[group - first_group])) |
956 | /* Skip faulty bitmaps */ | |
957 | continue; | |
958 | err = 0; | |
959 | ||
c9de560d AT |
960 | /* |
961 | * data carry information regarding this | |
962 | * particular group in the format specified | |
963 | * above | |
964 | * | |
965 | */ | |
966 | data = page_address(page) + (i * blocksize); | |
967 | bitmap = bh[group - first_group]->b_data; | |
968 | ||
969 | /* | |
970 | * We place the buddy block and bitmap block | |
971 | * close together | |
972 | */ | |
973 | if ((first_block + i) & 1) { | |
974 | /* this is block of buddy */ | |
975 | BUG_ON(incore == NULL); | |
d3df1453 | 976 | mb_debug(sb, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 977 | group, page->index, i * blocksize); |
f307333e | 978 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
979 | grinfo = ext4_get_group_info(sb, group); |
980 | grinfo->bb_fragments = 0; | |
981 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
982 | sizeof(*grinfo->bb_counters) * |
983 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
984 | /* |
985 | * incore got set to the group block bitmap below | |
986 | */ | |
7a2fcbf7 | 987 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
988 | /* init the buddy */ |
989 | memset(data, 0xff, blocksize); | |
c9de560d | 990 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 991 | ext4_unlock_group(sb, group); |
c9de560d AT |
992 | incore = NULL; |
993 | } else { | |
994 | /* this is block of bitmap */ | |
995 | BUG_ON(incore != NULL); | |
d3df1453 | 996 | mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 997 | group, page->index, i * blocksize); |
f307333e | 998 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
999 | |
1000 | /* see comments in ext4_mb_put_pa() */ | |
1001 | ext4_lock_group(sb, group); | |
1002 | memcpy(data, bitmap, blocksize); | |
1003 | ||
1004 | /* mark all preallocated blks used in in-core bitmap */ | |
1005 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 1006 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
1007 | ext4_unlock_group(sb, group); |
1008 | ||
1009 | /* set incore so that the buddy information can be | |
1010 | * generated using this | |
1011 | */ | |
1012 | incore = data; | |
1013 | } | |
1014 | } | |
1015 | SetPageUptodate(page); | |
1016 | ||
1017 | out: | |
1018 | if (bh) { | |
9b8b7d35 | 1019 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
1020 | brelse(bh[i]); |
1021 | if (bh != &bhs) | |
1022 | kfree(bh); | |
1023 | } | |
1024 | return err; | |
1025 | } | |
1026 | ||
eee4adc7 | 1027 | /* |
2de8807b AG |
1028 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
1029 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1030 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
1031 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 1032 | */ |
2de8807b | 1033 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1034 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1035 | { |
2de8807b AG |
1036 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1037 | int block, pnum, poff; | |
eee4adc7 | 1038 | int blocks_per_page; |
2de8807b AG |
1039 | struct page *page; |
1040 | ||
1041 | e4b->bd_buddy_page = NULL; | |
1042 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 1043 | |
09cbfeaf | 1044 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1045 | /* |
1046 | * the buddy cache inode stores the block bitmap | |
1047 | * and buddy information in consecutive blocks. | |
1048 | * So for each group we need two blocks. | |
1049 | */ | |
1050 | block = group * 2; | |
1051 | pnum = block / blocks_per_page; | |
2de8807b | 1052 | poff = block % blocks_per_page; |
adb7ef60 | 1053 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1054 | if (!page) |
c57ab39b | 1055 | return -ENOMEM; |
2de8807b AG |
1056 | BUG_ON(page->mapping != inode->i_mapping); |
1057 | e4b->bd_bitmap_page = page; | |
1058 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1059 | ||
1060 | if (blocks_per_page >= 2) { | |
1061 | /* buddy and bitmap are on the same page */ | |
1062 | return 0; | |
eee4adc7 | 1063 | } |
2de8807b AG |
1064 | |
1065 | block++; | |
1066 | pnum = block / blocks_per_page; | |
adb7ef60 | 1067 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1068 | if (!page) |
c57ab39b | 1069 | return -ENOMEM; |
2de8807b AG |
1070 | BUG_ON(page->mapping != inode->i_mapping); |
1071 | e4b->bd_buddy_page = page; | |
1072 | return 0; | |
eee4adc7 ES |
1073 | } |
1074 | ||
2de8807b | 1075 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1076 | { |
2de8807b AG |
1077 | if (e4b->bd_bitmap_page) { |
1078 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1079 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1080 | } |
1081 | if (e4b->bd_buddy_page) { | |
1082 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1083 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1084 | } |
eee4adc7 ES |
1085 | } |
1086 | ||
8a57d9d6 CW |
1087 | /* |
1088 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1089 | * block group lock of all groups for this page; do not hold the BG lock when | |
1090 | * calling this routine! | |
1091 | */ | |
b6a758ec | 1092 | static noinline_for_stack |
adb7ef60 | 1093 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1094 | { |
1095 | ||
b6a758ec | 1096 | struct ext4_group_info *this_grp; |
2de8807b AG |
1097 | struct ext4_buddy e4b; |
1098 | struct page *page; | |
1099 | int ret = 0; | |
b6a758ec | 1100 | |
b10a44c3 | 1101 | might_sleep(); |
d3df1453 | 1102 | mb_debug(sb, "init group %u\n", group); |
b6a758ec AK |
1103 | this_grp = ext4_get_group_info(sb, group); |
1104 | /* | |
08c3a813 AK |
1105 | * This ensures that we don't reinit the buddy cache |
1106 | * page which map to the group from which we are already | |
1107 | * allocating. If we are looking at the buddy cache we would | |
1108 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1109 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1110 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1111 | * page accessed. | |
b6a758ec | 1112 | */ |
adb7ef60 | 1113 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1114 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1115 | /* |
1116 | * somebody initialized the group | |
1117 | * return without doing anything | |
1118 | */ | |
b6a758ec AK |
1119 | goto err; |
1120 | } | |
2de8807b AG |
1121 | |
1122 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1123 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1124 | if (ret) |
1125 | goto err; | |
1126 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1127 | ret = -EIO; |
1128 | goto err; | |
1129 | } | |
b6a758ec | 1130 | |
2de8807b | 1131 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1132 | /* |
1133 | * If both the bitmap and buddy are in | |
1134 | * the same page we don't need to force | |
1135 | * init the buddy | |
1136 | */ | |
2de8807b AG |
1137 | ret = 0; |
1138 | goto err; | |
b6a758ec | 1139 | } |
2de8807b AG |
1140 | /* init buddy cache */ |
1141 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1142 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1143 | if (ret) |
1144 | goto err; | |
1145 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1146 | ret = -EIO; |
1147 | goto err; | |
1148 | } | |
b6a758ec | 1149 | err: |
2de8807b | 1150 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1151 | return ret; |
1152 | } | |
1153 | ||
8a57d9d6 CW |
1154 | /* |
1155 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1156 | * block group lock of all groups for this page; do not hold the BG lock when | |
1157 | * calling this routine! | |
1158 | */ | |
4ddfef7b | 1159 | static noinline_for_stack int |
adb7ef60 KK |
1160 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1161 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1162 | { |
c9de560d AT |
1163 | int blocks_per_page; |
1164 | int block; | |
1165 | int pnum; | |
1166 | int poff; | |
1167 | struct page *page; | |
fdf6c7a7 | 1168 | int ret; |
920313a7 AK |
1169 | struct ext4_group_info *grp; |
1170 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1171 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1172 | |
b10a44c3 | 1173 | might_sleep(); |
d3df1453 | 1174 | mb_debug(sb, "load group %u\n", group); |
c9de560d | 1175 | |
09cbfeaf | 1176 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1177 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1178 | |
1179 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1180 | e4b->bd_info = grp; |
c9de560d AT |
1181 | e4b->bd_sb = sb; |
1182 | e4b->bd_group = group; | |
1183 | e4b->bd_buddy_page = NULL; | |
1184 | e4b->bd_bitmap_page = NULL; | |
1185 | ||
f41c0750 | 1186 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1187 | /* |
1188 | * we need full data about the group | |
1189 | * to make a good selection | |
1190 | */ | |
adb7ef60 | 1191 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1192 | if (ret) |
1193 | return ret; | |
f41c0750 AK |
1194 | } |
1195 | ||
c9de560d AT |
1196 | /* |
1197 | * the buddy cache inode stores the block bitmap | |
1198 | * and buddy information in consecutive blocks. | |
1199 | * So for each group we need two blocks. | |
1200 | */ | |
1201 | block = group * 2; | |
1202 | pnum = block / blocks_per_page; | |
1203 | poff = block % blocks_per_page; | |
1204 | ||
1205 | /* we could use find_or_create_page(), but it locks page | |
1206 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1207 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1208 | if (page == NULL || !PageUptodate(page)) { |
1209 | if (page) | |
920313a7 AK |
1210 | /* |
1211 | * drop the page reference and try | |
1212 | * to get the page with lock. If we | |
1213 | * are not uptodate that implies | |
1214 | * somebody just created the page but | |
1215 | * is yet to initialize the same. So | |
1216 | * wait for it to initialize. | |
1217 | */ | |
09cbfeaf | 1218 | put_page(page); |
adb7ef60 | 1219 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1220 | if (page) { |
1221 | BUG_ON(page->mapping != inode->i_mapping); | |
1222 | if (!PageUptodate(page)) { | |
adb7ef60 | 1223 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1224 | if (ret) { |
1225 | unlock_page(page); | |
1226 | goto err; | |
1227 | } | |
c9de560d AT |
1228 | mb_cmp_bitmaps(e4b, page_address(page) + |
1229 | (poff * sb->s_blocksize)); | |
1230 | } | |
1231 | unlock_page(page); | |
1232 | } | |
1233 | } | |
c57ab39b YL |
1234 | if (page == NULL) { |
1235 | ret = -ENOMEM; | |
1236 | goto err; | |
1237 | } | |
1238 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1239 | ret = -EIO; |
c9de560d | 1240 | goto err; |
fdf6c7a7 | 1241 | } |
2457aec6 MG |
1242 | |
1243 | /* Pages marked accessed already */ | |
c9de560d AT |
1244 | e4b->bd_bitmap_page = page; |
1245 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1246 | |
1247 | block++; | |
1248 | pnum = block / blocks_per_page; | |
1249 | poff = block % blocks_per_page; | |
1250 | ||
2457aec6 | 1251 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1252 | if (page == NULL || !PageUptodate(page)) { |
1253 | if (page) | |
09cbfeaf | 1254 | put_page(page); |
adb7ef60 | 1255 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1256 | if (page) { |
1257 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1258 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1259 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1260 | gfp); | |
fdf6c7a7 SF |
1261 | if (ret) { |
1262 | unlock_page(page); | |
1263 | goto err; | |
1264 | } | |
1265 | } | |
c9de560d AT |
1266 | unlock_page(page); |
1267 | } | |
1268 | } | |
c57ab39b YL |
1269 | if (page == NULL) { |
1270 | ret = -ENOMEM; | |
1271 | goto err; | |
1272 | } | |
1273 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1274 | ret = -EIO; |
c9de560d | 1275 | goto err; |
fdf6c7a7 | 1276 | } |
2457aec6 MG |
1277 | |
1278 | /* Pages marked accessed already */ | |
c9de560d AT |
1279 | e4b->bd_buddy_page = page; |
1280 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d | 1281 | |
c9de560d AT |
1282 | return 0; |
1283 | ||
1284 | err: | |
26626f11 | 1285 | if (page) |
09cbfeaf | 1286 | put_page(page); |
c9de560d | 1287 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1288 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1289 | if (e4b->bd_buddy_page) |
09cbfeaf | 1290 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1291 | e4b->bd_buddy = NULL; |
1292 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1293 | return ret; |
c9de560d AT |
1294 | } |
1295 | ||
adb7ef60 KK |
1296 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1297 | struct ext4_buddy *e4b) | |
1298 | { | |
1299 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1300 | } | |
1301 | ||
e39e07fd | 1302 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1303 | { |
1304 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1305 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1306 | if (e4b->bd_buddy_page) |
09cbfeaf | 1307 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1308 | } |
1309 | ||
1310 | ||
1311 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1312 | { | |
1313 | int order = 1; | |
b5cb316c | 1314 | int bb_incr = 1 << (e4b->bd_blkbits - 1); |
c9de560d AT |
1315 | void *bb; |
1316 | ||
c5e8f3f3 | 1317 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1318 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1319 | ||
c5e8f3f3 | 1320 | bb = e4b->bd_buddy; |
c9de560d AT |
1321 | while (order <= e4b->bd_blkbits + 1) { |
1322 | block = block >> 1; | |
1323 | if (!mb_test_bit(block, bb)) { | |
1324 | /* this block is part of buddy of order 'order' */ | |
1325 | return order; | |
1326 | } | |
b5cb316c NS |
1327 | bb += bb_incr; |
1328 | bb_incr >>= 1; | |
c9de560d AT |
1329 | order++; |
1330 | } | |
1331 | return 0; | |
1332 | } | |
1333 | ||
955ce5f5 | 1334 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1335 | { |
1336 | __u32 *addr; | |
1337 | ||
1338 | len = cur + len; | |
1339 | while (cur < len) { | |
1340 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1341 | /* fast path: clear whole word at once */ | |
1342 | addr = bm + (cur >> 3); | |
1343 | *addr = 0; | |
1344 | cur += 32; | |
1345 | continue; | |
1346 | } | |
955ce5f5 | 1347 | mb_clear_bit(cur, bm); |
c9de560d AT |
1348 | cur++; |
1349 | } | |
1350 | } | |
1351 | ||
eabe0444 AS |
1352 | /* clear bits in given range |
1353 | * will return first found zero bit if any, -1 otherwise | |
1354 | */ | |
1355 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1356 | { | |
1357 | __u32 *addr; | |
1358 | int zero_bit = -1; | |
1359 | ||
1360 | len = cur + len; | |
1361 | while (cur < len) { | |
1362 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1363 | /* fast path: clear whole word at once */ | |
1364 | addr = bm + (cur >> 3); | |
1365 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1366 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1367 | *addr = 0; | |
1368 | cur += 32; | |
1369 | continue; | |
1370 | } | |
1371 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1372 | zero_bit = cur; | |
1373 | cur++; | |
1374 | } | |
1375 | ||
1376 | return zero_bit; | |
1377 | } | |
1378 | ||
c3e94d1d | 1379 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1380 | { |
1381 | __u32 *addr; | |
1382 | ||
1383 | len = cur + len; | |
1384 | while (cur < len) { | |
1385 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1386 | /* fast path: set whole word at once */ | |
1387 | addr = bm + (cur >> 3); | |
1388 | *addr = 0xffffffff; | |
1389 | cur += 32; | |
1390 | continue; | |
1391 | } | |
955ce5f5 | 1392 | mb_set_bit(cur, bm); |
c9de560d AT |
1393 | cur++; |
1394 | } | |
1395 | } | |
1396 | ||
eabe0444 AS |
1397 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
1398 | { | |
1399 | if (mb_test_bit(*bit + side, bitmap)) { | |
1400 | mb_clear_bit(*bit, bitmap); | |
1401 | (*bit) -= side; | |
1402 | return 1; | |
1403 | } | |
1404 | else { | |
1405 | (*bit) += side; | |
1406 | mb_set_bit(*bit, bitmap); | |
1407 | return -1; | |
1408 | } | |
1409 | } | |
1410 | ||
1411 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1412 | { | |
1413 | int max; | |
1414 | int order = 1; | |
1415 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1416 | ||
1417 | while (buddy) { | |
1418 | void *buddy2; | |
1419 | ||
1420 | /* Bits in range [first; last] are known to be set since | |
1421 | * corresponding blocks were allocated. Bits in range | |
1422 | * (first; last) will stay set because they form buddies on | |
1423 | * upper layer. We just deal with borders if they don't | |
1424 | * align with upper layer and then go up. | |
1425 | * Releasing entire group is all about clearing | |
1426 | * single bit of highest order buddy. | |
1427 | */ | |
1428 | ||
1429 | /* Example: | |
1430 | * --------------------------------- | |
1431 | * | 1 | 1 | 1 | 1 | | |
1432 | * --------------------------------- | |
1433 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1434 | * --------------------------------- | |
1435 | * 0 1 2 3 4 5 6 7 | |
1436 | * \_____________________/ | |
1437 | * | |
1438 | * Neither [1] nor [6] is aligned to above layer. | |
1439 | * Left neighbour [0] is free, so mark it busy, | |
1440 | * decrease bb_counters and extend range to | |
1441 | * [0; 6] | |
1442 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1443 | * mark [6] free, increase bb_counters and shrink range to | |
1444 | * [0; 5]. | |
1445 | * Then shift range to [0; 2], go up and do the same. | |
1446 | */ | |
1447 | ||
1448 | ||
1449 | if (first & 1) | |
1450 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1451 | if (!(last & 1)) | |
1452 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1453 | if (first > last) | |
1454 | break; | |
1455 | order++; | |
1456 | ||
1457 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1458 | mb_clear_bits(buddy, first, last - first + 1); | |
1459 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1460 | break; | |
1461 | } | |
1462 | first >>= 1; | |
1463 | last >>= 1; | |
1464 | buddy = buddy2; | |
1465 | } | |
1466 | } | |
1467 | ||
7e5a8cdd | 1468 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1469 | int first, int count) |
c9de560d | 1470 | { |
eabe0444 AS |
1471 | int left_is_free = 0; |
1472 | int right_is_free = 0; | |
1473 | int block; | |
1474 | int last = first + count - 1; | |
c9de560d AT |
1475 | struct super_block *sb = e4b->bd_sb; |
1476 | ||
c99d1e6e TT |
1477 | if (WARN_ON(count == 0)) |
1478 | return; | |
eabe0444 | 1479 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1480 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1481 | /* Don't bother if the block group is corrupt. */ |
1482 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1483 | return; | |
1484 | ||
c9de560d AT |
1485 | mb_check_buddy(e4b); |
1486 | mb_free_blocks_double(inode, e4b, first, count); | |
1487 | ||
07b5b8e1 | 1488 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1489 | e4b->bd_info->bb_free += count; |
1490 | if (first < e4b->bd_info->bb_first_free) | |
1491 | e4b->bd_info->bb_first_free = first; | |
1492 | ||
eabe0444 AS |
1493 | /* access memory sequentially: check left neighbour, |
1494 | * clear range and then check right neighbour | |
1495 | */ | |
c9de560d | 1496 | if (first != 0) |
eabe0444 AS |
1497 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1498 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1499 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1500 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1501 | ||
1502 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1503 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1504 | ext4_fsblk_t blocknr; |
1505 | ||
1506 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1507 | blocknr += EXT4_C2B(sbi, block); |
eabe0444 AS |
1508 | ext4_grp_locked_error(sb, e4b->bd_group, |
1509 | inode ? inode->i_ino : 0, | |
1510 | blocknr, | |
1511 | "freeing already freed block " | |
163a203d DW |
1512 | "(bit %u); block bitmap corrupt.", |
1513 | block); | |
db79e6d1 WS |
1514 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1515 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
eabe0444 AS |
1516 | mb_regenerate_buddy(e4b); |
1517 | goto done; | |
1518 | } | |
1519 | ||
1520 | /* let's maintain fragments counter */ | |
1521 | if (left_is_free && right_is_free) | |
c9de560d | 1522 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1523 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1524 | e4b->bd_info->bb_fragments++; |
1525 | ||
eabe0444 AS |
1526 | /* buddy[0] == bd_bitmap is a special case, so handle |
1527 | * it right away and let mb_buddy_mark_free stay free of | |
1528 | * zero order checks. | |
1529 | * Check if neighbours are to be coaleasced, | |
1530 | * adjust bitmap bb_counters and borders appropriately. | |
1531 | */ | |
1532 | if (first & 1) { | |
1533 | first += !left_is_free; | |
1534 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1535 | } | |
1536 | if (!(last & 1)) { | |
1537 | last -= !right_is_free; | |
1538 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1539 | } | |
c9de560d | 1540 | |
eabe0444 AS |
1541 | if (first <= last) |
1542 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1543 | |
eabe0444 | 1544 | done: |
8a57d9d6 | 1545 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1546 | mb_check_buddy(e4b); |
c9de560d AT |
1547 | } |
1548 | ||
15c006a2 | 1549 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1550 | int needed, struct ext4_free_extent *ex) |
1551 | { | |
1552 | int next = block; | |
15c006a2 | 1553 | int max, order; |
c9de560d AT |
1554 | void *buddy; |
1555 | ||
bc8e6740 | 1556 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1557 | BUG_ON(ex == NULL); |
1558 | ||
15c006a2 | 1559 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1560 | BUG_ON(buddy == NULL); |
1561 | BUG_ON(block >= max); | |
1562 | if (mb_test_bit(block, buddy)) { | |
1563 | ex->fe_len = 0; | |
1564 | ex->fe_start = 0; | |
1565 | ex->fe_group = 0; | |
1566 | return 0; | |
1567 | } | |
1568 | ||
15c006a2 RD |
1569 | /* find actual order */ |
1570 | order = mb_find_order_for_block(e4b, block); | |
1571 | block = block >> order; | |
c9de560d AT |
1572 | |
1573 | ex->fe_len = 1 << order; | |
1574 | ex->fe_start = block << order; | |
1575 | ex->fe_group = e4b->bd_group; | |
1576 | ||
1577 | /* calc difference from given start */ | |
1578 | next = next - ex->fe_start; | |
1579 | ex->fe_len -= next; | |
1580 | ex->fe_start += next; | |
1581 | ||
1582 | while (needed > ex->fe_len && | |
d8ec0c39 | 1583 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1584 | |
1585 | if (block + 1 >= max) | |
1586 | break; | |
1587 | ||
1588 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1589 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1590 | break; |
1591 | ||
b051d8dc | 1592 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1593 | |
c9de560d AT |
1594 | block = next >> order; |
1595 | ex->fe_len += 1 << order; | |
1596 | } | |
1597 | ||
31562b95 | 1598 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1599 | /* Should never happen! (but apparently sometimes does?!?) */ |
1600 | WARN_ON(1); | |
