]>
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 | * |
196e402a HS |
130 | * If "mb_optimize_scan" mount option is set, we maintain in memory group info |
131 | * structures in two data structures: | |
132 | * | |
133 | * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders) | |
134 | * | |
135 | * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks) | |
136 | * | |
137 | * This is an array of lists where the index in the array represents the | |
138 | * largest free order in the buddy bitmap of the participating group infos of | |
139 | * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total | |
140 | * number of buddy bitmap orders possible) number of lists. Group-infos are | |
141 | * placed in appropriate lists. | |
142 | * | |
83e80a6e | 143 | * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size) |
196e402a | 144 | * |
83e80a6e | 145 | * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks) |
196e402a | 146 | * |
83e80a6e JK |
147 | * This is an array of lists where in the i-th list there are groups with |
148 | * average fragment size >= 2^i and < 2^(i+1). The average fragment size | |
149 | * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments. | |
150 | * Note that we don't bother with a special list for completely empty groups | |
151 | * so we only have MB_NUM_ORDERS(sb) lists. | |
196e402a HS |
152 | * |
153 | * When "mb_optimize_scan" mount option is set, mballoc consults the above data | |
154 | * structures to decide the order in which groups are to be traversed for | |
155 | * fulfilling an allocation request. | |
156 | * | |
157 | * At CR = 0, we look for groups which have the largest_free_order >= the order | |
158 | * of the request. We directly look at the largest free order list in the data | |
159 | * structure (1) above where largest_free_order = order of the request. If that | |
160 | * list is empty, we look at remaining list in the increasing order of | |
161 | * largest_free_order. This allows us to perform CR = 0 lookup in O(1) time. | |
162 | * | |
163 | * At CR = 1, we only consider groups where average fragment size > request | |
164 | * size. So, we lookup a group which has average fragment size just above or | |
83e80a6e JK |
165 | * equal to request size using our average fragment size group lists (data |
166 | * structure 2) in O(1) time. | |
196e402a HS |
167 | * |
168 | * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in | |
169 | * linear order which requires O(N) search time for each CR 0 and CR 1 phase. | |
170 | * | |
d7a1fee1 | 171 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 172 | * |
b713a5ec TT |
173 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
174 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
175 | * /sys/fs/ext4/<partition>/mb_order2_req | |
196e402a | 176 | * /sys/fs/ext4/<partition>/mb_linear_limit |
c9de560d | 177 | * |
b713a5ec | 178 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
179 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
180 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 181 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 182 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
183 | * stripe size. This should result in better allocation on RAID setups. If |
184 | * not, we search in the specific group using bitmap for best extents. The | |
185 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 186 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 187 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
188 | * best extent in the found extents. Searching for the blocks starts with |
189 | * the group specified as the goal value in allocation context via | |
190 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 191 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
192 | * checked. |
193 | * | |
196e402a HS |
194 | * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not |
195 | * get traversed linearly. That may result in subsequent allocations being not | |
196 | * close to each other. And so, the underlying device may get filled up in a | |
197 | * non-linear fashion. While that may not matter on non-rotational devices, for | |
198 | * rotational devices that may result in higher seek times. "mb_linear_limit" | |
199 | * tells mballoc how many groups mballoc should search linearly before | |
200 | * performing consulting above data structures for more efficient lookups. For | |
201 | * non rotational devices, this value defaults to 0 and for rotational devices | |
202 | * this is set to MB_DEFAULT_LINEAR_LIMIT. | |
203 | * | |
c9de560d AT |
204 | * Both the prealloc space are getting populated as above. So for the first |
205 | * request we will hit the buddy cache which will result in this prealloc | |
206 | * space getting filled. The prealloc space is then later used for the | |
207 | * subsequent request. | |
208 | */ | |
209 | ||
210 | /* | |
211 | * mballoc operates on the following data: | |
212 | * - on-disk bitmap | |
213 | * - in-core buddy (actually includes buddy and bitmap) | |
214 | * - preallocation descriptors (PAs) | |
215 | * | |
216 | * there are two types of preallocations: | |
217 | * - inode | |
218 | * assiged to specific inode and can be used for this inode only. | |
219 | * it describes part of inode's space preallocated to specific | |
220 | * physical blocks. any block from that preallocated can be used | |
221 | * independent. the descriptor just tracks number of blocks left | |
222 | * unused. so, before taking some block from descriptor, one must | |
223 | * make sure corresponded logical block isn't allocated yet. this | |
224 | * also means that freeing any block within descriptor's range | |
225 | * must discard all preallocated blocks. | |
226 | * - locality group | |
227 | * assigned to specific locality group which does not translate to | |
228 | * permanent set of inodes: inode can join and leave group. space | |
229 | * from this type of preallocation can be used for any inode. thus | |
230 | * it's consumed from the beginning to the end. | |
231 | * | |
232 | * relation between them can be expressed as: | |
233 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
234 | * | |
235 | * this mean blocks mballoc considers used are: | |
236 | * - allocated blocks (persistent) | |
237 | * - preallocated blocks (non-persistent) | |
238 | * | |
239 | * consistency in mballoc world means that at any time a block is either | |
240 | * free or used in ALL structures. notice: "any time" should not be read | |
241 | * literally -- time is discrete and delimited by locks. | |
242 | * | |
243 | * to keep it simple, we don't use block numbers, instead we count number of | |
244 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
245 | * | |
246 | * all operations can be expressed as: | |
247 | * - init buddy: buddy = on-disk + PAs | |
248 | * - new PA: buddy += N; PA = N | |
249 | * - use inode PA: on-disk += N; PA -= N | |
250 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
251 | * - use locality group PA on-disk += N; PA -= N | |
252 | * - discard locality group PA buddy -= PA; PA = 0 | |
253 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
254 | * is used in real operation because we can't know actual used | |
255 | * bits from PA, only from on-disk bitmap | |
256 | * | |
257 | * if we follow this strict logic, then all operations above should be atomic. | |
258 | * given some of them can block, we'd have to use something like semaphores | |
259 | * killing performance on high-end SMP hardware. let's try to relax it using | |
260 | * the following knowledge: | |
261 | * 1) if buddy is referenced, it's already initialized | |
262 | * 2) while block is used in buddy and the buddy is referenced, | |
263 | * nobody can re-allocate that block | |
264 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
265 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
266 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
267 | * block | |
268 | * | |
269 | * so, now we're building a concurrency table: | |
270 | * - init buddy vs. | |
271 | * - new PA | |
272 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
273 | * until PA is linked to allocation group to avoid concurrent buddy init | |
274 | * - use inode PA | |
275 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
276 | * given (3) we care that PA-=N operation doesn't interfere with init | |
277 | * - discard inode PA | |
278 | * the simplest way would be to have buddy initialized by the discard | |
279 | * - use locality group PA | |