1601 | ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent " | |
1602 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1603 | block, order, needed, ex->fe_group, ex->fe_start, | |
1604 | ex->fe_len, ex->fe_logical); | |
1605 | ex->fe_len = 0; | |
1606 | ex->fe_start = 0; | |
1607 | ex->fe_group = 0; | |
1608 | } | |
c9de560d AT |
1609 | return ex->fe_len; |
1610 | } | |
1611 | ||
1612 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1613 | { | |
1614 | int ord; | |
1615 | int mlen = 0; | |
1616 | int max = 0; | |
1617 | int cur; | |
1618 | int start = ex->fe_start; | |
1619 | int len = ex->fe_len; | |
1620 | unsigned ret = 0; | |
1621 | int len0 = len; | |
1622 | void *buddy; | |
1623 | ||
1624 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1625 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1626 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1627 | mb_check_buddy(e4b); |
1628 | mb_mark_used_double(e4b, start, len); | |
1629 | ||
07b5b8e1 | 1630 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1631 | e4b->bd_info->bb_free -= len; |
1632 | if (e4b->bd_info->bb_first_free == start) | |
1633 | e4b->bd_info->bb_first_free += len; | |
1634 | ||
1635 | /* let's maintain fragments counter */ | |
1636 | if (start != 0) | |
c5e8f3f3 | 1637 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1638 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1639 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1640 | if (mlen && max) |
1641 | e4b->bd_info->bb_fragments++; | |
1642 | else if (!mlen && !max) | |
1643 | e4b->bd_info->bb_fragments--; | |
1644 | ||
1645 | /* let's maintain buddy itself */ | |
1646 | while (len) { | |
1647 | ord = mb_find_order_for_block(e4b, start); | |
1648 | ||
1649 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1650 | /* the whole chunk may be allocated at once! */ | |
1651 | mlen = 1 << ord; | |
1652 | buddy = mb_find_buddy(e4b, ord, &max); | |
1653 | BUG_ON((start >> ord) >= max); | |
1654 | mb_set_bit(start >> ord, buddy); | |
1655 | e4b->bd_info->bb_counters[ord]--; | |
1656 | start += mlen; | |
1657 | len -= mlen; | |
1658 | BUG_ON(len < 0); | |
1659 | continue; | |
1660 | } | |
1661 | ||
1662 | /* store for history */ | |
1663 | if (ret == 0) | |
1664 | ret = len | (ord << 16); | |
1665 | ||
1666 | /* we have to split large buddy */ | |
1667 | BUG_ON(ord <= 0); | |
1668 | buddy = mb_find_buddy(e4b, ord, &max); | |
1669 | mb_set_bit(start >> ord, buddy); | |
1670 | e4b->bd_info->bb_counters[ord]--; | |
1671 | ||
1672 | ord--; | |
1673 | cur = (start >> ord) & ~1U; | |
1674 | buddy = mb_find_buddy(e4b, ord, &max); | |
1675 | mb_clear_bit(cur, buddy); | |
1676 | mb_clear_bit(cur + 1, buddy); | |
1677 | e4b->bd_info->bb_counters[ord]++; | |
1678 | e4b->bd_info->bb_counters[ord]++; | |
1679 | } | |
8a57d9d6 | 1680 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1681 | |
c5e8f3f3 | 1682 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1683 | mb_check_buddy(e4b); |
1684 | ||
1685 | return ret; | |
1686 | } | |
1687 | ||
1688 | /* | |
1689 | * Must be called under group lock! | |
1690 | */ | |
1691 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1692 | struct ext4_buddy *e4b) | |
1693 | { | |
1694 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1695 | int ret; | |
1696 | ||
1697 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1698 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1699 | ||
1700 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1701 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1702 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1703 | ||
1704 | /* preallocation can change ac_b_ex, thus we store actually | |
1705 | * allocated blocks for history */ | |
1706 | ac->ac_f_ex = ac->ac_b_ex; | |
1707 | ||
1708 | ac->ac_status = AC_STATUS_FOUND; | |
1709 | ac->ac_tail = ret & 0xffff; | |
1710 | ac->ac_buddy = ret >> 16; | |
1711 | ||
c3a326a6 AK |
1712 | /* |
1713 | * take the page reference. We want the page to be pinned | |
1714 | * so that we don't get a ext4_mb_init_cache_call for this | |
1715 | * group until we update the bitmap. That would mean we | |
1716 | * double allocate blocks. The reference is dropped | |
1717 | * in ext4_mb_release_context | |
1718 | */ | |
c9de560d AT |
1719 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1720 | get_page(ac->ac_bitmap_page); | |
1721 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1722 | get_page(ac->ac_buddy_page); | |
c9de560d | 1723 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1724 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1725 | spin_lock(&sbi->s_md_lock); |
1726 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1727 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1728 | spin_unlock(&sbi->s_md_lock); | |
1729 | } | |
53f86b17 RH |
1730 | /* |
1731 | * As we've just preallocated more space than | |
1732 | * user requested originally, we store allocated | |
1733 | * space in a special descriptor. | |
1734 | */ | |
1735 | if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
1736 | ext4_mb_new_preallocation(ac); | |
1737 | ||
c9de560d AT |
1738 | } |
1739 | ||
c9de560d AT |
1740 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
1741 | struct ext4_buddy *e4b, | |
1742 | int finish_group) | |
1743 | { | |
1744 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1745 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1746 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1747 | struct ext4_free_extent ex; | |
1748 | int max; | |
1749 | ||
032115fc AK |
1750 | if (ac->ac_status == AC_STATUS_FOUND) |
1751 | return; | |
c9de560d AT |
1752 | /* |
1753 | * We don't want to scan for a whole year | |
1754 | */ | |
1755 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1756 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1757 | ac->ac_status = AC_STATUS_BREAK; | |
1758 | return; | |
1759 | } | |
1760 | ||
1761 | /* | |
1762 | * Haven't found good chunk so far, let's continue | |
1763 | */ | |
1764 | if (bex->fe_len < gex->fe_len) | |
1765 | return; | |
1766 | ||
1767 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1768 | && bex->fe_group == e4b->bd_group) { | |
1769 | /* recheck chunk's availability - we don't know | |
1770 | * when it was found (within this lock-unlock | |
1771 | * period or not) */ | |
15c006a2 | 1772 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1773 | if (max >= gex->fe_len) { |
1774 | ext4_mb_use_best_found(ac, e4b); | |
1775 | return; | |
1776 | } | |
1777 | } | |
1778 | } | |
1779 | ||
1780 | /* | |
1781 | * The routine checks whether found extent is good enough. If it is, | |
1782 | * then the extent gets marked used and flag is set to the context | |
1783 | * to stop scanning. Otherwise, the extent is compared with the | |
1784 | * previous found extent and if new one is better, then it's stored | |
1785 | * in the context. Later, the best found extent will be used, if | |
1786 | * mballoc can't find good enough extent. | |
1787 | * | |
1788 | * FIXME: real allocation policy is to be designed yet! | |
1789 | */ | |
1790 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1791 | struct ext4_free_extent *ex, | |
1792 | struct ext4_buddy *e4b) | |
1793 | { | |
1794 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1795 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1796 | ||
1797 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1798 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1799 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1800 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1801 | ||
1802 | ac->ac_found++; | |
1803 | ||
1804 | /* | |
1805 | * The special case - take what you catch first | |
1806 | */ | |
1807 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1808 | *bex = *ex; | |
1809 | ext4_mb_use_best_found(ac, e4b); | |
1810 | return; | |
1811 | } | |
1812 | ||
1813 | /* | |
1814 | * Let's check whether the chuck is good enough | |
1815 | */ | |
1816 | if (ex->fe_len == gex->fe_len) { | |
1817 | *bex = *ex; | |
1818 | ext4_mb_use_best_found(ac, e4b); | |
1819 | return; | |
1820 | } | |
1821 | ||
1822 | /* | |
1823 | * If this is first found extent, just store it in the context | |
1824 | */ | |
1825 | if (bex->fe_len == 0) { | |
1826 | *bex = *ex; | |
1827 | return; | |
1828 | } | |
1829 | ||
1830 | /* | |
1831 | * If new found extent is better, store it in the context | |
1832 | */ | |
1833 | if (bex->fe_len < gex->fe_len) { | |
1834 | /* if the request isn't satisfied, any found extent | |
1835 | * larger than previous best one is better */ | |
1836 | if (ex->fe_len > bex->fe_len) | |
1837 | *bex = *ex; | |
1838 | } else if (ex->fe_len > gex->fe_len) { | |
1839 | /* if the request is satisfied, then we try to find | |
1840 | * an extent that still satisfy the request, but is | |
1841 | * smaller than previous one */ | |
1842 | if (ex->fe_len < bex->fe_len) | |
1843 | *bex = *ex; | |
1844 | } | |
1845 | ||
1846 | ext4_mb_check_limits(ac, e4b, 0); | |
1847 | } | |
1848 | ||
089ceecc ES |
1849 | static noinline_for_stack |
1850 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1851 | struct ext4_buddy *e4b) |
1852 | { | |
1853 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1854 | ext4_group_t group = ex.fe_group; | |
1855 | int max; | |
1856 | int err; | |
1857 | ||
1858 | BUG_ON(ex.fe_len <= 0); | |
1859 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1860 | if (err) | |
1861 | return err; | |
1862 | ||
1863 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1864 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1865 | |
1866 | if (max > 0) { | |
1867 | ac->ac_b_ex = ex; | |
1868 | ext4_mb_use_best_found(ac, e4b); | |
1869 | } | |
1870 | ||
1871 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1872 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1873 | |
1874 | return 0; | |
1875 | } | |
1876 | ||
089ceecc ES |
1877 | static noinline_for_stack |
1878 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1879 | struct ext4_buddy *e4b) |
1880 | { | |
1881 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1882 | int max; | |
1883 | int err; | |
1884 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1885 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1886 | struct ext4_free_extent ex; |
1887 | ||
1888 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1889 | return 0; | |
838cd0cf YY |
1890 | if (grp->bb_free == 0) |
1891 | return 0; | |
c9de560d AT |
1892 | |
1893 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1894 | if (err) | |
1895 | return err; | |
1896 | ||
163a203d DW |
1897 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
1898 | ext4_mb_unload_buddy(e4b); | |
1899 | return 0; | |
1900 | } | |
1901 | ||
c9de560d | 1902 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 1903 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 1904 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 1905 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
1906 | |
1907 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1908 | ext4_fsblk_t start; | |
1909 | ||
5661bd68 AM |
1910 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1911 | ex.fe_start; | |
c9de560d AT |
1912 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1913 | if (do_div(start, sbi->s_stripe) == 0) { | |
1914 | ac->ac_found++; | |
1915 | ac->ac_b_ex = ex; | |
1916 | ext4_mb_use_best_found(ac, e4b); | |
1917 | } | |
1918 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1919 | BUG_ON(ex.fe_len <= 0); | |
1920 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1921 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1922 | ac->ac_found++; | |
1923 | ac->ac_b_ex = ex; | |
1924 | ext4_mb_use_best_found(ac, e4b); | |
1925 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1926 | /* Sometimes, caller may want to merge even small | |
1927 | * number of blocks to an existing extent */ | |
1928 | BUG_ON(ex.fe_len <= 0); | |
1929 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1930 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1931 | ac->ac_found++; | |
1932 | ac->ac_b_ex = ex; | |
1933 | ext4_mb_use_best_found(ac, e4b); | |
1934 | } | |
1935 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1936 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1937 | |
1938 | return 0; | |
1939 | } | |
1940 | ||
1941 | /* | |
1942 | * The routine scans buddy structures (not bitmap!) from given order | |
1943 | * to max order and tries to find big enough chunk to satisfy the req | |
1944 | */ | |
089ceecc ES |
1945 | static noinline_for_stack |
1946 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1947 | struct ext4_buddy *e4b) |
1948 | { | |
1949 | struct super_block *sb = ac->ac_sb; | |
1950 | struct ext4_group_info *grp = e4b->bd_info; | |
1951 | void *buddy; | |
1952 | int i; | |
1953 | int k; | |
1954 | int max; | |
1955 | ||
1956 | BUG_ON(ac->ac_2order <= 0); | |
1957 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1958 | if (grp->bb_counters[i] == 0) | |
1959 | continue; | |
1960 | ||
1961 | buddy = mb_find_buddy(e4b, i, &max); | |
1962 | BUG_ON(buddy == NULL); | |
1963 | ||
ffad0a44 | 1964 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
1965 | if (k >= max) { |
1966 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
1967 | "%d free clusters of order %d. But found 0", | |
1968 | grp->bb_counters[i], i); | |
1969 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
1970 | e4b->bd_group, | |
1971 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
1972 | break; | |
1973 | } | |
c9de560d AT |
1974 | ac->ac_found++; |
1975 | ||
1976 | ac->ac_b_ex.fe_len = 1 << i; | |
1977 | ac->ac_b_ex.fe_start = k << i; | |
1978 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1979 | ||
1980 | ext4_mb_use_best_found(ac, e4b); | |
1981 | ||
53f86b17 | 1982 | BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
c9de560d AT |
1983 | |
1984 | if (EXT4_SB(sb)->s_mb_stats) | |
1985 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1986 | ||
1987 | break; | |
1988 | } | |
1989 | } | |
1990 | ||
1991 | /* | |
1992 | * The routine scans the group and measures all found extents. | |
1993 | * In order to optimize scanning, caller must pass number of | |
1994 | * free blocks in the group, so the routine can know upper limit. | |
1995 | */ | |
089ceecc ES |
1996 | static noinline_for_stack |
1997 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1998 | struct ext4_buddy *e4b) |
1999 | { | |
2000 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 2001 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2002 | struct ext4_free_extent ex; |
2003 | int i; | |
2004 | int free; | |
2005 | ||
2006 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
2007 | if (WARN_ON(free <= 0)) |
2008 | return; | |
c9de560d AT |
2009 | |
2010 | i = e4b->bd_info->bb_first_free; | |
2011 | ||
2012 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 2013 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
2014 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
2015 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 2016 | /* |
e56eb659 | 2017 | * IF we have corrupt bitmap, we won't find any |
26346ff6 | 2018 | * free blocks even though group info says we |
b483bb77 | 2019 | * have free blocks |
26346ff6 | 2020 | */ |
e29136f8 | 2021 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2022 | "%d free clusters as per " |
fde4d95a | 2023 | "group info. But bitmap says 0", |
26346ff6 | 2024 | free); |
736dedbb WS |
2025 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2026 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
2027 | break; |
2028 | } | |
2029 | ||
15c006a2 | 2030 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2031 | if (WARN_ON(ex.fe_len <= 0)) |
2032 | break; | |
26346ff6 | 2033 | if (free < ex.fe_len) { |
e29136f8 | 2034 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2035 | "%d free clusters as per " |
fde4d95a | 2036 | "group info. But got %d blocks", |
26346ff6 | 2037 | free, ex.fe_len); |
736dedbb WS |
2038 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2039 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
2040 | /* |
2041 | * The number of free blocks differs. This mostly | |
2042 | * indicate that the bitmap is corrupt. So exit | |
2043 | * without claiming the space. | |
2044 | */ | |
2045 | break; | |
26346ff6 | 2046 | } |
ab0c00fc | 2047 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2048 | ext4_mb_measure_extent(ac, &ex, e4b); |
2049 | ||
2050 | i += ex.fe_len; | |
2051 | free -= ex.fe_len; | |
2052 | } | |
2053 | ||
2054 | ext4_mb_check_limits(ac, e4b, 1); | |
2055 | } | |
2056 | ||
2057 | /* | |
2058 | * This is a special case for storages like raid5 | |
506bf2d8 | 2059 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2060 | */ |
089ceecc ES |
2061 | static noinline_for_stack |
2062 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2063 | struct ext4_buddy *e4b) |
2064 | { | |
2065 | struct super_block *sb = ac->ac_sb; | |
2066 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2067 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2068 | struct ext4_free_extent ex; |
2069 | ext4_fsblk_t first_group_block; | |
2070 | ext4_fsblk_t a; | |
2071 | ext4_grpblk_t i; | |
2072 | int max; | |
2073 | ||
2074 | BUG_ON(sbi->s_stripe == 0); | |
2075 | ||
2076 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2077 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2078 | ||
c9de560d AT |
2079 | a = first_group_block + sbi->s_stripe - 1; |
2080 | do_div(a, sbi->s_stripe); | |
2081 | i = (a * sbi->s_stripe) - first_group_block; | |
2082 | ||
7137d7a4 | 2083 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2084 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2085 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2086 | if (max >= sbi->s_stripe) { |
2087 | ac->ac_found++; | |
ab0c00fc | 2088 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2089 | ac->ac_b_ex = ex; |
2090 | ext4_mb_use_best_found(ac, e4b); | |
2091 | break; | |
2092 | } | |
2093 | } | |
2094 | i += sbi->s_stripe; | |
2095 | } | |
2096 | } | |
2097 | ||
42ac1848 | 2098 | /* |
8ef123fe | 2099 | * This is also called BEFORE we load the buddy bitmap. |
42ac1848 | 2100 | * Returns either 1 or 0 indicating that the group is either suitable |
8ef123fe | 2101 | * for the allocation or not. |
42ac1848 | 2102 | */ |
8ef123fe | 2103 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
c9de560d AT |
2104 | ext4_group_t group, int cr) |
2105 | { | |
8ef123fe | 2106 | ext4_grpblk_t free, fragments; |
a4912123 | 2107 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2108 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2109 | ||
2110 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2111 | |
dddcd2f9 | 2112 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
8ef123fe | 2113 | return false; |
01fc48e8 | 2114 | |
dddcd2f9 | 2115 | free = grp->bb_free; |
2116 | if (free == 0) | |
8ef123fe | 2117 | return false; |
c9de560d | 2118 | |
c9de560d | 2119 | fragments = grp->bb_fragments; |
c9de560d | 2120 | if (fragments == 0) |
8ef123fe | 2121 | return false; |
c9de560d AT |
2122 | |
2123 | switch (cr) { | |
2124 | case 0: | |
2125 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2126 | |
a4912123 TT |
2127 | /* Avoid using the first bg of a flexgroup for data files */ |
2128 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2129 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2130 | ((group % flex_size) == 0)) | |
8ef123fe | 2131 | return false; |
a4912123 | 2132 | |
dddcd2f9 | 2133 | if (free < ac->ac_g_ex.fe_len) |
2134 | return false; | |
2135 | ||
2136 | if (ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) | |
8ef123fe | 2137 | return true; |
40ae3487 TT |
2138 | |
2139 | if (grp->bb_largest_free_order < ac->ac_2order) | |
8ef123fe | 2140 | return false; |
40ae3487 | 2141 | |
8ef123fe | 2142 | return true; |
c9de560d AT |
2143 | case 1: |
2144 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2145 | return true; |
c9de560d AT |
2146 | break; |
2147 | case 2: | |
2148 | if (free >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2149 | return true; |
c9de560d AT |
2150 | break; |
2151 | case 3: | |
8ef123fe | 2152 | return true; |
c9de560d AT |
2153 | default: |
2154 | BUG(); | |
2155 | } | |
2156 | ||
8ef123fe RH |
2157 | return false; |
2158 | } | |
2159 | ||
2160 | /* | |
2161 | * This could return negative error code if something goes wrong | |
2162 | * during ext4_mb_init_group(). This should not be called with | |
2163 | * ext4_lock_group() held. | |
2164 | */ | |
2165 | static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, | |
2166 | ext4_group_t group, int cr) | |
2167 | { | |
2168 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
99377830 | 2169 | struct super_block *sb = ac->ac_sb; |
c1d2c7d4 | 2170 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
99377830 | 2171 | bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
8ef123fe RH |
2172 | ext4_grpblk_t free; |
2173 | int ret = 0; | |
2174 | ||
99377830 RH |
2175 | if (should_lock) |
2176 | ext4_lock_group(sb, group); | |
8ef123fe RH |
2177 | free = grp->bb_free; |
2178 | if (free == 0) | |
2179 | goto out; | |
2180 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2181 | goto out; | |
2182 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
2183 | goto out; | |
99377830 RH |
2184 | if (should_lock) |
2185 | ext4_unlock_group(sb, group); | |
8ef123fe RH |
2186 | |
2187 | /* We only do this if the grp has never been initialized */ | |
2188 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
c1d2c7d4 AZ |
2189 | struct ext4_group_desc *gdp = |
2190 | ext4_get_group_desc(sb, group, NULL); | |
2191 | int ret; | |
2192 | ||
2193 | /* cr=0/1 is a very optimistic search to find large | |
2194 | * good chunks almost for free. If buddy data is not | |
2195 | * ready, then this optimization makes no sense. But | |
2196 | * we never skip the first block group in a flex_bg, | |
2197 | * since this gets used for metadata block allocation, | |
2198 | * and we want to make sure we locate metadata blocks | |
2199 | * in the first block group in the flex_bg if possible. | |
2200 | */ | |
2201 | if (cr < 2 && | |
2202 | (!sbi->s_log_groups_per_flex || | |
2203 | ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && | |
2204 | !(ext4_has_group_desc_csum(sb) && | |
2205 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) | |
2206 | return 0; | |
2207 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); | |
8ef123fe RH |
2208 | if (ret) |
2209 | return ret; | |
2210 | } | |
2211 | ||
99377830 RH |
2212 | if (should_lock) |
2213 | ext4_lock_group(sb, group); | |
8ef123fe RH |
2214 | ret = ext4_mb_good_group(ac, group, cr); |
2215 | out: | |
99377830 RH |
2216 | if (should_lock) |
2217 | ext4_unlock_group(sb, group); | |
8ef123fe | 2218 | return ret; |
c9de560d AT |
2219 | } |
2220 | ||
cfd73237 AZ |
2221 | /* |
2222 | * Start prefetching @nr block bitmaps starting at @group. | |
2223 | * Return the next group which needs to be prefetched. | |
2224 | */ | |
3d392b26 TT |
2225 | ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