280 | * again PA-=N must be serialized with init | |
281 | * - discard locality group PA | |
282 | * the simplest way would be to have buddy initialized by the discard | |
283 | * - new PA vs. | |
284 | * - use inode PA | |
285 | * i_data_sem serializes them | |
286 | * - discard inode PA | |
287 | * discard process must wait until PA isn't used by another process | |
288 | * - use locality group PA | |
289 | * some mutex should serialize them | |
290 | * - discard locality group PA | |
291 | * discard process must wait until PA isn't used by another process | |
292 | * - use inode PA | |
293 | * - use inode PA | |
294 | * i_data_sem or another mutex should serializes them | |
295 | * - discard inode PA | |
296 | * discard process must wait until PA isn't used by another process | |
297 | * - use locality group PA | |
298 | * nothing wrong here -- they're different PAs covering different blocks | |
299 | * - discard locality group PA | |
300 | * discard process must wait until PA isn't used by another process | |
301 | * | |
302 | * now we're ready to make few consequences: | |
303 | * - PA is referenced and while it is no discard is possible | |
304 | * - PA is referenced until block isn't marked in on-disk bitmap | |
305 | * - PA changes only after on-disk bitmap | |
306 | * - discard must not compete with init. either init is done before | |
307 | * any discard or they're serialized somehow | |
308 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
309 | * | |
310 | * a special case when we've used PA to emptiness. no need to modify buddy | |
311 | * in this case, but we should care about concurrent init | |
312 | * | |
313 | */ | |
314 | ||
315 | /* | |
316 | * Logic in few words: | |
317 | * | |
318 | * - allocation: | |
319 | * load group | |
320 | * find blocks | |
321 | * mark bits in on-disk bitmap | |
322 | * release group | |
323 | * | |
324 | * - use preallocation: | |
325 | * find proper PA (per-inode or group) | |
326 | * load group | |
327 | * mark bits in on-disk bitmap | |
328 | * release group | |
329 | * release PA | |
330 | * | |
331 | * - free: | |
332 | * load group | |
333 | * mark bits in on-disk bitmap | |
334 | * release group | |
335 | * | |
336 | * - discard preallocations in group: | |
337 | * mark PAs deleted | |
338 | * move them onto local list | |
339 | * load on-disk bitmap | |
340 | * load group | |
341 | * remove PA from object (inode or locality group) | |
342 | * mark free blocks in-core | |
343 | * | |
344 | * - discard inode's preallocations: | |
345 | */ | |
346 | ||
347 | /* | |
348 | * Locking rules | |
349 | * | |
350 | * Locks: | |
351 | * - bitlock on a group (group) | |
352 | * - object (inode/locality) (object) | |
353 | * - per-pa lock (pa) | |
196e402a HS |
354 | * - cr0 lists lock (cr0) |
355 | * - cr1 tree lock (cr1) | |
c9de560d AT |
356 | * |
357 | * Paths: | |
358 | * - new pa | |
359 | * object | |
360 | * group | |
361 | * | |
362 | * - find and use pa: | |
363 | * pa | |
364 | * | |
365 | * - release consumed pa: | |
366 | * pa | |
367 | * group | |
368 | * object | |
369 | * | |
370 | * - generate in-core bitmap: | |
371 | * group | |
372 | * pa | |
373 | * | |
374 | * - discard all for given object (inode, locality group): | |
375 | * object | |
376 | * pa | |
377 | * group | |
378 | * | |
379 | * - discard all for given group: | |
380 | * group | |
381 | * pa | |
382 | * group | |
383 | * object | |
384 | * | |
196e402a HS |
385 | * - allocation path (ext4_mb_regular_allocator) |
386 | * group | |
387 | * cr0/cr1 | |
c9de560d | 388 | */ |
c3a326a6 AK |
389 | static struct kmem_cache *ext4_pspace_cachep; |
390 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 391 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
392 | |
393 | /* We create slab caches for groupinfo data structures based on the | |
394 | * superblock block size. There will be one per mounted filesystem for | |
395 | * each unique s_blocksize_bits */ | |
2892c15d | 396 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
397 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
398 | ||
d6006186 | 399 | static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
2892c15d ES |
400 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", |
401 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