2226 | unsigned int nr, int *cnt) | |
cfd73237 AZ |
2227 | { |
2228 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2229 | struct buffer_head *bh; | |
2230 | struct blk_plug plug; | |
2231 | ||
2232 | blk_start_plug(&plug); | |
2233 | while (nr-- > 0) { | |
2234 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2235 | NULL); | |
2236 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2237 | ||
2238 | /* | |
2239 | * Prefetch block groups with free blocks; but don't | |
2240 | * bother if it is marked uninitialized on disk, since | |
2241 | * it won't require I/O to read. Also only try to | |
2242 | * prefetch once, so we avoid getblk() call, which can | |
2243 | * be expensive. | |
2244 | */ | |
2245 | if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) && | |
2246 | EXT4_MB_GRP_NEED_INIT(grp) && | |
2247 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2248 | !(ext4_has_group_desc_csum(sb) && | |
2249 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2250 | bh = ext4_read_block_bitmap_nowait(sb, group, true); | |
2251 | if (bh && !IS_ERR(bh)) { | |
2252 | if (!buffer_uptodate(bh) && cnt) | |
2253 | (*cnt)++; | |
2254 | brelse(bh); | |
2255 | } | |
2256 | } | |
2257 | if (++group >= ngroups) | |
2258 | group = 0; | |
2259 | } | |
2260 | blk_finish_plug(&plug); | |
2261 | return group; | |
2262 | } | |
2263 | ||
2264 | /* | |
2265 | * Prefetching reads the block bitmap into the buffer cache; but we | |
2266 | * need to make sure that the buddy bitmap in the page cache has been | |
2267 | * initialized. Note that ext4_mb_init_group() will block if the I/O | |
2268 | * is not yet completed, or indeed if it was not initiated by | |
2269 | * ext4_mb_prefetch did not start the I/O. | |
2270 | * | |
2271 | * TODO: We should actually kick off the buddy bitmap setup in a work | |
2272 | * queue when the buffer I/O is completed, so that we don't block | |
2273 | * waiting for the block allocation bitmap read to finish when | |
2274 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). | |
2275 | */ | |
3d392b26 TT |
2276 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
2277 | unsigned int nr) | |
cfd73237 AZ |
2278 | { |
2279 | while (nr-- > 0) { | |
2280 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2281 | NULL); | |
2282 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2283 | ||
2284 | if (!group) | |
2285 | group = ext4_get_groups_count(sb); | |
2286 | group--; | |
2287 | grp = ext4_get_group_info(sb, group); | |
2288 | ||
2289 | if (EXT4_MB_GRP_NEED_INIT(grp) && | |
2290 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2291 | !(ext4_has_group_desc_csum(sb) && | |
2292 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2293 | if (ext4_mb_init_group(sb, group, GFP_NOFS)) | |
2294 | break; | |
2295 | } | |
2296 | } | |
2297 | } | |
2298 | ||
4ddfef7b ES |
2299 | static noinline_for_stack int |
2300 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2301 | { |
cfd73237 | 2302 | ext4_group_t prefetch_grp = 0, ngroups, group, i; |
bbc4ec77 | 2303 | int cr = -1; |
42ac1848 | 2304 | int err = 0, first_err = 0; |
cfd73237 | 2305 | unsigned int nr = 0, prefetch_ios = 0; |
c9de560d AT |
2306 | struct ext4_sb_info *sbi; |
2307 | struct super_block *sb; | |
2308 | struct ext4_buddy e4b; | |
66d5e027 | 2309 | int lost; |
c9de560d AT |
2310 | |
2311 | sb = ac->ac_sb; | |
2312 | sbi = EXT4_SB(sb); | |
8df9675f | 2313 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2314 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2315 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2316 | ngroups = sbi->s_blockfile_groups; |
2317 | ||
c9de560d AT |
2318 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2319 | ||
2320 | /* first, try the goal */ | |
2321 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2322 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2323 | goto out; | |
2324 | ||
2325 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2326 | goto out; | |
2327 | ||
2328 | /* | |
e9a3cd48 | 2329 | * ac->ac_2order is set only if the fe_len is a power of 2 |
2330 | * if ac->ac_2order is set we also set criteria to 0 so that we | |
c9de560d AT |
2331 | * try exact allocation using buddy. |
2332 | */ | |
2333 | i = fls(ac->ac_g_ex.fe_len); | |
2334 | ac->ac_2order = 0; | |
2335 | /* | |
2336 | * We search using buddy data only if the order of the request | |
2337 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2338 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2339 | * We also support searching for power-of-two requests only for |
2340 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2341 | */ |
d9b22cf9 | 2342 | if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) { |
c9de560d AT |
2343 | /* |
2344 | * This should tell if fe_len is exactly power of 2 | |
2345 | */ | |
2346 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d JC |
2347 | ac->ac_2order = array_index_nospec(i - 1, |
2348 | sb->s_blocksize_bits + 2); | |
c9de560d AT |
2349 | } |
2350 | ||
4ba74d00 TT |
2351 | /* if stream allocation is enabled, use global goal */ |
2352 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2353 | /* TBD: may be hot point */ |
2354 | spin_lock(&sbi->s_md_lock); | |
2355 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2356 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2357 | spin_unlock(&sbi->s_md_lock); | |
2358 | } | |
4ba74d00 | 2359 | |
c9de560d AT |
2360 | /* Let's just scan groups to find more-less suitable blocks */ |
2361 | cr = ac->ac_2order ? 0 : 1; | |
2362 | /* | |
2363 | * cr == 0 try to get exact allocation, | |
2364 | * cr == 3 try to get anything | |
2365 | */ | |
2366 | repeat: | |
2367 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2368 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2369 | /* |
2370 | * searching for the right group start | |
2371 | * from the goal value specified | |
2372 | */ | |
2373 | group = ac->ac_g_ex.fe_group; | |
cfd73237 | 2374 | prefetch_grp = group; |
ed8f9c75 | 2375 | |
8df9675f | 2376 | for (i = 0; i < ngroups; group++, i++) { |
42ac1848 | 2377 | int ret = 0; |
2ed5724d | 2378 | cond_resched(); |
e6155736 LM |
2379 | /* |
2380 | * Artificially restricted ngroups for non-extent | |
2381 | * files makes group > ngroups possible on first loop. | |
2382 | */ | |
2383 | if (group >= ngroups) | |
c9de560d AT |
2384 | group = 0; |
2385 | ||
cfd73237 AZ |
2386 | /* |
2387 | * Batch reads of the block allocation bitmaps | |
2388 | * to get multiple READs in flight; limit | |
2389 | * prefetching at cr=0/1, otherwise mballoc can | |
2390 | * spend a lot of time loading imperfect groups | |
2391 | */ | |
2392 | if ((prefetch_grp == group) && | |
2393 | (cr > 1 || | |
2394 | prefetch_ios < sbi->s_mb_prefetch_limit)) { | |
2395 | unsigned int curr_ios = prefetch_ios; | |
2396 | ||
2397 | nr = sbi->s_mb_prefetch; | |
2398 | if (ext4_has_feature_flex_bg(sb)) { | |
2399 | nr = (group / sbi->s_mb_prefetch) * | |
2400 | sbi->s_mb_prefetch; | |
2401 | nr = nr + sbi->s_mb_prefetch - group; | |
2402 | } | |
2403 | prefetch_grp = ext4_mb_prefetch(sb, group, | |
2404 | nr, &prefetch_ios); | |
2405 | if (prefetch_ios == curr_ios) | |
2406 | nr = 0; | |
2407 | } | |
2408 | ||
8a57d9d6 | 2409 | /* This now checks without needing the buddy page */ |
8ef123fe | 2410 | ret = ext4_mb_good_group_nolock(ac, group, cr); |
42ac1848 LC |
2411 | if (ret <= 0) { |
2412 | if (!first_err) | |
2413 | first_err = ret; | |
c9de560d | 2414 | continue; |
42ac1848 | 2415 | } |
c9de560d | 2416 | |
c9de560d AT |
2417 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2418 | if (err) | |
2419 | goto out; | |
2420 | ||
2421 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2422 | |
2423 | /* | |
2424 | * We need to check again after locking the | |
2425 | * block group | |
2426 | */ | |
42ac1848 | 2427 | ret = ext4_mb_good_group(ac, group, cr); |
8ef123fe | 2428 | if (ret == 0) { |
c9de560d | 2429 | ext4_unlock_group(sb, group); |
e39e07fd | 2430 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2431 | continue; |
2432 | } | |
2433 | ||
2434 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2435 | if (cr == 0) |
c9de560d | 2436 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2437 | else if (cr == 1 && sbi->s_stripe && |
2438 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2439 | ext4_mb_scan_aligned(ac, &e4b); |
2440 | else | |
2441 | ext4_mb_complex_scan_group(ac, &e4b); | |
2442 | ||
2443 | ext4_unlock_group(sb, group); | |
e39e07fd | 2444 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2445 | |
2446 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2447 | break; | |
2448 | } | |
2449 | } | |
2450 | ||
2451 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2452 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2453 | /* | |
2454 | * We've been searching too long. Let's try to allocate | |
2455 | * the best chunk we've found so far | |
2456 | */ | |
c9de560d AT |
2457 | ext4_mb_try_best_found(ac, &e4b); |
2458 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2459 | /* | |
2460 | * Someone more lucky has already allocated it. | |
2461 | * The only thing we can do is just take first | |
2462 | * found block(s) | |
c9de560d | 2463 | */ |
66d5e027 | 2464 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
2465 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", | |
c55ee7d2 | 2466 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
2467 | ac->ac_b_ex.fe_len, lost); | |
2468 | ||
c9de560d AT |
2469 | ac->ac_b_ex.fe_group = 0; |
2470 | ac->ac_b_ex.fe_start = 0; | |
2471 | ac->ac_b_ex.fe_len = 0; | |
2472 | ac->ac_status = AC_STATUS_CONTINUE; | |
2473 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2474 | cr = 3; | |
c9de560d AT |
2475 | goto repeat; |
2476 | } | |
2477 | } | |
2478 | out: | |
42ac1848 LC |
2479 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2480 | err = first_err; | |
bbc4ec77 | 2481 | |
d3df1453 | 2482 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
bbc4ec77 RH |
2483 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
2484 | ac->ac_flags, cr, err); | |
cfd73237 AZ |
2485 | |
2486 | if (nr) | |
2487 | ext4_mb_prefetch_fini(sb, prefetch_grp, nr); | |
2488 | ||
c9de560d AT |
2489 | return err; |
2490 | } | |
2491 | ||
c9de560d AT |
2492 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2493 | { | |
247dbed8 | 2494 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2495 | ext4_group_t group; |
2496 | ||
8df9675f | 2497 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2498 | return NULL; |
c9de560d | 2499 | group = *pos + 1; |
a9df9a49 | 2500 | return (void *) ((unsigned long) group); |
c9de560d AT |
2501 | } |
2502 | ||
2503 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2504 | { | |
247dbed8 | 2505 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2506 | ext4_group_t group; |
2507 | ||
2508 | ++*pos; | |
8df9675f | 2509 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2510 | return NULL; |
2511 | group = *pos + 1; | |
a9df9a49 | 2512 | return (void *) ((unsigned long) group); |
c9de560d AT |
2513 | } |
2514 | ||
2515 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2516 | { | |
247dbed8 | 2517 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
a9df9a49 | 2518 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2519 | int i; |
1c8457ca | 2520 | int err, buddy_loaded = 0; |
c9de560d | 2521 | struct ext4_buddy e4b; |
1c8457ca | 2522 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2523 | unsigned char blocksize_bits = min_t(unsigned char, |
2524 | sb->s_blocksize_bits, | |
2525 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2526 | struct sg { |
2527 | struct ext4_group_info info; | |
b80b32b6 | 2528 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2529 | } sg; |
2530 | ||
2531 | group--; | |
2532 | if (group == 0) | |
97b4af2f RV |
2533 | seq_puts(seq, "#group: free frags first [" |
2534 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2535 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2536 | |
b80b32b6 TT |
2537 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2538 | sizeof(struct ext4_group_info); | |
2539 | ||
1c8457ca AK |
2540 | grinfo = ext4_get_group_info(sb, group); |
2541 | /* Load the group info in memory only if not already loaded. */ | |
2542 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2543 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2544 | if (err) { | |
2545 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2546 | return 0; | |
2547 | } | |
2548 | buddy_loaded = 1; | |
c9de560d | 2549 | } |
1c8457ca | 2550 | |
b80b32b6 | 2551 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2552 | |
2553 | if (buddy_loaded) | |
2554 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2555 | |
a9df9a49 | 2556 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2557 | sg.info.bb_fragments, sg.info.bb_first_free); |
2558 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2559 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d | 2560 | sg.info.bb_counters[i] : 0); |
e0d438c7 | 2561 | seq_puts(seq, " ]\n"); |
c9de560d AT |
2562 | |
2563 | return 0; | |
2564 | } | |
2565 | ||
2566 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2567 | { | |
2568 | } | |
2569 | ||
247dbed8 | 2570 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2571 | .start = ext4_mb_seq_groups_start, |
2572 | .next = ext4_mb_seq_groups_next, | |
2573 | .stop = ext4_mb_seq_groups_stop, | |
2574 | .show = ext4_mb_seq_groups_show, | |
2575 | }; | |
2576 | ||
fb1813f4 CW |
2577 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2578 | { | |
2579 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2580 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2581 | ||
2582 | BUG_ON(!cachep); | |
2583 | return cachep; | |
2584 | } | |
5f21b0e6 | 2585 | |
28623c2f TT |
2586 | /* |
2587 | * Allocate the top-level s_group_info array for the specified number | |
2588 | * of groups | |
2589 | */ | |
2590 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2591 | { | |
2592 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2593 | unsigned size; | |
df3da4ea | 2594 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
2595 | |
2596 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2597 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2598 | if (size <= sbi->s_group_info_size) | |
2599 | return 0; | |
2600 | ||
2601 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 2602 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
2603 | if (!new_groupinfo) { |
2604 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2605 | return -ENOMEM; | |
2606 | } | |
df3da4ea SJS |
2607 | rcu_read_lock(); |
2608 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
2609 | if (old_groupinfo) | |
2610 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 2611 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
2612 | rcu_read_unlock(); |
2613 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 2614 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
2615 | if (old_groupinfo) |
2616 | ext4_kvfree_array_rcu(old_groupinfo); | |
28623c2f TT |
2617 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
2618 | sbi->s_group_info_size); | |
2619 | return 0; | |
2620 | } | |
2621 | ||
5f21b0e6 | 2622 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2623 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2624 | struct ext4_group_desc *desc) |
2625 | { | |
fb1813f4 | 2626 | int i; |
5f21b0e6 | 2627 | int metalen = 0; |
df3da4ea | 2628 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
2629 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2630 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2631 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2632 | |
2633 | /* | |
2634 | * First check if this group is the first of a reserved block. | |
2635 | * If it's true, we have to allocate a new table of pointers | |
2636 | * to ext4_group_info structures | |
2637 | */ | |
2638 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2639 | metalen = sizeof(*meta_group_info) << | |
2640 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 2641 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 2642 | if (meta_group_info == NULL) { |
7f6a11e7 | 2643 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2644 | "for a buddy group"); |
5f21b0e6 FB |
2645 | goto exit_meta_group_info; |
2646 | } | |
df3da4ea SJS |
2647 | rcu_read_lock(); |
2648 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
2649 | rcu_read_unlock(); | |
5f21b0e6 FB |
2650 | } |
2651 | ||
df3da4ea | 2652 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
2653 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
2654 | ||
4fdb5543 | 2655 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 2656 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2657 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2658 | goto exit_group_info; |
2659 | } | |
2660 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2661 | &(meta_group_info[i]->bb_state)); | |
2662 | ||
2663 | /* | |
2664 | * initialize bb_free to be able to skip | |
2665 | * empty groups without initialization | |
2666 | */ | |
8844618d TT |
2667 | if (ext4_has_group_desc_csum(sb) && |
2668 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 2669 | meta_group_info[i]->bb_free = |
cff1dfd7 | 2670 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2671 | } else { |
2672 | meta_group_info[i]->bb_free = | |
021b65bb | 2673 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2674 | } |
2675 | ||
2676 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2677 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2678 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2679 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 | 2680 | |
a3450215 | 2681 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
2682 | return 0; |
2683 | ||
2684 | exit_group_info: | |
2685 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2686 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
2687 | struct ext4_group_info ***group_info; |
2688 | ||
2689 | rcu_read_lock(); | |
2690 | group_info = rcu_dereference(sbi->s_group_info); | |
2691 | kfree(group_info[idx]); | |
2692 | group_info[idx] = NULL; | |
2693 | rcu_read_unlock(); | |
caaf7a29 | 2694 | } |
5f21b0e6 FB |
2695 | exit_meta_group_info: |
2696 | return -ENOMEM; | |
2697 | } /* ext4_mb_add_groupinfo */ | |
2698 | ||
c9de560d AT |
2699 | static int ext4_mb_init_backend(struct super_block *sb) |
2700 | { | |
8df9675f | 2701 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2702 | ext4_group_t i; |
c9de560d | 2703 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2704 | int err; |
5f21b0e6 | 2705 | struct ext4_group_desc *desc; |
df3da4ea | 2706 | struct ext4_group_info ***group_info; |
fb1813f4 | 2707 | struct kmem_cache *cachep; |
5f21b0e6 | 2708 | |
28623c2f TT |
2709 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2710 | if (err) | |
2711 | return err; | |
c9de560d | 2712 | |
c9de560d AT |
2713 | sbi->s_buddy_cache = new_inode(sb); |
2714 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2715 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2716 | goto err_freesgi; |
2717 | } | |
48e6061b YJ |
2718 | /* To avoid potentially colliding with an valid on-disk inode number, |
2719 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2720 | * not in the inode hash, so it should never be found by iget(), but | |
2721 | * this will avoid confusion if it ever shows up during debugging. */ | |
2722 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2723 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2724 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 2725 | cond_resched(); |
c9de560d AT |
2726 | desc = ext4_get_group_desc(sb, i, NULL); |
2727 | if (desc == NULL) { | |
9d8b9ec4 | 2728 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2729 | goto err_freebuddy; |
2730 | } | |
5f21b0e6 FB |
2731 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2732 | goto err_freebuddy; | |
c9de560d AT |
2733 | } |
2734 | ||
cfd73237 AZ |
2735 | if (ext4_has_feature_flex_bg(sb)) { |
2736 | /* a single flex group is supposed to be read by a single IO */ | |
2737 | sbi->s_mb_prefetch = 1 << sbi->s_es->s_log_groups_per_flex; | |
2738 | sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ | |
2739 | } else { | |
2740 | sbi->s_mb_prefetch = 32; | |
2741 | } | |
2742 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) | |
2743 | sbi->s_mb_prefetch = ext4_get_groups_count(sb); | |
2744 | /* now many real IOs to prefetch within a single allocation at cr=0 | |
2745 | * given cr=0 is an CPU-related optimization we shouldn't try to | |
2746 | * load too many groups, at some point we should start to use what | |
2747 | * we've got in memory. | |
2748 | * with an average random access time 5ms, it'd take a second to get | |
2749 | * 200 groups (* N with flex_bg), so let's make this limit 4 | |
2750 | */ | |
2751 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; | |
2752 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) | |
2753 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); | |
2754 | ||
c9de560d AT |
2755 | return 0; |
2756 | ||
2757 | err_freebuddy: | |
fb1813f4 | 2758 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2759 | while (i-- > 0) |
fb1813f4 | 2760 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2761 | i = sbi->s_group_info_size; |
df3da4ea SJS |
2762 | rcu_read_lock(); |
2763 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 2764 | while (i-- > 0) |
df3da4ea SJS |
2765 | kfree(group_info[i]); |
2766 | rcu_read_unlock(); | |
c9de560d AT |
2767 | iput(sbi->s_buddy_cache); |
2768 | err_freesgi: | |
df3da4ea SJS |
2769 | rcu_read_lock(); |
2770 | kvfree(rcu_dereference(sbi->s_group_info)); | |
2771 | rcu_read_unlock(); | |
c9de560d AT |
2772 | return -ENOMEM; |
2773 | } | |
2774 | ||
2892c15d ES |
2775 | static void ext4_groupinfo_destroy_slabs(void) |
2776 | { | |
2777 | int i; | |
2778 | ||
2779 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 2780 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
2781 | ext4_groupinfo_caches[i] = NULL; |
2782 | } | |
2783 | } | |
2784 | ||
2785 | static int ext4_groupinfo_create_slab(size_t size) | |
2786 | { | |
2787 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2788 | int slab_size; | |
2789 | int blocksize_bits = order_base_2(size); | |
2790 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2791 | struct kmem_cache *cachep; | |
2792 | ||
2793 | if (cache_index >= NR_GRPINFO_CACHES) | |
2794 | return -EINVAL; | |
2795 | ||
2796 | if (unlikely(cache_index < 0)) | |
2797 | cache_index = 0; | |
2798 | ||
2799 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2800 | if (ext4_groupinfo_caches[cache_index]) { | |
2801 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2802 | return 0; /* Already created */ | |
2803 | } | |
2804 | ||
2805 | slab_size = offsetof(struct ext4_group_info, | |
2806 | bb_counters[blocksize_bits + 2]); | |
2807 | ||
2808 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2809 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2810 | NULL); | |
2811 | ||
823ba01f TM |
2812 | ext4_groupinfo_caches[cache_index] = cachep; |
2813 | ||
2892c15d ES |
2814 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2815 | if (!cachep) { | |
9d8b9ec4 TT |
2816 | printk(KERN_EMERG |
2817 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2818 | return -ENOMEM; |
2819 | } | |
2820 | ||
2892c15d ES |
2821 | return 0; |
2822 | } | |
2823 | ||
9d99012f | 2824 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2825 | { |
2826 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2827 | unsigned i, j; |
935244cd | 2828 | unsigned offset, offset_incr; |
c9de560d | 2829 | unsigned max; |
74767c5a | 2830 | int ret; |
c9de560d | 2831 | |
1927805e | 2832 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2833 | |
2834 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2835 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2836 | ret = -ENOMEM; |
2837 | goto out; | |
c9de560d | 2838 | } |
ff7ef329 | 2839 | |
1927805e | 2840 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2841 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2842 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2843 | ret = -ENOMEM; |
2844 | goto out; | |
2845 | } | |
2846 | ||
2892c15d ES |
2847 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2848 | if (ret < 0) | |
2849 | goto out; | |
c9de560d AT |
2850 | |
2851 | /* order 0 is regular bitmap */ | |
2852 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2853 | sbi->s_mb_offsets[0] = 0; | |
2854 | ||
2855 | i = 1; | |
2856 | offset = 0; | |
935244cd | 2857 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
2858 | max = sb->s_blocksize << 2; |
2859 | do { | |
2860 | sbi->s_mb_offsets[i] = offset; | |
2861 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
2862 | offset += offset_incr; |
2863 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
2864 | max = max >> 1; |
2865 | i++; | |
2866 | } while (i <= sb->s_blocksize_bits + 1); | |
2867 | ||
c9de560d | 2868 | spin_lock_init(&sbi->s_md_lock); |
c9de560d | 2869 | spin_lock_init(&sbi->s_bal_lock); |
d08854f5 | 2870 | sbi->s_mb_free_pending = 0; |
a0154344 | 2871 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
c9de560d AT |
2872 | |
2873 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2874 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2875 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2876 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2877 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27bc446e | 2878 | sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC; |
27baebb8 TT |
2879 | /* |
2880 | * The default group preallocation is 512, which for 4k block | |
2881 | * sizes translates to 2 megabytes. However for bigalloc file | |
2882 | * systems, this is probably too big (i.e, if the cluster size | |
2883 | * is 1 megabyte, then group preallocation size becomes half a | |
2884 | * gigabyte!). As a default, we will keep a two megabyte | |
2885 | * group pralloc size for cluster sizes up to 64k, and after | |
2886 | * that, we will force a minimum group preallocation size of | |
2887 | * 32 clusters. This translates to 8 megs when the cluster | |
2888 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2889 | * which seems reasonable as a default. | |
2890 | */ | |
2891 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2892 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2893 | /* |
2894 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2895 | * to the lowest multiple of s_stripe which is bigger than | |
2896 | * the s_mb_group_prealloc as determined above. We want | |
2897 | * the preallocation size to be an exact multiple of the | |
2898 | * RAID stripe size so that preallocations don't fragment | |
2899 | * the stripes. | |
2900 | */ | |
2901 | if (sbi->s_stripe > 1) { | |
2902 | sbi->s_mb_group_prealloc = roundup( | |
2903 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2904 | } | |
c9de560d | 2905 | |
730c213c | 2906 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2907 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2908 | ret = -ENOMEM; |
029b10c5 | 2909 | goto out; |
c9de560d | 2910 | } |
730c213c | 2911 | for_each_possible_cpu(i) { |
c9de560d | 2912 | struct ext4_locality_group *lg; |
730c213c | 2913 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2914 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2915 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2916 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2917 | spin_lock_init(&lg->lg_prealloc_lock); |
2918 | } | |
2919 | ||
79a77c5a YJ |
2920 | /* init file for buddy data */ |
2921 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2922 | if (ret != 0) |
2923 | goto out_free_locality_groups; | |
79a77c5a | 2924 | |
7aa0baea TM |
2925 | return 0; |
2926 | ||
2927 | out_free_locality_groups: | |
2928 | free_percpu(sbi->s_locality_groups); | |
2929 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 2930 | out: |
7aa0baea TM |
2931 | kfree(sbi->s_mb_offsets); |
2932 | sbi->s_mb_offsets = NULL; | |
2933 | kfree(sbi->s_mb_maxs); | |
2934 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2935 | return ret; |
c9de560d AT |
2936 | } |
2937 | ||
955ce5f5 | 2938 | /* need to called with the ext4 group lock held */ |
d3df1453 | 2939 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
c9de560d AT |
2940 | { |
2941 | struct ext4_prealloc_space *pa; | |
2942 | struct list_head *cur, *tmp; | |
2943 | int count = 0; | |
2944 | ||
2945 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2946 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2947 | list_del(&pa->pa_group_list); | |
2948 | count++; | |
688f05a0 | 2949 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d | 2950 | } |
d3df1453 | 2951 | return count; |
c9de560d AT |
2952 | } |
2953 | ||
2954 | int ext4_mb_release(struct super_block *sb) | |
2955 | { | |
8df9675f | 2956 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2957 | ext4_group_t i; |
2958 | int num_meta_group_infos; | |
df3da4ea | 2959 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 2960 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 2961 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
d3df1453 | 2962 | int count; |
c9de560d | 2963 | |
c9de560d | 2964 | if (sbi->s_group_info) { |
8df9675f | 2965 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 2966 | cond_resched(); |
c9de560d | 2967 | grinfo = ext4_get_group_info(sb, i); |
a3450215 | 2968 | mb_group_bb_bitmap_free(grinfo); |
c9de560d | 2969 | ext4_lock_group(sb, i); |
d3df1453 RH |
2970 | count = ext4_mb_cleanup_pa(grinfo); |
2971 | if (count) | |
2972 | mb_debug(sb, "mballoc: %d PAs left\n", | |
2973 | count); | |
c9de560d | 2974 | ext4_unlock_group(sb, i); |
fb1813f4 | 2975 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2976 | } |
8df9675f | 2977 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2978 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2979 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
2980 | rcu_read_lock(); |
2981 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 2982 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
2983 | kfree(group_info[i]); |
2984 | kvfree(group_info); | |
2985 | rcu_read_unlock(); | |
c9de560d AT |
2986 | } |
2987 | kfree(sbi->s_mb_offsets); | |
2988 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 2989 | iput(sbi->s_buddy_cache); |
c9de560d | 2990 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
2991 | ext4_msg(sb, KERN_INFO, |
2992 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2993 | atomic_read(&sbi->s_bal_allocated), |
2994 | atomic_read(&sbi->s_bal_reqs), | |
2995 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2996 | ext4_msg(sb, KERN_INFO, |
2997 | "mballoc: %u extents scanned, %u goal hits, " | |
2998 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2999 | atomic_read(&sbi->s_bal_ex_scanned), |
3000 | atomic_read(&sbi->s_bal_goals), | |
3001 | atomic_read(&sbi->s_bal_2orders), | |
3002 | atomic_read(&sbi->s_bal_breaks), | |
3003 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
3004 | ext4_msg(sb, KERN_INFO, |
3005 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 3006 | sbi->s_mb_buddies_generated, |
c9de560d | 3007 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
3008 | ext4_msg(sb, KERN_INFO, |
3009 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
3010 | atomic_read(&sbi->s_mb_preallocated), |
3011 | atomic_read(&sbi->s_mb_discarded)); | |
3012 | } | |
3013 | ||
730c213c | 3014 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
3015 | |
3016 | return 0; | |
3017 | } | |
3018 | ||
77ca6cdf | 3019 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
3020 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
3021 | struct bio **biop) | |
5c521830 | 3022 | { |
5c521830 JZ |
3023 | ext4_fsblk_t discard_block; |
3024 | ||
84130193 TT |
3025 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
3026 | ext4_group_first_block_no(sb, block_group)); | |
3027 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
3028 | trace_ext4_discard_blocks(sb, |
3029 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
3030 | if (biop) { |
3031 | return __blkdev_issue_discard(sb->s_bdev, | |
3032 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
3033 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
3034 | GFP_NOFS, 0, biop); | |
3035 | } else | |
3036 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
3037 | } |
3038 | ||
a0154344 DJ |
3039 | static void ext4_free_data_in_buddy(struct super_block *sb, |
3040 | struct ext4_free_data *entry) | |
c9de560d | 3041 | { |
c9de560d | 3042 | struct ext4_buddy e4b; |
c894058d | 3043 | struct ext4_group_info *db; |
d9f34504 | 3044 | int err, count = 0, count2 = 0; |
c9de560d | 3045 | |
d3df1453 | 3046 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
18aadd47 | 3047 | entry->efd_count, entry->efd_group, entry); |
c9de560d | 3048 | |
18aadd47 BJ |
3049 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
3050 | /* we expect to find existing buddy because it's pinned */ | |
3051 | BUG_ON(err != 0); | |
b90f6870 | 3052 | |
d08854f5 TT |
3053 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
3054 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
3055 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 3056 | |
18aadd47 BJ |
3057 | db = e4b.bd_info; |
3058 | /* there are blocks to put in buddy to make them really free */ | |
3059 | count += entry->efd_count; | |
3060 | count2++; | |
3061 | ext4_lock_group(sb, entry->efd_group); | |
3062 | /* Take it out of per group rb tree */ | |
3063 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
3064 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 3065 | |
18aadd47 BJ |
3066 | /* |
3067 | * Clear the trimmed flag for the group so that the next | |
3068 | * ext4_trim_fs can trim it. | |
3069 | * If the volume is mounted with -o discard, online discard | |
3070 | * is supported and the free blocks will be trimmed online. | |
3071 | */ | |
3072 | if (!test_opt(sb, DISCARD)) | |
3073 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 3074 | |
18aadd47 BJ |
3075 | if (!db->bb_free_root.rb_node) { |
3076 | /* No more items in the per group rb tree | |
3077 | * balance refcounts from ext4_mb_free_metadata() | |
3078 | */ | |
09cbfeaf KS |
3079 | put_page(e4b.bd_buddy_page); |
3080 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 3081 | } |
18aadd47 BJ |
3082 | ext4_unlock_group(sb, entry->efd_group); |
3083 | kmem_cache_free(ext4_free_data_cachep, entry); | |
3084 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 3085 | |
d3df1453 RH |
3086 | mb_debug(sb, "freed %d blocks in %d structures\n", count, |
3087 | count2); | |
c9de560d AT |
3088 | } |
3089 | ||
a0154344 DJ |
3090 | /* |
3091 | * This function is called by the jbd2 layer once the commit has finished, | |
3092 | * so we know we can free the blocks that were released with that commit. | |
3093 | */ | |
3094 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
3095 | { | |
3096 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3097 | struct ext4_free_data *entry, *tmp; | |
3098 | struct bio *discard_bio = NULL; | |
3099 | struct list_head freed_data_list; | |
3100 | struct list_head *cut_pos = NULL; | |
3101 | int err; | |
3102 | ||
3103 | INIT_LIST_HEAD(&freed_data_list); | |
3104 | ||
3105 | spin_lock(&sbi->s_md_lock); | |
3106 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
3107 | if (entry->efd_tid != commit_tid) | |
3108 | break; | |
3109 | cut_pos = &entry->efd_list; | |
3110 | } | |
3111 | if (cut_pos) | |
3112 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
3113 | cut_pos); | |
3114 | spin_unlock(&sbi->s_md_lock); | |
3115 | ||
3116 | if (test_opt(sb, DISCARD)) { | |
3117 | list_for_each_entry(entry, &freed_data_list, efd_list) { | |
3118 | err = ext4_issue_discard(sb, entry->efd_group, | |
3119 | entry->efd_start_cluster, | |
3120 | entry->efd_count, | |
3121 | &discard_bio); | |
3122 | if (err && err != -EOPNOTSUPP) { | |
3123 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
3124 | " group:%d block:%d count:%d failed" | |
3125 | " with %d", entry->efd_group, | |
3126 | entry->efd_start_cluster, | |
3127 | entry->efd_count, err); | |
3128 | } else if (err == -EOPNOTSUPP) | |
3129 | break; | |
3130 | } | |
3131 | ||
e4510577 | 3132 | if (discard_bio) { |
a0154344 | 3133 | submit_bio_wait(discard_bio); |
e4510577 DJ |
3134 | bio_put(discard_bio); |
3135 | } | |
a0154344 DJ |
3136 | } |
3137 | ||
3138 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
3139 | ext4_free_data_in_buddy(sb, entry); | |
3140 | } | |
3141 | ||
5dabfc78 | 3142 | int __init ext4_init_mballoc(void) |
c9de560d | 3143 | { |
16828088 TT |
3144 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
3145 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 3146 | if (ext4_pspace_cachep == NULL) |
f283529a | 3147 | goto out; |
c9de560d | 3148 | |
16828088 TT |
3149 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
3150 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3151 | if (ext4_ac_cachep == NULL) |
3152 | goto out_pa_free; | |
c894058d | 3153 | |
18aadd47 BJ |
3154 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
3155 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3156 | if (ext4_free_data_cachep == NULL) |
3157 | goto out_ac_free; | |
3158 | ||
c9de560d | 3159 | return 0; |
f283529a RH |
3160 | |
3161 | out_ac_free: | |
3162 | kmem_cache_destroy(ext4_ac_cachep); | |
3163 | out_pa_free: | |
3164 | kmem_cache_destroy(ext4_pspace_cachep); | |
3165 | out: | |
3166 | return -ENOMEM; | |
c9de560d AT |
3167 | } |
3168 | ||
5dabfc78 | 3169 | void ext4_exit_mballoc(void) |
c9de560d | 3170 | { |
60e6679e | 3171 | /* |
3e03f9ca JDB |
3172 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
3173 | * before destroying the slab cache. | |
3174 | */ | |
3175 | rcu_barrier(); | |
c9de560d | 3176 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 3177 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 3178 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 3179 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
3180 | } |
3181 | ||
3182 | ||
3183 | /* | |
73b2c716 | 3184 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
3185 | * Returns 0 if success or error code |
3186 | */ | |
4ddfef7b ES |
3187 | static noinline_for_stack int |
3188 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 3189 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
3190 | { |
3191 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
3192 | struct ext4_group_desc *gdp; |
3193 | struct buffer_head *gdp_bh; | |
3194 | struct ext4_sb_info *sbi; | |
3195 | struct super_block *sb; | |
3196 | ext4_fsblk_t block; | |
519deca0 | 3197 | int err, len; |
c9de560d AT |
3198 | |
3199 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3200 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
3201 | ||
3202 | sb = ac->ac_sb; | |
3203 | sbi = EXT4_SB(sb); | |
c9de560d | 3204 | |
574ca174 | 3205 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
3206 | if (IS_ERR(bitmap_bh)) { |
3207 | err = PTR_ERR(bitmap_bh); | |
3208 | bitmap_bh = NULL; | |
c9de560d | 3209 | goto out_err; |
9008a58e | 3210 | } |
c9de560d | 3211 | |
5d601255 | 3212 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
c9de560d AT |
3213 | err = ext4_journal_get_write_access(handle, bitmap_bh); |
3214 | if (err) | |
3215 | goto out_err; | |
3216 | ||
3217 | err = -EIO; | |
3218 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3219 | if (!gdp) | |
3220 | goto out_err; | |
3221 | ||
a9df9a49 | 3222 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3223 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3224 | |
5d601255 | 3225 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
c9de560d AT |
3226 | err = ext4_journal_get_write_access(handle, gdp_bh); |
3227 | if (err) | |
3228 | goto out_err; | |
3229 | ||
bda00de7 | 3230 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3231 | |
53accfa9 | 3232 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
ce9f24cc | 3233 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
12062ddd | 3234 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3235 | "fs metadata", block, block+len); |
519deca0 | 3236 | /* File system mounted not to panic on error |
554a5ccc | 3237 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3238 | * We leak some of the blocks here. |
3239 | */ | |
955ce5f5 | 3240 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
3241 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3242 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 3243 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3244 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3245 | if (!err) |
554a5ccc | 3246 | err = -EFSCORRUPTED; |
519deca0 | 3247 | goto out_err; |
c9de560d | 3248 | } |
955ce5f5 AK |
3249 | |
3250 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3251 | #ifdef AGGRESSIVE_CHECK |
3252 | { | |
3253 | int i; | |
3254 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3255 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3256 | bitmap_bh->b_data)); | |
3257 | } | |
3258 | } | |
3259 | #endif | |
c3e94d1d YY |
3260 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3261 | ac->ac_b_ex.fe_len); | |
8844618d TT |
3262 | if (ext4_has_group_desc_csum(sb) && |
3263 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3264 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3265 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3266 | ext4_free_clusters_after_init(sb, |
021b65bb | 3267 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3268 | } |
021b65bb TT |
3269 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3270 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3271 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3272 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3273 | |
3274 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3275 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3276 | /* |
6bc6e63f | 3277 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3278 | */ |
6bc6e63f AK |
3279 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3280 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3281 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3282 | reserv_clstrs); | |
c9de560d | 3283 | |
772cb7c8 JS |
3284 | if (sbi->s_log_groups_per_flex) { |
3285 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3286 | ac->ac_b_ex.fe_group); | |
90ba983f | 3287 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3288 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3289 | flex_group)->free_clusters); | |
772cb7c8 JS |
3290 | } |
3291 | ||
0390131b | 3292 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3293 | if (err) |
3294 | goto out_err; | |
0390131b | 3295 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3296 | |
3297 | out_err: | |
42a10add | 3298 | brelse(bitmap_bh); |
c9de560d AT |
3299 | return err; |
3300 | } | |
3301 | ||
3302 | /* | |
3303 | * here we normalize request for locality group | |
d7a1fee1 DE |
3304 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3305 | * s_strip if we set the same via mount option. | |
3306 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3307 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3308 | * |
3309 | * XXX: should we try to preallocate more than the group has now? | |
3310 | */ | |
3311 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3312 | { | |
3313 | struct super_block *sb = ac->ac_sb; | |
3314 | struct ext4_locality_group *lg = ac->ac_lg; | |
3315 | ||
3316 | BUG_ON(lg == NULL); | |
d7a1fee1 | 3317 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
d3df1453 | 3318 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
c9de560d AT |
3319 | } |
3320 | ||
3321 | /* | |
3322 | * Normalization means making request better in terms of | |
3323 | * size and alignment | |
3324 | */ | |
4ddfef7b ES |
3325 | static noinline_for_stack void |
3326 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3327 | struct ext4_allocation_request *ar) |
3328 | { | |
53accfa9 | 3329 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3330 | int bsbits, max; |
3331 | ext4_lblk_t end; | |
1592d2c5 CW |
3332 | loff_t size, start_off; |
3333 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3334 | ext4_lblk_t start; |
c9de560d | 3335 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 3336 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3337 | |
3338 | /* do normalize only data requests, metadata requests | |
3339 | do not need preallocation */ | |
3340 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3341 | return; | |
3342 | ||
3343 | /* sometime caller may want exact blocks */ | |
3344 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3345 | return; | |
3346 | ||
3347 | /* caller may indicate that preallocation isn't | |
3348 | * required (it's a tail, for example) */ | |
3349 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3350 | return; | |
3351 | ||
3352 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3353 | ext4_mb_normalize_group_request(ac); | |
3354 | return ; | |
3355 | } | |
3356 | ||
3357 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3358 | ||
3359 | /* first, let's learn actual file size | |
3360 | * given current request is allocated */ | |
53accfa9 | 3361 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
3362 | size = size << bsbits; |
3363 | if (size < i_size_read(ac->ac_inode)) | |
3364 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 3365 | orig_size = size; |
c9de560d | 3366 | |
1930479c VC |
3367 | /* max size of free chunks */ |
3368 | max = 2 << bsbits; | |
c9de560d | 3369 | |
1930479c VC |
3370 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3371 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3372 | |
3373 | /* first, try to predict filesize */ | |
3374 | /* XXX: should this table be tunable? */ | |
3375 | start_off = 0; | |
3376 | if (size <= 16 * 1024) { | |
3377 | size = 16 * 1024; | |
3378 | } else if (size <= 32 * 1024) { | |
3379 | size = 32 * 1024; | |
3380 | } else if (size <= 64 * 1024) { | |