402 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
403 | }; | |
404 | ||
c3a326a6 AK |
405 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
406 | ext4_group_t group); | |
7a2fcbf7 AK |
407 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
408 | ext4_group_t group); | |
53f86b17 | 409 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); |
c3a326a6 | 410 | |
196e402a HS |
411 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
412 | ext4_group_t group, int cr); | |
413 | ||
55cdd0af WJ |
414 | static int ext4_try_to_trim_range(struct super_block *sb, |
415 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
416 | ext4_grpblk_t max, ext4_grpblk_t minblocks); | |
417 | ||
07b5b8e1 RH |
418 | /* |
419 | * The algorithm using this percpu seq counter goes below: | |
420 | * 1. We sample the percpu discard_pa_seq counter before trying for block | |
421 | * allocation in ext4_mb_new_blocks(). | |
422 | * 2. We increment this percpu discard_pa_seq counter when we either allocate | |
423 | * or free these blocks i.e. while marking those blocks as used/free in | |
424 | * mb_mark_used()/mb_free_blocks(). | |
425 | * 3. We also increment this percpu seq counter when we successfully identify | |
426 | * that the bb_prealloc_list is not empty and hence proceed for discarding | |
427 | * of those PAs inside ext4_mb_discard_group_preallocations(). | |
428 | * | |
429 | * Now to make sure that the regular fast path of block allocation is not | |
430 | * affected, as a small optimization we only sample the percpu seq counter | |
431 | * on that cpu. Only when the block allocation fails and when freed blocks | |
432 | * found were 0, that is when we sample percpu seq counter for all cpus using | |
433 | * below function ext4_get_discard_pa_seq_sum(). This happens after making | |
434 | * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. | |
435 | */ | |
436 | static DEFINE_PER_CPU(u64, discard_pa_seq); | |
437 | static inline u64 ext4_get_discard_pa_seq_sum(void) | |
438 | { | |
439 | int __cpu; | |
440 | u64 __seq = 0; | |
441 | ||
442 | for_each_possible_cpu(__cpu) | |
443 | __seq += per_cpu(discard_pa_seq, __cpu); | |
444 | return __seq; | |
445 | } | |
446 | ||
ffad0a44 AK |
447 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
448 | { | |
c9de560d | 449 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
450 | *bit += ((unsigned long) addr & 7UL) << 3; |
451 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 452 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
453 | *bit += ((unsigned long) addr & 3UL) << 3; |
454 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
455 | #else |
456 | #error "how many bits you are?!" | |
457 | #endif | |
ffad0a44 AK |
458 | return addr; |
459 | } | |
c9de560d AT |
460 | |
461 | static inline int mb_test_bit(int bit, void *addr) | |
462 | { | |
463 | /* | |
464 | * ext4_test_bit on architecture like powerpc | |
465 | * needs unsigned long aligned address | |
466 | */ | |
ffad0a44 | 467 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
468 | return ext4_test_bit(bit, addr); |
469 | } | |
470 | ||
471 | static inline void mb_set_bit(int bit, void *addr) | |
472 | { | |
ffad0a44 | 473 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
474 | ext4_set_bit(bit, addr); |
475 | } | |
476 | ||
c9de560d AT |
477 | static inline void mb_clear_bit(int bit, void *addr) |
478 | { | |
ffad0a44 | 479 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
480 | ext4_clear_bit(bit, addr); |
481 | } | |
482 | ||
eabe0444 AS |
483 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
484 | { | |
485 | addr = mb_correct_addr_and_bit(&bit, addr); | |
486 | return ext4_test_and_clear_bit(bit, addr); | |
487 | } | |
488 | ||
ffad0a44 AK |
489 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
490 | { | |
e7dfb246 | 491 | int fix = 0, ret, tmpmax; |
ffad0a44 | 492 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 493 | tmpmax = max + fix; |
ffad0a44 AK |
494 | start += fix; |
495 | ||
e7dfb246 AK |
496 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
497 | if (ret > max) | |
498 | return max; | |
499 | return ret; | |
ffad0a44 AK |
500 | } |
501 | ||
502 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
503 | { | |
e7dfb246 | 504 | int fix = 0, ret, tmpmax; |
ffad0a44 | 505 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 506 | tmpmax = max + fix; |
ffad0a44 AK |
507 | start += fix; |
508 | ||
e7dfb246 AK |
509 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
510 | if (ret > max) | |
511 | return max; | |
512 | return ret; | |
ffad0a44 AK |
513 | } |
514 | ||
c9de560d AT |
515 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
516 | { | |
517 | char *bb; | |
518 | ||
c5e8f3f3 | 519 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
520 | BUG_ON(max == NULL); |
521 | ||
522 | if (order > e4b->bd_blkbits + 1) { | |
523 | *max = 0; | |
524 | return NULL; | |
525 | } | |
526 | ||
527 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
528 | if (order == 0) { |
529 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 530 | return e4b->bd_bitmap; |
84b775a3 | 531 | } |
c9de560d | 532 | |
c5e8f3f3 | 533 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
534 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
535 | ||
536 | return bb; | |
537 | } | |
538 | ||
539 | #ifdef DOUBLE_CHECK | |
540 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
541 | int first, int count) | |
542 | { | |
543 | int i; | |
544 | struct super_block *sb = e4b->bd_sb; | |
545 | ||
546 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
547 | return; | |
bc8e6740 | 548 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
549 | for (i = 0; i < count; i++) { |
550 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
551 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
552 | |
553 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 554 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 555 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
556 | inode ? inode->i_ino : 0, |
557 | blocknr, | |
558 | "freeing block already freed " | |
559 | "(bit %u)", | |
560 | first + i); | |
736dedbb WS |
561 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
562 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
563 | } |
564 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
565 | } | |
566 | } | |
567 | ||
568 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
569 | { | |
570 | int i; | |
571 | ||
572 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
573 | return; | |
bc8e6740 | 574 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
575 | for (i = 0; i < count; i++) { |
576 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
577 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
578 | } | |
579 | } | |
580 | ||
581 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
582 | { | |
eb2b8ebb RH |
583 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
584 | return; | |
c9de560d AT |
585 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { |
586 | unsigned char *b1, *b2; | |
587 | int i; | |
588 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
589 | b2 = (unsigned char *) bitmap; | |
590 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
591 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
592 | ext4_msg(e4b->bd_sb, KERN_ERR, |
593 | "corruption in group %u " | |
594 | "at byte %u(%u): %x in copy != %x " | |
595 | "on disk/prealloc", | |
596 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
597 | BUG(); |
598 | } | |
599 | } | |
600 | } | |
601 | } | |
602 | ||
a3450215 RH |
603 | static void mb_group_bb_bitmap_alloc(struct super_block *sb, |
604 | struct ext4_group_info *grp, ext4_group_t group) | |
605 | { | |
606 | struct buffer_head *bh; | |
607 | ||
608 | grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); | |
eb2b8ebb RH |
609 | if (!grp->bb_bitmap) |
610 | return; | |
a3450215 RH |
611 | |
612 | bh = ext4_read_block_bitmap(sb, group); | |
eb2b8ebb RH |
613 | if (IS_ERR_OR_NULL(bh)) { |
614 | kfree(grp->bb_bitmap); | |
615 | grp->bb_bitmap = NULL; | |
616 | return; | |
617 | } | |
a3450215 RH |
618 | |
619 | memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); | |
620 | put_bh(bh); | |
621 | } | |
622 | ||
623 | static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
624 | { | |
625 | kfree(grp->bb_bitmap); | |
626 | } | |
627 | ||
c9de560d AT |
628 | #else |
629 | static inline void mb_free_blocks_double(struct inode *inode, | |
630 | struct ext4_buddy *e4b, int first, int count) | |