3381 | size = 64 * 1024; | |
3382 | } else if (size <= 128 * 1024) { | |
3383 | size = 128 * 1024; | |
3384 | } else if (size <= 256 * 1024) { | |
3385 | size = 256 * 1024; | |
3386 | } else if (size <= 512 * 1024) { | |
3387 | size = 512 * 1024; | |
3388 | } else if (size <= 1024 * 1024) { | |
3389 | size = 1024 * 1024; | |
1930479c | 3390 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3391 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3392 | (21 - bsbits)) << 21; |
3393 | size = 2 * 1024 * 1024; | |
3394 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3395 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3396 | (22 - bsbits)) << 22; | |
3397 | size = 4 * 1024 * 1024; | |
3398 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3399 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3400 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3401 | (23 - bsbits)) << 23; | |
3402 | size = 8 * 1024 * 1024; | |
3403 | } else { | |
b27b1535 XW |
3404 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
3405 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
3406 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 3407 | } |
5a0790c2 AK |
3408 | size = size >> bsbits; |
3409 | start = start_off >> bsbits; | |
c9de560d AT |
3410 | |
3411 | /* don't cover already allocated blocks in selected range */ | |
3412 | if (ar->pleft && start <= ar->lleft) { | |
3413 | size -= ar->lleft + 1 - start; | |
3414 | start = ar->lleft + 1; | |
3415 | } | |
3416 | if (ar->pright && start + size - 1 >= ar->lright) | |
3417 | size -= start + size - ar->lright; | |
3418 | ||
cd648b8a JK |
3419 | /* |
3420 | * Trim allocation request for filesystems with artificially small | |
3421 | * groups. | |
3422 | */ | |
3423 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
3424 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
3425 | ||
c9de560d AT |
3426 | end = start + size; |
3427 | ||
3428 | /* check we don't cross already preallocated blocks */ | |
3429 | rcu_read_lock(); | |
9a0762c5 | 3430 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3431 | ext4_lblk_t pa_end; |
c9de560d | 3432 | |
c9de560d AT |
3433 | if (pa->pa_deleted) |
3434 | continue; | |
3435 | spin_lock(&pa->pa_lock); | |
3436 | if (pa->pa_deleted) { | |
3437 | spin_unlock(&pa->pa_lock); | |
3438 | continue; | |
3439 | } | |
3440 | ||
53accfa9 TT |
3441 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3442 | pa->pa_len); | |
c9de560d AT |
3443 | |
3444 | /* PA must not overlap original request */ | |
3445 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3446 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3447 | ||
38877f4e ES |
3448 | /* skip PAs this normalized request doesn't overlap with */ |
3449 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3450 | spin_unlock(&pa->pa_lock); |
3451 | continue; | |
3452 | } | |
3453 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3454 | ||
38877f4e | 3455 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3456 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3457 | BUG_ON(pa_end < start); | |
3458 | start = pa_end; | |
38877f4e | 3459 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3460 | BUG_ON(pa->pa_lstart > end); |
3461 | end = pa->pa_lstart; | |
3462 | } | |
3463 | spin_unlock(&pa->pa_lock); | |
3464 | } | |
3465 | rcu_read_unlock(); | |
3466 | size = end - start; | |
3467 | ||
3468 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3469 | rcu_read_lock(); | |
9a0762c5 | 3470 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3471 | ext4_lblk_t pa_end; |
53accfa9 | 3472 | |
c9de560d AT |
3473 | spin_lock(&pa->pa_lock); |
3474 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3475 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3476 | pa->pa_len); | |
c9de560d AT |
3477 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3478 | } | |
3479 | spin_unlock(&pa->pa_lock); | |
3480 | } | |
3481 | rcu_read_unlock(); | |
3482 | ||
3483 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3484 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3485 | ext4_msg(ac->ac_sb, KERN_ERR, |
3486 | "start %lu, size %lu, fe_logical %lu", | |
3487 | (unsigned long) start, (unsigned long) size, | |
3488 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 3489 | BUG(); |
c9de560d | 3490 | } |
b5b60778 | 3491 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3492 | |
3493 | /* now prepare goal request */ | |
3494 | ||
3495 | /* XXX: is it better to align blocks WRT to logical | |
3496 | * placement or satisfy big request as is */ | |
3497 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3498 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3499 | |
3500 | /* define goal start in order to merge */ | |
3501 | if (ar->pright && (ar->lright == (start + size))) { | |
3502 | /* merge to the right */ | |
3503 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3504 | &ac->ac_f_ex.fe_group, | |
3505 | &ac->ac_f_ex.fe_start); | |
3506 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3507 | } | |
3508 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3509 | /* merge to the left */ | |
3510 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3511 | &ac->ac_f_ex.fe_group, | |
3512 | &ac->ac_f_ex.fe_start); | |
3513 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3514 | } | |
3515 | ||
d3df1453 RH |
3516 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
3517 | orig_size, start); | |
c9de560d AT |
3518 | } |
3519 | ||
3520 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3521 | { | |
3522 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3523 | ||
3524 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3525 | atomic_inc(&sbi->s_bal_reqs); | |
3526 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3527 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3528 | atomic_inc(&sbi->s_bal_success); |
3529 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3530 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3531 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3532 | atomic_inc(&sbi->s_bal_goals); | |
3533 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3534 | atomic_inc(&sbi->s_bal_breaks); | |
3535 | } | |
3536 | ||
296c355c TT |
3537 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3538 | trace_ext4_mballoc_alloc(ac); | |
3539 | else | |
3540 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3541 | } |
3542 | ||
b844167e CW |
3543 | /* |
3544 | * Called on failure; free up any blocks from the inode PA for this | |
3545 | * context. We don't need this for MB_GROUP_PA because we only change | |
3546 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3547 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3548 | */ | |
3549 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3550 | { | |
3551 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
3552 | struct ext4_buddy e4b; |
3553 | int err; | |
b844167e | 3554 | |
86f0afd4 | 3555 | if (pa == NULL) { |
c99d1e6e TT |
3556 | if (ac->ac_f_ex.fe_len == 0) |
3557 | return; | |
86f0afd4 TT |
3558 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
3559 | if (err) { | |
3560 | /* | |
3561 | * This should never happen since we pin the | |
3562 | * pages in the ext4_allocation_context so | |
3563 | * ext4_mb_load_buddy() should never fail. | |
3564 | */ | |
3565 | WARN(1, "mb_load_buddy failed (%d)", err); | |
3566 | return; | |
3567 | } | |
3568 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
3569 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
3570 | ac->ac_f_ex.fe_len); | |
3571 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 3572 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
3573 | return; |
3574 | } | |
3575 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 3576 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
3577 | } |
3578 | ||
c9de560d AT |
3579 | /* |
3580 | * use blocks preallocated to inode | |
3581 | */ | |
3582 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3583 | struct ext4_prealloc_space *pa) | |
3584 | { | |
53accfa9 | 3585 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3586 | ext4_fsblk_t start; |
3587 | ext4_fsblk_t end; | |
3588 | int len; | |
3589 | ||
3590 | /* found preallocated blocks, use them */ | |
3591 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3592 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3593 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3594 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3595 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3596 | &ac->ac_b_ex.fe_start); | |
3597 | ac->ac_b_ex.fe_len = len; | |
3598 | ac->ac_status = AC_STATUS_FOUND; | |
3599 | ac->ac_pa = pa; | |
3600 | ||
3601 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3602 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3603 | BUG_ON(pa->pa_free < len); |
3604 | pa->pa_free -= len; | |
3605 | ||
d3df1453 | 3606 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
3607 | } |
3608 | ||
3609 | /* | |
3610 | * use blocks preallocated to locality group | |
3611 | */ | |
3612 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3613 | struct ext4_prealloc_space *pa) | |
3614 | { | |
03cddb80 | 3615 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3616 | |
c9de560d AT |
3617 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3618 | &ac->ac_b_ex.fe_group, | |
3619 | &ac->ac_b_ex.fe_start); | |
3620 | ac->ac_b_ex.fe_len = len; | |
3621 | ac->ac_status = AC_STATUS_FOUND; | |
3622 | ac->ac_pa = pa; | |
3623 | ||
3624 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3625 | * possible race when the group is being loaded concurrently |
c9de560d | 3626 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3627 | * in on-disk bitmap -- see ext4_mb_release_context() |
3628 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3629 | */ |
d3df1453 RH |
3630 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
3631 | pa->pa_lstart-len, len, pa); | |
c9de560d AT |
3632 | } |
3633 | ||
5e745b04 AK |
3634 | /* |
3635 | * Return the prealloc space that have minimal distance | |
3636 | * from the goal block. @cpa is the prealloc | |
3637 | * space that is having currently known minimal distance | |
3638 | * from the goal block. | |
3639 | */ | |
3640 | static struct ext4_prealloc_space * | |
3641 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3642 | struct ext4_prealloc_space *pa, | |
3643 | struct ext4_prealloc_space *cpa) | |
3644 | { | |
3645 | ext4_fsblk_t cur_distance, new_distance; | |
3646 | ||
3647 | if (cpa == NULL) { | |
3648 | atomic_inc(&pa->pa_count); | |
3649 | return pa; | |
3650 | } | |
79211c8e AM |
3651 | cur_distance = abs(goal_block - cpa->pa_pstart); |
3652 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 3653 | |
5a54b2f1 | 3654 | if (cur_distance <= new_distance) |
5e745b04 AK |
3655 | return cpa; |
3656 | ||
3657 | /* drop the previous reference */ | |
3658 | atomic_dec(&cpa->pa_count); | |
3659 | atomic_inc(&pa->pa_count); | |
3660 | return pa; | |
3661 | } | |
3662 | ||
c9de560d AT |
3663 | /* |
3664 | * search goal blocks in preallocated space | |
3665 | */ | |
4fca8f07 | 3666 | static noinline_for_stack bool |
4ddfef7b | 3667 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 3668 | { |
53accfa9 | 3669 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3670 | int order, i; |
c9de560d AT |
3671 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3672 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3673 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3674 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3675 | |
3676 | /* only data can be preallocated */ | |
3677 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 3678 | return false; |
c9de560d AT |
3679 | |
3680 | /* first, try per-file preallocation */ | |
3681 | rcu_read_lock(); | |
9a0762c5 | 3682 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3683 | |
3684 | /* all fields in this condition don't change, | |
3685 | * so we can skip locking for them */ | |
3686 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3687 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3688 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3689 | continue; |
3690 | ||
fb0a387d | 3691 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3692 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3693 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3694 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3695 | continue; |
3696 | ||
c9de560d AT |
3697 | /* found preallocated blocks, use them */ |
3698 | spin_lock(&pa->pa_lock); | |
3699 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3700 | atomic_inc(&pa->pa_count); | |
3701 | ext4_mb_use_inode_pa(ac, pa); | |
3702 | spin_unlock(&pa->pa_lock); | |
3703 | ac->ac_criteria = 10; | |
3704 | rcu_read_unlock(); | |
4fca8f07 | 3705 | return true; |
c9de560d AT |
3706 | } |
3707 | spin_unlock(&pa->pa_lock); | |
3708 | } | |
3709 | rcu_read_unlock(); | |
3710 | ||
3711 | /* can we use group allocation? */ | |
3712 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 3713 | return false; |
c9de560d AT |
3714 | |
3715 | /* inode may have no locality group for some reason */ | |
3716 | lg = ac->ac_lg; | |
3717 | if (lg == NULL) | |
4fca8f07 | 3718 | return false; |
6be2ded1 AK |
3719 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3720 | if (order > PREALLOC_TB_SIZE - 1) | |
3721 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3722 | order = PREALLOC_TB_SIZE - 1; | |
3723 | ||
bda00de7 | 3724 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3725 | /* |
3726 | * search for the prealloc space that is having | |
3727 | * minimal distance from the goal block. | |
3728 | */ | |
6be2ded1 AK |
3729 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3730 | rcu_read_lock(); | |
3731 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3732 | pa_inode_list) { | |
3733 | spin_lock(&pa->pa_lock); | |
3734 | if (pa->pa_deleted == 0 && | |
3735 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3736 | |
3737 | cpa = ext4_mb_check_group_pa(goal_block, | |
3738 | pa, cpa); | |
6be2ded1 | 3739 | } |
c9de560d | 3740 | spin_unlock(&pa->pa_lock); |
c9de560d | 3741 | } |
6be2ded1 | 3742 | rcu_read_unlock(); |
c9de560d | 3743 | } |
5e745b04 AK |
3744 | if (cpa) { |
3745 | ext4_mb_use_group_pa(ac, cpa); | |
3746 | ac->ac_criteria = 20; | |
4fca8f07 | 3747 | return true; |
5e745b04 | 3748 | } |
4fca8f07 | 3749 | return false; |
c9de560d AT |
3750 | } |
3751 | ||
7a2fcbf7 AK |
3752 | /* |
3753 | * the function goes through all block freed in the group | |
3754 | * but not yet committed and marks them used in in-core bitmap. | |
3755 | * buddy must be generated from this bitmap | |
955ce5f5 | 3756 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3757 | */ |
3758 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3759 | ext4_group_t group) | |
3760 | { | |
3761 | struct rb_node *n; | |
3762 | struct ext4_group_info *grp; | |
3763 | struct ext4_free_data *entry; | |
3764 | ||
3765 | grp = ext4_get_group_info(sb, group); | |
3766 | n = rb_first(&(grp->bb_free_root)); | |
3767 | ||
3768 | while (n) { | |
18aadd47 BJ |
3769 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3770 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3771 | n = rb_next(n); |
3772 | } | |
3773 | return; | |
3774 | } | |
3775 | ||
c9de560d AT |
3776 | /* |
3777 | * the function goes through all preallocation in this group and marks them | |
3778 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3779 | * Need to be called with ext4 group lock held |
c9de560d | 3780 | */ |
089ceecc ES |
3781 | static noinline_for_stack |
3782 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3783 | ext4_group_t group) |
3784 | { | |
3785 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3786 | struct ext4_prealloc_space *pa; | |
3787 | struct list_head *cur; | |
3788 | ext4_group_t groupnr; | |
3789 | ext4_grpblk_t start; | |
3790 | int preallocated = 0; | |
c9de560d AT |
3791 | int len; |
3792 | ||
3793 | /* all form of preallocation discards first load group, | |
3794 | * so the only competing code is preallocation use. | |
3795 | * we don't need any locking here | |
3796 | * notice we do NOT ignore preallocations with pa_deleted | |
3797 | * otherwise we could leave used blocks available for | |
3798 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3799 | * is dropping preallocation | |
3800 | */ | |
3801 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3802 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3803 | spin_lock(&pa->pa_lock); | |
3804 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3805 | &groupnr, &start); | |
3806 | len = pa->pa_len; | |
3807 | spin_unlock(&pa->pa_lock); | |
3808 | if (unlikely(len == 0)) | |
3809 | continue; | |
3810 | BUG_ON(groupnr != group); | |
c3e94d1d | 3811 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3812 | preallocated += len; |
c9de560d | 3813 | } |
d3df1453 | 3814 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
3815 | } |
3816 | ||
27bc446e | 3817 | static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
3818 | struct ext4_prealloc_space *pa) | |
3819 | { | |
3820 | struct ext4_inode_info *ei; | |
3821 | ||
3822 | if (pa->pa_deleted) { | |
3823 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", | |
3824 | pa->pa_type, pa->pa_pstart, pa->pa_lstart, | |
3825 | pa->pa_len); | |
3826 | return; | |
3827 | } | |
3828 | ||
3829 | pa->pa_deleted = 1; | |
3830 | ||
3831 | if (pa->pa_type == MB_INODE_PA) { | |
3832 | ei = EXT4_I(pa->pa_inode); | |
3833 | atomic_dec(&ei->i_prealloc_active); | |
3834 | } | |
3835 | } | |
3836 | ||
c9de560d AT |
3837 | static void ext4_mb_pa_callback(struct rcu_head *head) |
3838 | { | |
3839 | struct ext4_prealloc_space *pa; | |
3840 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
3841 | |
3842 | BUG_ON(atomic_read(&pa->pa_count)); | |
3843 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
3844 | kmem_cache_free(ext4_pspace_cachep, pa); |
3845 | } | |
3846 | ||
3847 | /* | |
3848 | * drops a reference to preallocated space descriptor | |
3849 | * if this was the last reference and the space is consumed | |
3850 | */ | |
3851 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3852 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3853 | { | |
a9df9a49 | 3854 | ext4_group_t grp; |
d33a1976 | 3855 | ext4_fsblk_t grp_blk; |
c9de560d | 3856 | |
c9de560d AT |
3857 | /* in this short window concurrent discard can set pa_deleted */ |
3858 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
3859 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
3860 | spin_unlock(&pa->pa_lock); | |
3861 | return; | |
3862 | } | |
3863 | ||
c9de560d AT |
3864 | if (pa->pa_deleted == 1) { |
3865 | spin_unlock(&pa->pa_lock); | |
3866 | return; | |
3867 | } | |
3868 | ||
27bc446e | 3869 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
3870 | spin_unlock(&pa->pa_lock); |
3871 | ||
d33a1976 | 3872 | grp_blk = pa->pa_pstart; |
60e6679e | 3873 | /* |
cc0fb9ad AK |
3874 | * If doing group-based preallocation, pa_pstart may be in the |
3875 | * next group when pa is used up | |
3876 | */ | |
3877 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3878 | grp_blk--; |
3879 | ||
bd86298e | 3880 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
3881 | |
3882 | /* | |
3883 | * possible race: | |
3884 | * | |
3885 | * P1 (buddy init) P2 (regular allocation) | |
3886 | * find block B in PA | |
3887 | * copy on-disk bitmap to buddy | |
3888 | * mark B in on-disk bitmap | |
3889 | * drop PA from group | |
3890 | * mark all PAs in buddy | |
3891 | * | |
3892 | * thus, P1 initializes buddy with B available. to prevent this | |
3893 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3894 | * against that pair | |
3895 | */ | |
3896 | ext4_lock_group(sb, grp); | |
3897 | list_del(&pa->pa_group_list); | |
3898 | ext4_unlock_group(sb, grp); | |
3899 | ||
3900 | spin_lock(pa->pa_obj_lock); | |
3901 | list_del_rcu(&pa->pa_inode_list); | |
3902 | spin_unlock(pa->pa_obj_lock); | |
3903 | ||
3904 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3905 | } | |
3906 | ||
3907 | /* | |
3908 | * creates new preallocated space for given inode | |
3909 | */ | |
53f86b17 | 3910 | static noinline_for_stack void |
4ddfef7b | 3911 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
3912 | { |
3913 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3914 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3915 | struct ext4_prealloc_space *pa; |
3916 | struct ext4_group_info *grp; | |
3917 | struct ext4_inode_info *ei; | |
3918 | ||
3919 | /* preallocate only when found space is larger then requested */ | |
3920 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3921 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3922 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 3923 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 3924 | |
53f86b17 | 3925 | pa = ac->ac_pa; |
c9de560d AT |
3926 | |
3927 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3928 | int winl; | |
3929 | int wins; | |
3930 | int win; | |
3931 | int offs; | |
3932 | ||
3933 | /* we can't allocate as much as normalizer wants. | |
3934 | * so, found space must get proper lstart | |
3935 | * to cover original request */ | |
3936 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3937 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3938 | ||
3939 | /* we're limited by original request in that | |
3940 | * logical block must be covered any way | |
3941 | * winl is window we can move our chunk within */ | |
3942 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3943 | ||
3944 | /* also, we should cover whole original request */ | |
53accfa9 | 3945 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3946 | |
3947 | /* the smallest one defines real window */ | |
3948 | win = min(winl, wins); | |
3949 | ||
53accfa9 TT |
3950 | offs = ac->ac_o_ex.fe_logical % |
3951 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3952 | if (offs && offs < win) |
3953 | win = offs; | |
3954 | ||
53accfa9 | 3955 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3956 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3957 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3958 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3959 | } | |
3960 | ||
3961 | /* preallocation can change ac_b_ex, thus we store actually | |
3962 | * allocated blocks for history */ | |
3963 | ac->ac_f_ex = ac->ac_b_ex; | |
3964 | ||
3965 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3966 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3967 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3968 | pa->pa_free = pa->pa_len; | |
c9de560d | 3969 | spin_lock_init(&pa->pa_lock); |
d794bf8e AK |
3970 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3971 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3972 | pa->pa_deleted = 0; |
cc0fb9ad | 3973 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3974 | |
d3df1453 RH |
3975 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
3976 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 3977 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3978 | |
3979 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3980 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3981 | |
3982 | ei = EXT4_I(ac->ac_inode); | |
3983 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3984 | ||
3985 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3986 | pa->pa_inode = ac->ac_inode; | |
3987 | ||
c9de560d | 3988 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d AT |
3989 | |
3990 | spin_lock(pa->pa_obj_lock); | |
3991 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3992 | spin_unlock(pa->pa_obj_lock); | |
27bc446e | 3993 | atomic_inc(&ei->i_prealloc_active); |
c9de560d AT |
3994 | } |
3995 | ||
3996 | /* | |
3997 | * creates new preallocated space for locality group inodes belongs to | |
3998 | */ | |
53f86b17 | 3999 | static noinline_for_stack void |
4ddfef7b | 4000 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4001 | { |
4002 | struct super_block *sb = ac->ac_sb; | |
4003 | struct ext4_locality_group *lg; | |
4004 | struct ext4_prealloc_space *pa; | |
4005 | struct ext4_group_info *grp; | |
4006 | ||
4007 | /* preallocate only when found space is larger then requested */ | |
4008 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4009 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4010 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4011 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4012 | |
53f86b17 | 4013 | pa = ac->ac_pa; |
c9de560d AT |
4014 | |
4015 | /* preallocation can change ac_b_ex, thus we store actually | |
4016 | * allocated blocks for history */ | |
4017 | ac->ac_f_ex = ac->ac_b_ex; | |
4018 | ||
4019 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4020 | pa->pa_lstart = pa->pa_pstart; | |
4021 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4022 | pa->pa_free = pa->pa_len; | |
c9de560d | 4023 | spin_lock_init(&pa->pa_lock); |
6be2ded1 | 4024 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 4025 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 4026 | pa->pa_deleted = 0; |
cc0fb9ad | 4027 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 4028 | |
d3df1453 RH |
4029 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4030 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4031 | trace_ext4_mb_new_group_pa(ac, pa); |
c9de560d AT |
4032 | |
4033 | ext4_mb_use_group_pa(ac, pa); | |
4034 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
4035 | ||
4036 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4037 | lg = ac->ac_lg; | |
4038 | BUG_ON(lg == NULL); | |
4039 | ||
4040 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
4041 | pa->pa_inode = NULL; | |
4042 | ||
c9de560d | 4043 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 4044 | |
6be2ded1 AK |
4045 | /* |
4046 | * We will later add the new pa to the right bucket | |
4047 | * after updating the pa_free in ext4_mb_release_context | |
4048 | */ | |
c9de560d AT |
4049 | } |
4050 | ||
53f86b17 | 4051 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
c9de560d | 4052 | { |
c9de560d | 4053 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
53f86b17 | 4054 | ext4_mb_new_group_pa(ac); |
c9de560d | 4055 | else |
53f86b17 | 4056 | ext4_mb_new_inode_pa(ac); |
c9de560d AT |
4057 | } |
4058 | ||
4059 | /* | |
4060 | * finds all unused blocks in on-disk bitmap, frees them in | |
4061 | * in-core bitmap and buddy. | |
4062 | * @pa must be unlinked from inode and group lists, so that | |
4063 | * nobody else can find/use it. | |
4064 | * the caller MUST hold group/inode locks. | |
4065 | * TODO: optimize the case when there are no in-core structures yet | |
4066 | */ | |
4ddfef7b ES |
4067 | static noinline_for_stack int |
4068 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 4069 | struct ext4_prealloc_space *pa) |
c9de560d | 4070 | { |
c9de560d AT |
4071 | struct super_block *sb = e4b->bd_sb; |
4072 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
4073 | unsigned int end; |
4074 | unsigned int next; | |
c9de560d AT |
4075 | ext4_group_t group; |
4076 | ext4_grpblk_t bit; | |
ba80b101 | 4077 | unsigned long long grp_blk_start; |
c9de560d AT |
4078 | int free = 0; |
4079 | ||
4080 | BUG_ON(pa->pa_deleted == 0); | |
4081 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 4082 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
4083 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
4084 | end = bit + pa->pa_len; | |
4085 | ||
c9de560d | 4086 | while (bit < end) { |
ffad0a44 | 4087 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
4088 | if (bit >= end) |
4089 | break; | |
ffad0a44 | 4090 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
d3df1453 | 4091 | mb_debug(sb, "free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
4092 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
4093 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
4094 | free += next - bit; |
4095 | ||
3e1e5f50 | 4096 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
4097 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
4098 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 4099 | next - bit); |
c9de560d AT |
4100 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
4101 | bit = next + 1; | |
4102 | } | |
4103 | if (free != pa->pa_free) { | |
9d8b9ec4 | 4104 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 4105 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
4106 | pa, (unsigned long) pa->pa_lstart, |
4107 | (unsigned long) pa->pa_pstart, | |
36bad423 | 4108 | pa->pa_len); |
e29136f8 | 4109 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 4110 | free, pa->pa_free); |
e56eb659 AK |
4111 | /* |
4112 | * pa is already deleted so we use the value obtained | |
4113 | * from the bitmap and continue. | |
4114 | */ | |
c9de560d | 4115 | } |
c9de560d AT |
4116 | atomic_add(free, &sbi->s_mb_discarded); |
4117 | ||
863c37fc | 4118 | return 0; |
c9de560d AT |
4119 | } |
4120 | ||
4ddfef7b ES |
4121 | static noinline_for_stack int |
4122 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 4123 | struct ext4_prealloc_space *pa) |
c9de560d | 4124 | { |
c9de560d AT |
4125 | struct super_block *sb = e4b->bd_sb; |
4126 | ext4_group_t group; | |
4127 | ext4_grpblk_t bit; | |
4128 | ||
60e07cf5 | 4129 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
4130 | BUG_ON(pa->pa_deleted == 0); |
4131 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
4132 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
4133 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
4134 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 4135 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
4136 | |
4137 | return 0; | |
4138 | } | |
4139 | ||
4140 | /* | |
4141 | * releases all preallocations in given group | |
4142 | * | |
4143 | * first, we need to decide discard policy: | |
4144 | * - when do we discard | |
4145 | * 1) ENOSPC | |
4146 | * - how many do we discard | |
4147 | * 1) how many requested | |
4148 | */ | |
4ddfef7b ES |
4149 | static noinline_for_stack int |
4150 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
4151 | ext4_group_t group, int needed) |
4152 | { | |
4153 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4154 | struct buffer_head *bitmap_bh = NULL; | |
4155 | struct ext4_prealloc_space *pa, *tmp; | |
4156 | struct list_head list; | |
4157 | struct ext4_buddy e4b; | |
4158 | int err; | |
4159 | int busy = 0; | |
5b3dc19d | 4160 | int free, free_total = 0; |
c9de560d | 4161 | |
d3df1453 | 4162 | mb_debug(sb, "discard preallocation for group %u\n", group); |
c9de560d | 4163 | if (list_empty(&grp->bb_prealloc_list)) |
bbc4ec77 | 4164 | goto out_dbg; |
c9de560d | 4165 | |
574ca174 | 4166 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4167 | if (IS_ERR(bitmap_bh)) { |
4168 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4169 | ext4_error_err(sb, -err, |
4170 | "Error %d reading block bitmap for %u", | |
4171 | err, group); | |
bbc4ec77 | 4172 | goto out_dbg; |
c9de560d AT |
4173 | } |
4174 | ||
4175 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 4176 | if (err) { |
9651e6b2 KK |
4177 | ext4_warning(sb, "Error %d loading buddy information for %u", |
4178 | err, group); | |
ce89f46c | 4179 | put_bh(bitmap_bh); |
bbc4ec77 | 4180 | goto out_dbg; |
ce89f46c | 4181 | } |
c9de560d AT |
4182 | |
4183 | if (needed == 0) | |
7137d7a4 | 4184 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 4185 | |
c9de560d | 4186 | INIT_LIST_HEAD(&list); |
c9de560d | 4187 | repeat: |
5b3dc19d | 4188 | free = 0; |
c9de560d AT |
4189 | ext4_lock_group(sb, group); |
4190 | list_for_each_entry_safe(pa, tmp, | |
4191 | &grp->bb_prealloc_list, pa_group_list) { | |
4192 | spin_lock(&pa->pa_lock); | |
4193 | if (atomic_read(&pa->pa_count)) { | |
4194 | spin_unlock(&pa->pa_lock); | |
4195 | busy = 1; | |
4196 | continue; | |
4197 | } | |
4198 | if (pa->pa_deleted) { | |
4199 | spin_unlock(&pa->pa_lock); | |
4200 | continue; | |
4201 | } | |
4202 | ||
4203 | /* seems this one can be freed ... */ | |
27bc446e | 4204 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 4205 | |
70022da8 YB |
4206 | if (!free) |
4207 | this_cpu_inc(discard_pa_seq); | |
4208 | ||
c9de560d AT |
4209 | /* we can trust pa_free ... */ |
4210 | free += pa->pa_free; | |
4211 | ||
4212 | spin_unlock(&pa->pa_lock); | |
4213 | ||
4214 | list_del(&pa->pa_group_list); | |
4215 | list_add(&pa->u.pa_tmp_list, &list); | |
4216 | } | |
4217 | ||
c9de560d AT |
4218 | /* now free all selected PAs */ |
4219 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
4220 | ||
4221 | /* remove from object (inode or locality group) */ | |
4222 | spin_lock(pa->pa_obj_lock); | |
4223 | list_del_rcu(&pa->pa_inode_list); | |
4224 | spin_unlock(pa->pa_obj_lock); | |
4225 | ||
cc0fb9ad | 4226 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 4227 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 4228 | else |
3e1e5f50 | 4229 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4230 | |
4231 | list_del(&pa->u.pa_tmp_list); | |
4232 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4233 | } | |
4234 | ||
5b3dc19d JK |
4235 | free_total += free; |
4236 | ||
4237 | /* if we still need more blocks and some PAs were used, try again */ | |
4238 | if (free_total < needed && busy) { | |
4239 | ext4_unlock_group(sb, group); | |
4240 | cond_resched(); | |
4241 | busy = 0; | |
4242 | goto repeat; | |
4243 | } | |
c9de560d | 4244 | ext4_unlock_group(sb, group); |
e39e07fd | 4245 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4246 | put_bh(bitmap_bh); |
bbc4ec77 | 4247 | out_dbg: |
d3df1453 | 4248 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
5b3dc19d JK |
4249 | free_total, group, grp->bb_free); |
4250 | return free_total; | |
c9de560d AT |
4251 | } |
4252 | ||
4253 | /* | |
4254 | * releases all non-used preallocated blocks for given inode | |
4255 | * | |
4256 | * It's important to discard preallocations under i_data_sem | |
4257 | * We don't want another block to be served from the prealloc | |
4258 | * space when we are discarding the inode prealloc space. | |
4259 | * | |
4260 | * FIXME!! Make sure it is valid at all the call sites | |
4261 | */ | |
27bc446e | 4262 | void ext4_discard_preallocations(struct inode *inode, unsigned int needed) |
c9de560d AT |
4263 | { |
4264 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4265 | struct super_block *sb = inode->i_sb; | |
4266 | struct buffer_head *bitmap_bh = NULL; | |
4267 | struct ext4_prealloc_space *pa, *tmp; | |
4268 | ext4_group_t group = 0; | |
4269 | struct list_head list; | |
4270 | struct ext4_buddy e4b; | |
4271 | int err; | |
4272 | ||
c2ea3fde | 4273 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4274 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4275 | return; | |
4276 | } | |
4277 | ||
d3df1453 RH |
4278 | mb_debug(sb, "discard preallocation for inode %lu\n", |
4279 | inode->i_ino); | |
27bc446e | 4280 | trace_ext4_discard_preallocations(inode, |
4281 | atomic_read(&ei->i_prealloc_active), needed); | |
c9de560d AT |
4282 | |
4283 | INIT_LIST_HEAD(&list); | |
4284 | ||
27bc446e | 4285 | if (needed == 0) |
4286 | needed = UINT_MAX; | |
4287 | ||
c9de560d AT |
4288 | repeat: |
4289 | /* first, collect all pa's in the inode */ | |
4290 | spin_lock(&ei->i_prealloc_lock); | |
27bc446e | 4291 | while (!list_empty(&ei->i_prealloc_list) && needed) { |
4292 | pa = list_entry(ei->i_prealloc_list.prev, | |
c9de560d AT |
4293 | struct ext4_prealloc_space, pa_inode_list); |
4294 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4295 | spin_lock(&pa->pa_lock); | |
4296 | if (atomic_read(&pa->pa_count)) { | |
4297 | /* this shouldn't happen often - nobody should | |
4298 | * use preallocation while we're discarding it */ | |
4299 | spin_unlock(&pa->pa_lock); | |
4300 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
4301 | ext4_msg(sb, KERN_ERR, |
4302 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
4303 | WARN_ON(1); |
4304 | schedule_timeout_uninterruptible(HZ); | |
4305 | goto repeat; | |
4306 | ||
4307 | } | |
4308 | if (pa->pa_deleted == 0) { | |
27bc446e | 4309 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4310 | spin_unlock(&pa->pa_lock); |
4311 | list_del_rcu(&pa->pa_inode_list); | |
4312 | list_add(&pa->u.pa_tmp_list, &list); | |
27bc446e | 4313 | needed--; |
c9de560d AT |
4314 | continue; |
4315 | } | |
4316 | ||
4317 | /* someone is deleting pa right now */ | |
4318 | spin_unlock(&pa->pa_lock); | |
4319 | spin_unlock(&ei->i_prealloc_lock); | |
4320 | ||
4321 | /* we have to wait here because pa_deleted | |
4322 | * doesn't mean pa is already unlinked from | |
4323 | * the list. as we might be called from | |
4324 | * ->clear_inode() the inode will get freed | |
4325 | * and concurrent thread which is unlinking | |
4326 | * pa from inode's list may access already | |
4327 | * freed memory, bad-bad-bad */ | |
4328 | ||
4329 | /* XXX: if this happens too often, we can | |
4330 | * add a flag to force wait only in case | |
4331 | * of ->clear_inode(), but not in case of | |
4332 | * regular truncate */ | |
4333 | schedule_timeout_uninterruptible(HZ); | |
4334 | goto repeat; | |
4335 | } | |
4336 | spin_unlock(&ei->i_prealloc_lock); | |
4337 | ||
4338 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4339 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 4340 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 4341 | |
9651e6b2 KK |
4342 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4343 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 4344 | if (err) { |
54d3adbc TT |
4345 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
4346 | err, group); | |
ce89f46c AK |
4347 | continue; |
4348 | } | |
c9de560d | 4349 | |
574ca174 | 4350 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4351 | if (IS_ERR(bitmap_bh)) { |
4352 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4353 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
4354 | err, group); | |
e39e07fd | 4355 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 4356 | continue; |
c9de560d AT |
4357 | } |
4358 | ||
4359 | ext4_lock_group(sb, group); | |
4360 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4361 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4362 | ext4_unlock_group(sb, group); |
4363 | ||
e39e07fd | 4364 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4365 | put_bh(bitmap_bh); |
4366 | ||
4367 | list_del(&pa->u.pa_tmp_list); | |
4368 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4369 | } | |
4370 | } | |
4371 | ||
53f86b17 RH |
4372 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
4373 | { | |
4374 | struct ext4_prealloc_space *pa; | |
4375 | ||
4376 | BUG_ON(ext4_pspace_cachep == NULL); | |
4377 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); | |
4378 | if (!pa) | |
4379 | return -ENOMEM; | |
4380 | atomic_set(&pa->pa_count, 1); | |
4381 | ac->ac_pa = pa; | |
4382 | return 0; | |
4383 | } | |
4384 | ||
4385 | static void ext4_mb_pa_free(struct ext4_allocation_context *ac) | |
4386 | { | |
4387 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
4388 | ||
4389 | BUG_ON(!pa); | |
4390 | ac->ac_pa = NULL; | |
4391 | WARN_ON(!atomic_dec_and_test(&pa->pa_count)); | |
4392 | kmem_cache_free(ext4_pspace_cachep, pa); | |
4393 | } | |
4394 | ||
6ba495e9 | 4395 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 4396 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 4397 | { |
e68cf40c | 4398 | ext4_group_t i, ngroups; |
c9de560d | 4399 | |
d3df1453 | 4400 | if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED) |
e3570639 ES |
4401 | return; |
4402 | ||
8df9675f | 4403 | ngroups = ext4_get_groups_count(sb); |
d3df1453 | 4404 | mb_debug(sb, "groups: "); |
8df9675f | 4405 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
4406 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4407 | struct ext4_prealloc_space *pa; | |
4408 | ext4_grpblk_t start; | |
4409 | struct list_head *cur; | |
4410 | ext4_lock_group(sb, i); | |
4411 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4412 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4413 | pa_group_list); | |
4414 | spin_lock(&pa->pa_lock); | |
4415 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4416 | NULL, &start); | |
4417 | spin_unlock(&pa->pa_lock); | |
d3df1453 RH |
4418 | mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
4419 | pa->pa_len); | |
c9de560d | 4420 | } |
60bd63d1 | 4421 | ext4_unlock_group(sb, i); |
d3df1453 RH |
4422 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
4423 | grp->bb_fragments); | |
c9de560d | 4424 | } |
c9de560d | 4425 | } |
e68cf40c RH |
4426 | |
4427 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4428 | { | |
4429 | struct super_block *sb = ac->ac_sb; | |
4430 | ||
d3df1453 | 4431 | if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED) |
e68cf40c RH |
4432 | return; |
4433 | ||
d3df1453 | 4434 | mb_debug(sb, "Can't allocate:" |
e68cf40c | 4435 | " Allocation context details:"); |
d3df1453 | 4436 | mb_debug(sb, "status %u flags 0x%x", |
e68cf40c | 4437 | ac->ac_status, ac->ac_flags); |
d3df1453 | 4438 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
e68cf40c RH |
4439 | "goal %lu/%lu/%lu@%lu, " |
4440 | "best %lu/%lu/%lu@%lu cr %d", | |
4441 | (unsigned long)ac->ac_o_ex.fe_group, | |
4442 | (unsigned long)ac->ac_o_ex.fe_start, | |
4443 | (unsigned long)ac->ac_o_ex.fe_len, | |
4444 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4445 | (unsigned long)ac->ac_g_ex.fe_group, | |
4446 | (unsigned long)ac->ac_g_ex.fe_start, | |
4447 | (unsigned long)ac->ac_g_ex.fe_len, | |
4448 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4449 | (unsigned long)ac->ac_b_ex.fe_group, | |
4450 | (unsigned long)ac->ac_b_ex.fe_start, | |
4451 | (unsigned long)ac->ac_b_ex.fe_len, | |
4452 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4453 | (int)ac->ac_criteria); | |
d3df1453 | 4454 | mb_debug(sb, "%u found", ac->ac_found); |
e68cf40c RH |
4455 | ext4_mb_show_pa(sb); |
4456 | } | |
c9de560d | 4457 | #else |
e68cf40c RH |
4458 | static inline void ext4_mb_show_pa(struct super_block *sb) |
4459 | { | |
4460 | return; | |
4461 | } | |
c9de560d AT |
4462 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
4463 | { | |
e68cf40c | 4464 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
4465 | return; |
4466 | } | |
4467 | #endif | |
4468 | ||
4469 | /* | |
4470 | * We use locality group preallocation for small size file. The size of the | |
4471 | * file is determined by the current size or the resulting size after | |
4472 | * allocation which ever is larger | |
4473 | * | |
b713a5ec | 4474 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4475 | */ |
4476 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4477 | { | |
4478 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4479 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4480 | loff_t size, isize; | |
4481 | ||
4482 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4483 | return; | |
4484 | ||
4ba74d00 TT |
4485 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4486 | return; | |
4487 | ||
53accfa9 | 4488 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
4489 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4490 | >> bsbits; | |
c9de560d | 4491 | |
82dd124c NB |
4492 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
4493 | !inode_is_open_for_write(ac->ac_inode)) { | |
50797481 TT |
4494 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; |
4495 | return; | |
4496 | } | |
4497 | ||
ebbe0277 RD |
4498 | if (sbi->s_mb_group_prealloc <= 0) { |
4499 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
4500 | return; | |
4501 | } | |
4502 | ||
c9de560d | 4503 | /* don't use group allocation for large files */ |
71780577 | 4504 | size = max(size, isize); |
cc483f10 | 4505 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 4506 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 4507 | return; |
4ba74d00 | 4508 | } |
c9de560d AT |
4509 | |
4510 | BUG_ON(ac->ac_lg != NULL); | |
4511 | /* | |
4512 | * locality group prealloc space are per cpu. The reason for having | |
4513 | * per cpu locality group is to reduce the contention between block | |
4514 | * request from multiple CPUs. | |
4515 | */ | |
a0b6bc63 | 4516 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
4517 | |
4518 | /* we're going to use group allocation */ | |
4519 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4520 | ||
4521 | /* serialize all allocations in the group */ | |
4522 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4523 | } | |
4524 | ||
4ddfef7b ES |
4525 | static noinline_for_stack int |
4526 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4527 | struct ext4_allocation_request *ar) |
4528 | { | |
4529 | struct super_block *sb = ar->inode->i_sb; | |
4530 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4531 | struct ext4_super_block *es = sbi->s_es; | |
4532 | ext4_group_t group; | |
498e5f24 TT |
4533 | unsigned int len; |
4534 | ext4_fsblk_t goal; | |
c9de560d AT |
4535 | ext4_grpblk_t block; |
4536 | ||
4537 | /* we can't allocate > group size */ | |
4538 | len = ar->len; | |
4539 | ||
4540 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
4541 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
4542 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4543 | |
4544 | /* start searching from the goal */ | |
4545 | goal = ar->goal; | |
4546 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4547 | goal >= ext4_blocks_count(es)) | |
4548 | goal = le32_to_cpu(es->s_first_data_block); | |
4549 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4550 | ||
4551 | /* set up allocation goals */ | |
f5a44db5 | 4552 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 4553 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4554 | ac->ac_sb = sb; |
4555 | ac->ac_inode = ar->inode; | |
53accfa9 | 4556 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4557 | ac->ac_o_ex.fe_group = group; |
4558 | ac->ac_o_ex.fe_start = block; | |
4559 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4560 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4561 | ac->ac_flags = ar->flags; |
c9de560d | 4562 | |