631 | { | |
632 | return; | |
633 | } | |
634 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
635 | int first, int count) | |
636 | { | |
637 | return; | |
638 | } | |
639 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
640 | { | |
641 | return; | |
642 | } | |
a3450215 RH |
643 | |
644 | static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, | |
645 | struct ext4_group_info *grp, ext4_group_t group) | |
646 | { | |
647 | return; | |
648 | } | |
649 | ||
650 | static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
651 | { | |
652 | return; | |
653 | } | |
c9de560d AT |
654 | #endif |
655 | ||
656 | #ifdef AGGRESSIVE_CHECK | |
657 | ||
658 | #define MB_CHECK_ASSERT(assert) \ | |
659 | do { \ | |
660 | if (!(assert)) { \ | |
661 | printk(KERN_EMERG \ | |
662 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
663 | function, file, line, # assert); \ | |
664 | BUG(); \ | |
665 | } \ | |
666 | } while (0) | |
667 | ||
668 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
669 | const char *function, int line) | |
670 | { | |
671 | struct super_block *sb = e4b->bd_sb; | |
672 | int order = e4b->bd_blkbits + 1; | |
673 | int max; | |
674 | int max2; | |
675 | int i; | |
676 | int j; | |
677 | int k; | |
678 | int count; | |
679 | struct ext4_group_info *grp; | |
680 | int fragments = 0; | |
681 | int fstart; | |
682 | struct list_head *cur; | |
683 | void *buddy; | |
684 | void *buddy2; | |
685 | ||
addd752c CX |
686 | if (e4b->bd_info->bb_check_counter++ % 10) |
687 | return 0; | |
c9de560d AT |
688 | |
689 | while (order > 1) { | |
690 | buddy = mb_find_buddy(e4b, order, &max); | |
691 | MB_CHECK_ASSERT(buddy); | |
692 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
693 | MB_CHECK_ASSERT(buddy2); | |
694 | MB_CHECK_ASSERT(buddy != buddy2); | |
695 | MB_CHECK_ASSERT(max * 2 == max2); | |
696 | ||
697 | count = 0; | |
698 | for (i = 0; i < max; i++) { | |
699 | ||
700 | if (mb_test_bit(i, buddy)) { | |
af2b3275 | 701 | /* only single bit in buddy2 may be 0 */ |
c9de560d AT |
702 | if (!mb_test_bit(i << 1, buddy2)) { |
703 | MB_CHECK_ASSERT( | |
704 | mb_test_bit((i<<1)+1, buddy2)); | |
c9de560d AT |
705 | } |
706 | continue; | |
707 | } | |
708 | ||
0a10da73 | 709 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
710 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
711 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
712 | ||
713 | for (j = 0; j < (1 << order); j++) { | |
714 | k = (i * (1 << order)) + j; | |
715 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 716 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
717 | } |
718 | count++; | |
719 | } | |
720 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
721 | order--; | |
722 | } | |
723 | ||
724 | fstart = -1; | |
725 | buddy = mb_find_buddy(e4b, 0, &max); | |
726 | for (i = 0; i < max; i++) { | |
727 | if (!mb_test_bit(i, buddy)) { | |
728 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
729 | if (fstart == -1) { | |
730 | fragments++; | |
731 | fstart = i; | |
732 | } | |
733 | continue; | |
734 | } | |
735 | fstart = -1; | |
736 | /* check used bits only */ | |
737 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
738 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
739 | k = i >> j; | |
740 | MB_CHECK_ASSERT(k < max2); | |
741 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
742 | } | |
743 | } | |
744 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
745 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
746 | ||
747 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
748 | list_for_each(cur, &grp->bb_prealloc_list) { |
749 | ext4_group_t groupnr; | |
750 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
751 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
752 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 753 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 754 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
755 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
756 | } | |
757 | return 0; | |
758 | } | |
759 | #undef MB_CHECK_ASSERT | |
760 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 761 | __FILE__, __func__, __LINE__) |
c9de560d AT |
762 | #else |
763 | #define mb_check_buddy(e4b) | |
764 | #endif | |
765 | ||
7c786059 CL |