3cb77bd2 | 4563 | /* we have to define context: we'll work with a file or |
c9de560d AT |
4564 | * locality group. this is a policy, actually */ |
4565 | ext4_mb_group_or_file(ac); | |
4566 | ||
d3df1453 | 4567 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
c9de560d AT |
4568 | "left: %u/%u, right %u/%u to %swritable\n", |
4569 | (unsigned) ar->len, (unsigned) ar->logical, | |
4570 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4571 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4572 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 4573 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
4574 | return 0; |
4575 | ||
4576 | } | |
4577 | ||
6be2ded1 AK |
4578 | static noinline_for_stack void |
4579 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4580 | struct ext4_locality_group *lg, | |
4581 | int order, int total_entries) | |
4582 | { | |
4583 | ext4_group_t group = 0; | |
4584 | struct ext4_buddy e4b; | |
4585 | struct list_head discard_list; | |
4586 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4587 | |
d3df1453 | 4588 | mb_debug(sb, "discard locality group preallocation\n"); |
6be2ded1 AK |
4589 | |
4590 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4591 | |
4592 | spin_lock(&lg->lg_prealloc_lock); | |
4593 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
92e9c58c MB |
4594 | pa_inode_list, |
4595 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
4596 | spin_lock(&pa->pa_lock); |
4597 | if (atomic_read(&pa->pa_count)) { | |
4598 | /* | |
4599 | * This is the pa that we just used | |
4600 | * for block allocation. So don't | |
4601 | * free that | |
4602 | */ | |
4603 | spin_unlock(&pa->pa_lock); | |
4604 | continue; | |
4605 | } | |
4606 | if (pa->pa_deleted) { | |
4607 | spin_unlock(&pa->pa_lock); | |
4608 | continue; | |
4609 | } | |
4610 | /* only lg prealloc space */ | |
cc0fb9ad | 4611 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4612 | |
4613 | /* seems this one can be freed ... */ | |
27bc446e | 4614 | ext4_mb_mark_pa_deleted(sb, pa); |
6be2ded1 AK |
4615 | spin_unlock(&pa->pa_lock); |
4616 | ||
4617 | list_del_rcu(&pa->pa_inode_list); | |
4618 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4619 | ||
4620 | total_entries--; | |
4621 | if (total_entries <= 5) { | |
4622 | /* | |
4623 | * we want to keep only 5 entries | |
4624 | * allowing it to grow to 8. This | |
4625 | * mak sure we don't call discard | |
4626 | * soon for this list. | |
4627 | */ | |
4628 | break; | |
4629 | } | |
4630 | } | |
4631 | spin_unlock(&lg->lg_prealloc_lock); | |
4632 | ||
4633 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 4634 | int err; |
6be2ded1 | 4635 | |
bd86298e | 4636 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
4637 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4638 | GFP_NOFS|__GFP_NOFAIL); | |
4639 | if (err) { | |
54d3adbc TT |
4640 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
4641 | err, group); | |
6be2ded1 AK |
4642 | continue; |
4643 | } | |
4644 | ext4_lock_group(sb, group); | |
4645 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4646 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4647 | ext4_unlock_group(sb, group); |
4648 | ||
e39e07fd | 4649 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4650 | list_del(&pa->u.pa_tmp_list); |
4651 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4652 | } | |
6be2ded1 AK |
4653 | } |
4654 | ||
4655 | /* | |
4656 | * We have incremented pa_count. So it cannot be freed at this | |
4657 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4658 | * possible from this lg. That means pa_free cannot be updated. | |
4659 | * | |
4660 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4661 | * which can cause the lg_prealloc_list to be updated. | |
4662 | */ | |
4663 | ||
4664 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4665 | { | |
4666 | int order, added = 0, lg_prealloc_count = 1; | |
4667 | struct super_block *sb = ac->ac_sb; | |
4668 | struct ext4_locality_group *lg = ac->ac_lg; | |
4669 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4670 | ||
4671 | order = fls(pa->pa_free) - 1; | |
4672 | if (order > PREALLOC_TB_SIZE - 1) | |
4673 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4674 | order = PREALLOC_TB_SIZE - 1; | |
4675 | /* Add the prealloc space to lg */ | |
f1167009 | 4676 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 4677 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
92e9c58c MB |
4678 | pa_inode_list, |
4679 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
4680 | spin_lock(&tmp_pa->pa_lock); |
4681 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4682 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4683 | continue; |
4684 | } | |
4685 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4686 | /* Add to the tail of the previous entry */ | |
4687 | list_add_tail_rcu(&pa->pa_inode_list, | |
4688 | &tmp_pa->pa_inode_list); | |
4689 | added = 1; | |
4690 | /* | |
4691 | * we want to count the total | |
4692 | * number of entries in the list | |
4693 | */ | |
4694 | } | |
4695 | spin_unlock(&tmp_pa->pa_lock); | |
4696 | lg_prealloc_count++; | |
4697 | } | |
4698 | if (!added) | |
4699 | list_add_tail_rcu(&pa->pa_inode_list, | |
4700 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4701 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4702 | |
4703 | /* Now trim the list to be not more than 8 elements */ | |
4704 | if (lg_prealloc_count > 8) { | |
4705 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4706 | order, lg_prealloc_count); |
6be2ded1 AK |
4707 | return; |
4708 | } | |
4709 | return ; | |
4710 | } | |
4711 | ||
27bc446e | 4712 | /* |
4713 | * if per-inode prealloc list is too long, trim some PA | |
4714 | */ | |
4715 | static void ext4_mb_trim_inode_pa(struct inode *inode) | |
4716 | { | |
4717 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4718 | struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); | |
4719 | int count, delta; | |
4720 | ||
4721 | count = atomic_read(&ei->i_prealloc_active); | |
4722 | delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1; | |
4723 | if (count > sbi->s_mb_max_inode_prealloc + delta) { | |
4724 | count -= sbi->s_mb_max_inode_prealloc; | |
4725 | ext4_discard_preallocations(inode, count); | |
4726 | } | |
4727 | } | |
4728 | ||
c9de560d AT |
4729 | /* |
4730 | * release all resource we used in allocation | |
4731 | */ | |
4732 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4733 | { | |
27bc446e | 4734 | struct inode *inode = ac->ac_inode; |
4735 | struct ext4_inode_info *ei = EXT4_I(inode); | |
53accfa9 | 4736 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4737 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4738 | if (pa) { | |
cc0fb9ad | 4739 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4740 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4741 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4742 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4743 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4744 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4745 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4746 | spin_unlock(&pa->pa_lock); | |
66d5e027 | 4747 | |
4748 | /* | |
4749 | * We want to add the pa to the right bucket. | |
4750 | * Remove it from the list and while adding | |
4751 | * make sure the list to which we are adding | |
4752 | * doesn't grow big. | |
4753 | */ | |
4754 | if (likely(pa->pa_free)) { | |
4755 | spin_lock(pa->pa_obj_lock); | |
4756 | list_del_rcu(&pa->pa_inode_list); | |
4757 | spin_unlock(pa->pa_obj_lock); | |
4758 | ext4_mb_add_n_trim(ac); | |
4759 | } | |
ba443916 | 4760 | } |
27bc446e | 4761 | |
4762 | if (pa->pa_type == MB_INODE_PA) { | |
4763 | /* | |
4764 | * treat per-inode prealloc list as a lru list, then try | |
4765 | * to trim the least recently used PA. | |
4766 | */ | |
4767 | spin_lock(pa->pa_obj_lock); | |
4768 | list_move(&pa->pa_inode_list, &ei->i_prealloc_list); | |
4769 | spin_unlock(pa->pa_obj_lock); | |
4770 | } | |
4771 | ||
ba443916 AK |
4772 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
4773 | } | |
c9de560d | 4774 | if (ac->ac_bitmap_page) |
09cbfeaf | 4775 | put_page(ac->ac_bitmap_page); |
c9de560d | 4776 | if (ac->ac_buddy_page) |
09cbfeaf | 4777 | put_page(ac->ac_buddy_page); |
c9de560d AT |
4778 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
4779 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4780 | ext4_mb_collect_stats(ac); | |
27bc446e | 4781 | ext4_mb_trim_inode_pa(inode); |
c9de560d AT |
4782 | return 0; |
4783 | } | |
4784 | ||
4785 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4786 | { | |
8df9675f | 4787 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4788 | int ret; |
4789 | int freed = 0; | |
4790 | ||
9bffad1e | 4791 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4792 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4793 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4794 | freed += ret; | |
4795 | needed -= ret; | |
4796 | } | |
4797 | ||
4798 | return freed; | |
4799 | } | |
4800 | ||
cf5e2ca6 | 4801 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
07b5b8e1 | 4802 | struct ext4_allocation_context *ac, u64 *seq) |
cf5e2ca6 RH |
4803 | { |
4804 | int freed; | |
07b5b8e1 RH |
4805 | u64 seq_retry = 0; |
4806 | bool ret = false; | |
cf5e2ca6 RH |
4807 | |
4808 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
07b5b8e1 RH |
4809 | if (freed) { |
4810 | ret = true; | |
4811 | goto out_dbg; | |
4812 | } | |
4813 | seq_retry = ext4_get_discard_pa_seq_sum(); | |
99377830 RH |
4814 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
4815 | ac->ac_flags |= EXT4_MB_STRICT_CHECK; | |
07b5b8e1 RH |
4816 | *seq = seq_retry; |
4817 | ret = true; | |
4818 | } | |
4819 | ||
4820 | out_dbg: | |
4821 | mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); | |
4822 | return ret; | |
cf5e2ca6 RH |
4823 | } |
4824 | ||
c9de560d AT |
4825 | /* |
4826 | * Main entry point into mballoc to allocate blocks | |
4827 | * it tries to use preallocation first, then falls back | |
4828 | * to usual allocation | |
4829 | */ | |
4830 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4831 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4832 | { |
256bdb49 | 4833 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4834 | struct ext4_sb_info *sbi; |
4835 | struct super_block *sb; | |
4836 | ext4_fsblk_t block = 0; | |
60e58e0f | 4837 | unsigned int inquota = 0; |
53accfa9 | 4838 | unsigned int reserv_clstrs = 0; |
07b5b8e1 | 4839 | u64 seq; |
c9de560d | 4840 | |
b10a44c3 | 4841 | might_sleep(); |
c9de560d AT |
4842 | sb = ar->inode->i_sb; |
4843 | sbi = EXT4_SB(sb); | |
4844 | ||
9bffad1e | 4845 | trace_ext4_request_blocks(ar); |
ba80b101 | 4846 | |
45dc63e7 | 4847 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 4848 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
4849 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
4850 | ||
e3cf5d5d | 4851 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
4852 | /* Without delayed allocation we need to verify |
4853 | * there is enough free blocks to do block allocation | |
4854 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4855 | */ |
55f020db | 4856 | while (ar->len && |
e7d5f315 | 4857 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4858 | |
030ba6bc | 4859 | /* let others to free the space */ |
bb8b20ed | 4860 | cond_resched(); |
030ba6bc AK |
4861 | ar->len = ar->len >> 1; |
4862 | } | |
4863 | if (!ar->len) { | |
bbc4ec77 | 4864 | ext4_mb_show_pa(sb); |
a30d542a AK |
4865 | *errp = -ENOSPC; |
4866 | return 0; | |
4867 | } | |
53accfa9 | 4868 | reserv_clstrs = ar->len; |
55f020db | 4869 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4870 | dquot_alloc_block_nofail(ar->inode, |
4871 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4872 | } else { |
4873 | while (ar->len && | |
53accfa9 TT |
4874 | dquot_alloc_block(ar->inode, |
4875 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4876 | |
4877 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4878 | ar->len--; | |
4879 | } | |
60e58e0f MC |
4880 | } |
4881 | inquota = ar->len; | |
4882 | if (ar->len == 0) { | |
4883 | *errp = -EDQUOT; | |
6c7a120a | 4884 | goto out; |
60e58e0f | 4885 | } |
07031431 | 4886 | } |
d2a17637 | 4887 | |
85556c9a | 4888 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4889 | if (!ac) { |
363d4251 | 4890 | ar->len = 0; |
256bdb49 | 4891 | *errp = -ENOMEM; |
6c7a120a | 4892 | goto out; |
256bdb49 ES |
4893 | } |
4894 | ||
256bdb49 | 4895 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4896 | if (*errp) { |
4897 | ar->len = 0; | |
6c7a120a | 4898 | goto out; |
c9de560d AT |
4899 | } |
4900 | ||
256bdb49 | 4901 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
81198536 | 4902 | seq = this_cpu_read(discard_pa_seq); |
256bdb49 | 4903 | if (!ext4_mb_use_preallocated(ac)) { |
256bdb49 ES |
4904 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4905 | ext4_mb_normalize_request(ac, ar); | |
53f86b17 RH |
4906 | |
4907 | *errp = ext4_mb_pa_alloc(ac); | |
4908 | if (*errp) | |
4909 | goto errout; | |
c9de560d AT |
4910 | repeat: |
4911 | /* allocate space in core */ | |
6c7a120a | 4912 | *errp = ext4_mb_regular_allocator(ac); |
53f86b17 RH |
4913 | /* |
4914 | * pa allocated above is added to grp->bb_prealloc_list only | |
4915 | * when we were able to allocate some block i.e. when | |
4916 | * ac->ac_status == AC_STATUS_FOUND. | |
4917 | * And error from above mean ac->ac_status != AC_STATUS_FOUND | |
4918 | * So we have to free this pa here itself. | |
4919 | */ | |
2c00ef3e | 4920 | if (*errp) { |
53f86b17 | 4921 | ext4_mb_pa_free(ac); |
2c00ef3e AK |
4922 | ext4_discard_allocated_blocks(ac); |
4923 | goto errout; | |
4924 | } | |
53f86b17 RH |
4925 | if (ac->ac_status == AC_STATUS_FOUND && |
4926 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) | |
4927 | ext4_mb_pa_free(ac); | |
c9de560d | 4928 | } |
256bdb49 | 4929 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4930 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 4931 | if (*errp) { |
b844167e | 4932 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4933 | goto errout; |
4934 | } else { | |
519deca0 AK |
4935 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4936 | ar->len = ac->ac_b_ex.fe_len; | |
4937 | } | |
c9de560d | 4938 | } else { |
07b5b8e1 | 4939 | if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) |
c9de560d | 4940 | goto repeat; |
53f86b17 RH |
4941 | /* |
4942 | * If block allocation fails then the pa allocated above | |
4943 | * needs to be freed here itself. | |
4944 | */ | |
4945 | ext4_mb_pa_free(ac); | |
c9de560d | 4946 | *errp = -ENOSPC; |
6c7a120a AK |
4947 | } |
4948 | ||
6d138ced | 4949 | errout: |
6c7a120a | 4950 | if (*errp) { |
256bdb49 | 4951 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4952 | ar->len = 0; |
256bdb49 | 4953 | ext4_mb_show_ac(ac); |
c9de560d | 4954 | } |
256bdb49 | 4955 | ext4_mb_release_context(ac); |
6c7a120a AK |
4956 | out: |
4957 | if (ac) | |
4958 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4959 | if (inquota && ar->len < inquota) |
53accfa9 | 4960 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4961 | if (!ar->len) { |
e3cf5d5d | 4962 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 4963 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4964 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4965 | reserv_clstrs); |
0087d9fb | 4966 | } |
c9de560d | 4967 | |
9bffad1e | 4968 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4969 | |
c9de560d AT |
4970 | return block; |
4971 | } | |
c9de560d | 4972 | |
c894058d AK |
4973 | /* |
4974 | * We can merge two free data extents only if the physical blocks | |
4975 | * are contiguous, AND the extents were freed by the same transaction, | |
4976 | * AND the blocks are associated with the same group. | |
4977 | */ | |
a0154344 DJ |
4978 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
4979 | struct ext4_free_data *entry, | |
4980 | struct ext4_free_data *new_entry, | |
4981 | struct rb_root *entry_rb_root) | |
c894058d | 4982 | { |
a0154344 DJ |
4983 | if ((entry->efd_tid != new_entry->efd_tid) || |
4984 | (entry->efd_group != new_entry->efd_group)) | |
4985 | return; | |
4986 | if (entry->efd_start_cluster + entry->efd_count == | |
4987 | new_entry->efd_start_cluster) { | |
4988 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
4989 | new_entry->efd_count += entry->efd_count; | |
4990 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
4991 | entry->efd_start_cluster) { | |
4992 | new_entry->efd_count += entry->efd_count; | |
4993 | } else | |
4994 | return; | |
4995 | spin_lock(&sbi->s_md_lock); | |
4996 | list_del(&entry->efd_list); | |
4997 | spin_unlock(&sbi->s_md_lock); | |
4998 | rb_erase(&entry->efd_node, entry_rb_root); | |
4999 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
5000 | } |
5001 | ||
4ddfef7b ES |
5002 | static noinline_for_stack int |
5003 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 5004 | struct ext4_free_data *new_entry) |
c9de560d | 5005 | { |
e29136f8 | 5006 | ext4_group_t group = e4b->bd_group; |
84130193 | 5007 | ext4_grpblk_t cluster; |
d08854f5 | 5008 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 5009 | struct ext4_free_data *entry; |
c9de560d AT |
5010 | struct ext4_group_info *db = e4b->bd_info; |
5011 | struct super_block *sb = e4b->bd_sb; | |
5012 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
5013 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
5014 | struct rb_node *parent = NULL, *new_node; | |
5015 | ||
0390131b | 5016 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
5017 | BUG_ON(e4b->bd_bitmap_page == NULL); |
5018 | BUG_ON(e4b->bd_buddy_page == NULL); | |
5019 | ||
18aadd47 BJ |
5020 | new_node = &new_entry->efd_node; |
5021 | cluster = new_entry->efd_start_cluster; | |
c894058d | 5022 | |
c894058d AK |
5023 | if (!*n) { |
5024 | /* first free block exent. We need to | |
5025 | protect buddy cache from being freed, | |
5026 | * otherwise we'll refresh it from | |
5027 | * on-disk bitmap and lose not-yet-available | |
5028 | * blocks */ | |
09cbfeaf KS |
5029 | get_page(e4b->bd_buddy_page); |
5030 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
5031 | } |
5032 | while (*n) { | |
5033 | parent = *n; | |
18aadd47 BJ |
5034 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
5035 | if (cluster < entry->efd_start_cluster) | |
c894058d | 5036 | n = &(*n)->rb_left; |
18aadd47 | 5037 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
5038 | n = &(*n)->rb_right; |
5039 | else { | |
e29136f8 | 5040 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
5041 | ext4_group_first_block_no(sb, group) + |
5042 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 5043 | "Block already on to-be-freed list"); |
c894058d | 5044 | return 0; |
c9de560d | 5045 | } |
c894058d | 5046 | } |
c9de560d | 5047 | |
c894058d AK |
5048 | rb_link_node(new_node, parent, n); |
5049 | rb_insert_color(new_node, &db->bb_free_root); | |
5050 | ||
5051 | /* Now try to see the extent can be merged to left and right */ | |
5052 | node = rb_prev(new_node); | |
5053 | if (node) { | |
18aadd47 | 5054 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5055 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5056 | &(db->bb_free_root)); | |
c894058d | 5057 | } |
c9de560d | 5058 | |
c894058d AK |
5059 | node = rb_next(new_node); |
5060 | if (node) { | |
18aadd47 | 5061 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5062 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5063 | &(db->bb_free_root)); | |
c9de560d | 5064 | } |
a0154344 | 5065 | |
d08854f5 | 5066 | spin_lock(&sbi->s_md_lock); |
a0154344 | 5067 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
5068 | sbi->s_mb_free_pending += clusters; |
5069 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
5070 | return 0; |
5071 | } | |
5072 | ||
44338711 TT |
5073 | /** |
5074 | * ext4_free_blocks() -- Free given blocks and update quota | |
5075 | * @handle: handle for this transaction | |
5076 | * @inode: inode | |
c60990b3 TT |
5077 | * @bh: optional buffer of the block to be freed |
5078 | * @block: starting physical block to be freed | |
5079 | * @count: number of blocks to be freed | |
5def1360 | 5080 | * @flags: flags used by ext4_free_blocks |
c9de560d | 5081 | */ |
44338711 | 5082 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
5083 | struct buffer_head *bh, ext4_fsblk_t block, |
5084 | unsigned long count, int flags) | |
c9de560d | 5085 | { |
26346ff6 | 5086 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 5087 | struct super_block *sb = inode->i_sb; |
c9de560d | 5088 | struct ext4_group_desc *gdp; |
498e5f24 | 5089 | unsigned int overflow; |
c9de560d AT |
5090 | ext4_grpblk_t bit; |
5091 | struct buffer_head *gd_bh; | |
5092 | ext4_group_t block_group; | |
5093 | struct ext4_sb_info *sbi; | |
5094 | struct ext4_buddy e4b; | |
84130193 | 5095 | unsigned int count_clusters; |
c9de560d AT |
5096 | int err = 0; |
5097 | int ret; | |
5098 | ||
b10a44c3 | 5099 | might_sleep(); |
e6362609 TT |
5100 | if (bh) { |
5101 | if (block) | |
5102 | BUG_ON(block != bh->b_blocknr); | |
5103 | else | |
5104 | block = bh->b_blocknr; | |
5105 | } | |
c9de560d | 5106 | |
c9de560d | 5107 | sbi = EXT4_SB(sb); |
1f2acb60 | 5108 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
ce9f24cc | 5109 | !ext4_inode_block_valid(inode, block, count)) { |
12062ddd | 5110 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 5111 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
5112 | goto error_return; |
5113 | } | |
5114 | ||
0610b6e9 | 5115 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
5116 | trace_ext4_free_blocks(inode, block, count, flags); |
5117 | ||
9c02ac97 DJ |
5118 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
5119 | BUG_ON(count > 1); | |
e6362609 | 5120 | |
9c02ac97 DJ |
5121 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
5122 | inode, bh, block); | |
e6362609 TT |
5123 | } |
5124 | ||
84130193 TT |
5125 | /* |
5126 | * If the extent to be freed does not begin on a cluster | |
5127 | * boundary, we need to deal with partial clusters at the | |
5128 | * beginning and end of the extent. Normally we will free | |
5129 | * blocks at the beginning or the end unless we are explicitly | |
5130 | * requested to avoid doing so. | |
5131 | */ | |
f5a44db5 | 5132 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
5133 | if (overflow) { |
5134 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
5135 | overflow = sbi->s_cluster_ratio - overflow; | |
5136 | block += overflow; | |
5137 | if (count > overflow) | |
5138 | count -= overflow; | |
5139 | else | |
5140 | return; | |
5141 | } else { | |
5142 | block -= overflow; | |
5143 | count += overflow; | |
5144 | } | |
5145 | } | |
f5a44db5 | 5146 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
5147 | if (overflow) { |
5148 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
5149 | if (count > overflow) | |
5150 | count -= overflow; | |
5151 | else | |
5152 | return; | |
5153 | } else | |
5154 | count += sbi->s_cluster_ratio - overflow; | |
5155 | } | |
5156 | ||
9c02ac97 DJ |
5157 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
5158 | int i; | |
f96c450d | 5159 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; |
9c02ac97 DJ |
5160 | |
5161 | for (i = 0; i < count; i++) { | |
5162 | cond_resched(); | |
f96c450d DJ |
5163 | if (is_metadata) |
5164 | bh = sb_find_get_block(inode->i_sb, block + i); | |
5165 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
9c02ac97 DJ |
5166 | } |
5167 | } | |
5168 | ||
c9de560d AT |
5169 | do_more: |
5170 | overflow = 0; | |
5171 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
5172 | ||
163a203d DW |
5173 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
5174 | ext4_get_group_info(sb, block_group)))) | |
5175 | return; | |
5176 | ||
c9de560d AT |
5177 | /* |
5178 | * Check to see if we are freeing blocks across a group | |
5179 | * boundary. | |
5180 | */ | |
84130193 TT |
5181 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
5182 | overflow = EXT4_C2B(sbi, bit) + count - | |
5183 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
5184 | count -= overflow; |
5185 | } | |
810da240 | 5186 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 5187 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5188 | if (IS_ERR(bitmap_bh)) { |
5189 | err = PTR_ERR(bitmap_bh); | |
5190 | bitmap_bh = NULL; | |
c9de560d | 5191 | goto error_return; |
ce89f46c | 5192 | } |
c9de560d | 5193 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
5194 | if (!gdp) { |
5195 | err = -EIO; | |
c9de560d | 5196 | goto error_return; |
ce89f46c | 5197 | } |
c9de560d AT |
5198 | |
5199 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
5200 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
5201 | in_range(block, ext4_inode_table(sb, gdp), | |
49598e04 | 5202 | sbi->s_itb_per_group) || |
c9de560d | 5203 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
49598e04 | 5204 | sbi->s_itb_per_group)) { |
c9de560d | 5205 | |
12062ddd | 5206 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 5207 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
5208 | /* err = 0. ext4_std_error should be a no op */ |
5209 | goto error_return; | |
c9de560d AT |
5210 | } |
5211 | ||
5212 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
5213 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
5214 | if (err) | |
5215 | goto error_return; | |
5216 | ||
5217 | /* | |
5218 | * We are about to modify some metadata. Call the journal APIs | |
5219 | * to unshare ->b_data if a currently-committing transaction is | |
5220 | * using it | |
5221 | */ | |
5222 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
5223 | err = ext4_journal_get_write_access(handle, gd_bh); | |
5224 | if (err) | |
5225 | goto error_return; | |
c9de560d AT |
5226 | #ifdef AGGRESSIVE_CHECK |
5227 | { | |
5228 | int i; | |
84130193 | 5229 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
5230 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
5231 | } | |
5232 | #endif | |
84130193 | 5233 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 5234 | |
adb7ef60 KK |
5235 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
5236 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
5237 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
5238 | if (err) |
5239 | goto error_return; | |
e6362609 | 5240 | |
f96c450d DJ |
5241 | /* |
5242 | * We need to make sure we don't reuse the freed block until after the | |
5243 | * transaction is committed. We make an exception if the inode is to be | |
5244 | * written in writeback mode since writeback mode has weak data | |
5245 | * consistency guarantees. | |
5246 | */ | |
5247 | if (ext4_handle_valid(handle) && | |
5248 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
5249 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
5250 | struct ext4_free_data *new_entry; |
5251 | /* | |
7444a072 MH |
5252 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
5253 | * to fail. | |
7a2fcbf7 | 5254 | */ |
7444a072 MH |
5255 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
5256 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
5257 | new_entry->efd_start_cluster = bit; |
5258 | new_entry->efd_group = block_group; | |
5259 | new_entry->efd_count = count_clusters; | |
5260 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 5261 | |
7a2fcbf7 | 5262 | ext4_lock_group(sb, block_group); |
84130193 | 5263 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 5264 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 5265 | } else { |
7a2fcbf7 AK |
5266 | /* need to update group_info->bb_free and bitmap |
5267 | * with group lock held. generate_buddy look at | |
5268 | * them with group lock_held | |
5269 | */ | |
d71c1ae2 | 5270 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
5271 | err = ext4_issue_discard(sb, block_group, bit, count, |
5272 | NULL); | |
d71c1ae2 LC |
5273 | if (err && err != -EOPNOTSUPP) |
5274 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
5275 | " group:%d block:%d count:%lu failed" | |
5276 | " with %d", block_group, bit, count, | |
5277 | err); | |
8f9ff189 LC |
5278 | } else |
5279 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 5280 | |
955ce5f5 | 5281 | ext4_lock_group(sb, block_group); |
84130193 TT |
5282 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
5283 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
5284 | } |
5285 | ||
021b65bb TT |
5286 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
5287 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 5288 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 5289 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 5290 | ext4_unlock_group(sb, block_group); |
c9de560d | 5291 | |
772cb7c8 JS |
5292 | if (sbi->s_log_groups_per_flex) { |
5293 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 5294 | atomic64_add(count_clusters, |
7c990728 SJS |
5295 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
5296 | flex_group)->free_clusters); | |
772cb7c8 JS |
5297 | } |
5298 | ||
9fe67149 EW |
5299 | /* |
5300 | * on a bigalloc file system, defer the s_freeclusters_counter | |
5301 | * update to the caller (ext4_remove_space and friends) so they | |
5302 | * can determine if a cluster freed here should be rereserved | |
5303 | */ | |
5304 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
5305 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
5306 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
5307 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
5308 | count_clusters); | |
5309 | } | |
7d734532 JK |
5310 | |
5311 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 5312 | |
7a2fcbf7 AK |
5313 | /* We dirtied the bitmap block */ |
5314 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5315 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5316 | ||
c9de560d AT |
5317 | /* And the group descriptor block */ |
5318 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 5319 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
5320 | if (!err) |
5321 | err = ret; | |
5322 | ||
5323 | if (overflow && !err) { | |
5324 | block += count; | |
5325 | count = overflow; | |
5326 | put_bh(bitmap_bh); | |
5327 | goto do_more; | |
5328 | } | |
c9de560d AT |
5329 | error_return: |
5330 | brelse(bitmap_bh); | |
5331 | ext4_std_error(sb, err); | |
5332 | return; | |
5333 | } | |
7360d173 | 5334 | |
2846e820 | 5335 | /** |
0529155e | 5336 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
5337 | * @handle: handle to this transaction |
5338 | * @sb: super block | |
4907cb7b | 5339 | * @block: start physical block to add to the block group |
2846e820 AG |
5340 | * @count: number of blocks to free |
5341 | * | |
e73a347b | 5342 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 5343 | */ |
cc7365df | 5344 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
5345 | ext4_fsblk_t block, unsigned long count) |
5346 | { | |
5347 | struct buffer_head *bitmap_bh = NULL; | |
5348 | struct buffer_head *gd_bh; | |
5349 | ext4_group_t block_group; | |
5350 | ext4_grpblk_t bit; | |
5351 | unsigned int i; | |
5352 | struct ext4_group_desc *desc; | |
5353 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 5354 | struct ext4_buddy e4b; |
d77147ff | 5355 | int err = 0, ret, free_clusters_count; |
5356 | ext4_grpblk_t clusters_freed; | |
5357 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
5358 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
5359 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
5360 | |
5361 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
5362 | ||
4740b830 YY |
5363 | if (count == 0) |
5364 | return 0; | |
5365 | ||
2846e820 | 5366 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
5367 | /* |
5368 | * Check to see if we are freeing blocks across a group | |
5369 | * boundary. | |
5370 | */ | |
d77147ff | 5371 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
5372 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
5373 | block_group); |
5374 | err = -EINVAL; | |
2846e820 | 5375 | goto error_return; |
cc7365df | 5376 | } |
2cd05cc3 | 5377 | |
2846e820 | 5378 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5379 | if (IS_ERR(bitmap_bh)) { |
5380 | err = PTR_ERR(bitmap_bh); | |
5381 | bitmap_bh = NULL; | |
2846e820 | 5382 | goto error_return; |
cc7365df YY |
5383 | } |
5384 | ||
2846e820 | 5385 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
5386 | if (!desc) { |
5387 | err = -EIO; | |
2846e820 | 5388 | goto error_return; |
cc7365df | 5389 | } |
2846e820 AG |
5390 | |
5391 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
5392 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
5393 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
5394 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
5395 | sbi->s_itb_per_group)) { | |
5396 | ext4_error(sb, "Adding blocks in system zones - " | |
5397 | "Block = %llu, count = %lu", | |
5398 | block, count); | |
cc7365df | 5399 | err = -EINVAL; |
2846e820 AG |
5400 | goto error_return; |
5401 | } | |
5402 | ||
2cd05cc3 TT |
5403 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
5404 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
5405 | if (err) |
5406 | goto error_return; | |
5407 | ||
5408 | /* | |
5409 | * We are about to modify some metadata. Call the journal APIs | |
5410 | * to unshare ->b_data if a currently-committing transaction is | |
5411 | * using it | |
5412 | */ | |
5413 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
5414 | err = ext4_journal_get_write_access(handle, gd_bh); | |
5415 | if (err) | |
5416 | goto error_return; | |
e73a347b | 5417 | |
d77147ff | 5418 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 5419 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 5420 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
5421 | ext4_error(sb, "bit already cleared for block %llu", |
5422 | (ext4_fsblk_t)(block + i)); | |
5423 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
5424 | } else { | |
d77147ff | 5425 | clusters_freed++; |
2846e820 AG |
5426 | } |
5427 | } | |
e73a347b AG |
5428 | |
5429 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
5430 | if (err) | |
5431 | goto error_return; | |
5432 | ||
5433 | /* | |
5434 | * need to update group_info->bb_free and bitmap | |
5435 | * with group lock held. generate_buddy look at | |
5436 | * them with group lock_held | |
5437 | */ | |
2846e820 | 5438 | ext4_lock_group(sb, block_group); |
d77147ff | 5439 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
5440 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
5441 | free_clusters_count = clusters_freed + | |
5442 | ext4_free_group_clusters(sb, desc); | |
5443 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
79f1ba49 | 5444 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 5445 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 5446 | ext4_unlock_group(sb, block_group); |
57042651 | 5447 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 5448 | clusters_freed); |
2846e820 AG |
5449 | |
5450 | if (sbi->s_log_groups_per_flex) { | |
5451 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 5452 | atomic64_add(clusters_freed, |
7c990728 SJS |
5453 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
5454 | flex_group)->free_clusters); | |
2846e820 | 5455 | } |
e73a347b AG |
5456 | |
5457 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
5458 | |
5459 | /* We dirtied the bitmap block */ | |
5460 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5461 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5462 | ||
5463 | /* And the group descriptor block */ | |
5464 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
5465 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
5466 | if (!err) | |
5467 | err = ret; | |
5468 | ||
5469 | error_return: | |
5470 | brelse(bitmap_bh); | |
5471 | ext4_std_error(sb, err); | |
cc7365df | 5472 | return err; |
2846e820 AG |
5473 | } |
5474 | ||
7360d173 LC |
5475 | /** |
5476 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
5477 | * @sb: super block for the file system | |
5478 | * @start: starting block of the free extent in the alloc. group | |
5479 | * @count: number of blocks to TRIM | |
5480 | * @group: alloc. group we are working with | |
5481 | * @e4b: ext4 buddy for the group | |
5482 | * | |
5483 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
5484 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
5485 | * be called with under the group lock. | |
5486 | */ | |
d71c1ae2 | 5487 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
d9f34504 | 5488 | ext4_group_t group, struct ext4_buddy *e4b) |
e2cbd587 | 5489 | __releases(bitlock) |
5490 | __acquires(bitlock) | |
7360d173 LC |
5491 | { |
5492 | struct ext4_free_extent ex; | |
d71c1ae2 | 5493 | int ret = 0; |
7360d173 | 5494 | |
b3d4c2b1 TM |
5495 | trace_ext4_trim_extent(sb, group, start, count); |
5496 | ||
7360d173 LC |
5497 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
5498 | ||
5499 | ex.fe_start = start; | |
5500 | ex.fe_group = group; | |
5501 | ex.fe_len = count; | |
5502 | ||
5503 | /* | |
5504 | * Mark blocks used, so no one can reuse them while | |
5505 | * being trimmed. | |
5506 | */ | |
5507 | mb_mark_used(e4b, &ex); | |
5508 | ext4_unlock_group(sb, group); | |
a0154344 | 5509 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
5510 | ext4_lock_group(sb, group); |
5511 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 5512 | return ret; |
7360d173 LC |
5513 | } |
5514 | ||
5515 | /** | |
5516 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
5517 | * @sb: super block for file system | |
22612283 | 5518 | * @group: group to be trimmed |
7360d173 LC |
5519 | * @start: first group block to examine |
5520 | * @max: last group block to examine | |
5521 | * @minblocks: minimum extent block count | |
5522 | * | |
5523 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
5524 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
5525 | * the extent. | |
5526 | * | |
5527 | * | |
5528 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
5529 | * extents. When the free extent is found, mark it as used in group buddy | |
5530 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
5531 | * the group buddy bitmap. This is done until whole group is scanned. | |
5532 | */ | |
0b75a840 | 5533 | static ext4_grpblk_t |
78944086 LC |
5534 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
5535 | ext4_grpblk_t start, ext4_grpblk_t max, | |
5536 | ext4_grpblk_t minblocks) | |
7360d173 LC |
5537 | { |
5538 | void *bitmap; | |
169ddc3e | 5539 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 5540 | struct ext4_buddy e4b; |
d71c1ae2 | 5541 | int ret = 0; |
7360d173 | 5542 | |
b3d4c2b1 TM |
5543 | trace_ext4_trim_all_free(sb, group, start, max); |
5544 | ||
78944086 LC |
5545 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
5546 | if (ret) { | |
9651e6b2 KK |
5547 | ext4_warning(sb, "Error %d loading buddy information for %u", |
5548 | ret, group); | |
78944086 LC |
5549 | return ret; |
5550 | } | |
78944086 | 5551 | bitmap = e4b.bd_bitmap; |
28739eea LC |
5552 | |
5553 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
5554 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
5555 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
5556 | goto out; | |
5557 | ||
78944086 LC |
5558 | start = (e4b.bd_info->bb_first_free > start) ? |
5559 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 5560 | |
913eed83 LC |
5561 | while (start <= max) { |
5562 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
5563 | if (start > max) | |
7360d173 | 5564 | break; |
913eed83 | 5565 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
5566 | |
5567 | if ((next - start) >= minblocks) { | |
d71c1ae2 LC |
5568 | ret = ext4_trim_extent(sb, start, |
5569 | next - start, group, &e4b); | |
5570 | if (ret && ret != -EOPNOTSUPP) | |
5571 | break; | |
5572 | ret = 0; | |
7360d173 LC |
5573 | count += next - start; |
5574 | } | |
169ddc3e | 5575 | free_count += next - start; |
7360d173 LC |
5576 | start = next + 1; |
5577 | ||
5578 | if (fatal_signal_pending(current)) { | |
5579 | count = -ERESTARTSYS; | |
5580 | break; | |
5581 | } | |
5582 | ||
5583 | if (need_resched()) { | |
5584 | ext4_unlock_group(sb, group); | |
5585 | cond_resched(); | |
5586 | ext4_lock_group(sb, group); | |
5587 | } | |
5588 | ||
169ddc3e | 5589 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
5590 | break; |
5591 | } | |
3d56b8d2 | 5592 | |
d71c1ae2 LC |
5593 | if (!ret) { |
5594 | ret = count; | |
3d56b8d2 | 5595 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 5596 | } |
3d56b8d2 | 5597 | out: |
7360d173 | 5598 | ext4_unlock_group(sb, group); |
78944086 | 5599 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
5600 | |
5601 | ext4_debug("trimmed %d blocks in the group %d\n", | |
5602 | count, group); | |
5603 | ||
d71c1ae2 | 5604 | return ret; |
7360d173 LC |
5605 | } |
5606 | ||
5607 | /** | |
5608 | * ext4_trim_fs() -- trim ioctl handle function | |
5609 | * @sb: superblock for filesystem | |
5610 | * @range: fstrim_range structure | |
5611 | * | |
5612 | * start: First Byte to trim | |
5613 | * len: number of Bytes to trim from start | |
5614 | * minlen: minimum extent length in Bytes | |
5615 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
5616 | * start to start+len. For each such a group ext4_trim_all_free function | |
5617 | * is invoked to trim all free space. | |
5618 | */ | |
5619 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
5620 | { | |
78944086 | 5621 | struct ext4_group_info *grp; |
913eed83 | 5622 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 5623 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 5624 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
5625 | ext4_fsblk_t first_data_blk = |
5626 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 5627 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
5628 | int ret = 0; |
5629 | ||
5630 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 5631 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
5632 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
5633 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 5634 | |
5de35e8d LC |
5635 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
5636 | start >= max_blks || | |
5637 | range->len < sb->s_blocksize) | |
7360d173 | 5638 | return -EINVAL; |
913eed83 LC |
5639 | if (end >= max_blks) |
5640 | end = max_blks - 1; | |
5641 | if (end <= first_data_blk) | |
22f10457 | 5642 | goto out; |
913eed83 | 5643 | if (start < first_data_blk) |
0f0a25bf | 5644 | start = first_data_blk; |
7360d173 | 5645 | |
913eed83 | 5646 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5647 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5648 | &first_group, &first_cluster); |
913eed83 | 5649 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5650 | &last_group, &last_cluster); |
7360d173 | 5651 | |
913eed83 LC |
5652 | /* end now represents the last cluster to discard in this group */ |
5653 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5654 | |
5655 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5656 | grp = ext4_get_group_info(sb, group); |
5657 | /* We only do this if the grp has never been initialized */ | |
5658 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 5659 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
5660 | if (ret) |
5661 | break; | |
7360d173 LC |
5662 | } |
5663 | ||
0ba08517 | 5664 | /* |
913eed83 LC |
5665 | * For all the groups except the last one, last cluster will |
5666 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5667 | * change it for the last group, note that last_cluster is | |
5668 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5669 | */ |
913eed83 LC |
5670 | if (group == last_group) |
5671 | end = last_cluster; | |
7360d173 | 5672 | |
78944086 | 5673 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5674 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 5675 | end, minlen); |
7360d173 LC |
5676 | if (cnt < 0) { |
5677 | ret = cnt; | |
7360d173 LC |
5678 | break; |
5679 | } | |
21e7fd22 | 5680 | trimmed += cnt; |
7360d173 | 5681 | } |
913eed83 LC |
5682 | |
5683 | /* | |
5684 | * For every group except the first one, we are sure | |
5685 | * that the first cluster to discard will be cluster #0. | |
5686 | */ | |
7137d7a4 | 5687 | first_cluster = 0; |
7360d173 | 5688 | } |
7360d173 | 5689 | |
3d56b8d2 TM |
5690 | if (!ret) |
5691 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5692 | ||
22f10457 | 5693 | out: |
aaf7d73e | 5694 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5695 | return ret; |
5696 | } | |
0c9ec4be DW |
5697 | |
5698 | /* Iterate all the free extents in the group. */ | |
5699 | int | |
5700 | ext4_mballoc_query_range( | |
5701 | struct super_block *sb, | |
5702 | ext4_group_t group, | |
5703 | ext4_grpblk_t start, | |
5704 | ext4_grpblk_t end, | |
5705 | ext4_mballoc_query_range_fn formatter, | |
5706 | void *priv) | |
5707 | { | |
5708 | void *bitmap; | |
5709 | ext4_grpblk_t next; | |
5710 | struct ext4_buddy e4b; | |
5711 | int error; | |
5712 | ||
5713 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
5714 | if (error) | |
5715 | return error; | |
5716 | bitmap = e4b.bd_bitmap; | |
5717 | ||
5718 | ext4_lock_group(sb, group); | |
5719 | ||
5720 | start = (e4b.bd_info->bb_first_free > start) ? | |
5721 | e4b.bd_info->bb_first_free : start; | |
5722 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
5723 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
5724 | ||
5725 | while (start <= end) { | |
5726 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
5727 | if (start > end) | |
5728 | break; | |
5729 | next = mb_find_next_bit(bitmap, end + 1, start); | |
5730 | ||
5731 | ext4_unlock_group(sb, group); | |
5732 | error = formatter(sb, group, start, next - start, priv); | |
5733 | if (error) | |
5734 | goto out_unload; | |
5735 | ext4_lock_group(sb, group); | |
5736 | ||
5737 | start = next + 1; | |
5738 | } | |
5739 | ||
5740 | ext4_unlock_group(sb, group); | |
5741 | out_unload: | |
5742 | ext4_mb_unload_buddy(&e4b); | |
5743 | ||
5744 | return error; | |
5745 | } |