766 | /* |
767 | * Divide blocks started from @first with length @len into | |
768 | * smaller chunks with power of 2 blocks. | |
769 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
770 | * then increase bb_counters[] for corresponded chunk size. | |
771 | */ | |
c9de560d | 772 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 773 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
774 | struct ext4_group_info *grp) |
775 | { | |
776 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
777 | ext4_grpblk_t min; |
778 | ext4_grpblk_t max; | |
779 | ext4_grpblk_t chunk; | |
69e43e8c | 780 | unsigned int border; |
c9de560d | 781 | |
7137d7a4 | 782 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
783 | |
784 | border = 2 << sb->s_blocksize_bits; | |
785 | ||
786 | while (len > 0) { | |
787 | /* find how many blocks can be covered since this position */ | |
788 | max = ffs(first | border) - 1; | |
789 | ||
790 | /* find how many blocks of power 2 we need to mark */ | |
791 | min = fls(len) - 1; | |
792 | ||
793 | if (max < min) | |
794 | min = max; | |
795 | chunk = 1 << min; | |
796 | ||
797 | /* mark multiblock chunks only */ | |
798 | grp->bb_counters[min]++; | |
799 | if (min > 0) | |
800 | mb_clear_bit(first >> min, | |
801 | buddy + sbi->s_mb_offsets[min]); | |
802 | ||
803 | len -= chunk; | |
804 | first += chunk; | |
805 | } | |
806 | } | |
807 | ||
83e80a6e | 808 | static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) |
196e402a | 809 | { |
83e80a6e | 810 | int order; |
196e402a | 811 | |
83e80a6e JK |
812 | /* |
813 | * We don't bother with a special lists groups with only 1 block free | |
814 | * extents and for completely empty groups. | |
815 | */ | |
816 | order = fls(len) - 2; | |
817 | if (order < 0) | |
818 | return 0; | |
819 | if (order == MB_NUM_ORDERS(sb)) | |
820 | order--; | |
821 | return order; | |
196e402a HS |
822 | } |
823 | ||
83e80a6e | 824 | /* Move group to appropriate avg_fragment_size list */ |
196e402a HS |
825 | static void |
826 | mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) | |
827 | { | |
828 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
83e80a6e | 829 | int new_order; |
196e402a HS |
830 | |
831 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) | |
832 | return; | |
833 | ||
83e80a6e JK |
834 | new_order = mb_avg_fragment_size_order(sb, |
835 | grp->bb_free / grp->bb_fragments); | |
836 | if (new_order == grp->bb_avg_fragment_size_order) | |
837 | return; | |
196e402a | 838 | |
83e80a6e JK |
839 | if (grp->bb_avg_fragment_size_order != -1) { |
840 | write_lock(&sbi->s_mb_avg_fragment_size_locks[ | |
841 | grp->bb_avg_fragment_size_order]); | |
842 | list_del(&grp->bb_avg_fragment_size_node); | |
843 | write_unlock(&sbi->s_mb_avg_fragment_size_locks[ | |
844 | grp->bb_avg_fragment_size_order]); | |
845 | } | |
846 | grp->bb_avg_fragment_size_order = new_order; | |
847 | write_lock(&sbi->s_mb_avg_fragment_size_locks[ | |
848 | grp->bb_avg_fragment_size_order]); | |
849 | list_add_tail(&grp->bb_avg_fragment_size_node, | |
850 | &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]); | |
851 | write_unlock(&sbi->s_mb_avg_fragment_size_locks[ | |
852 | grp->bb_avg_fragment_size_order]); | |
196e402a HS |
853 | } |
854 | ||
855 | /* | |
856 | * Choose next group by traversing largest_free_order lists. Updates *new_cr if | |
857 | * cr level needs an update. | |
858 | */ | |
859 | static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, | |
860 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
861 | { | |
862 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
863 | struct ext4_group_info *iter, *grp; | |
864 | int i; | |
865 | ||
866 | if (ac->ac_status == AC_STATUS_FOUND) | |
867 | return; | |
868 | ||
869 | if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED)) | |
870 | atomic_inc(&sbi->s_bal_cr0_bad_suggestions); | |
871 | ||
872 | grp = NULL; | |
873 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
874 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) | |
875 | continue; | |
876 | read_lock(&sbi->s_mb_largest_free_orders_locks[i]); | |
877 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) { | |
878 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
879 | continue; | |
880 | } | |
881 | grp = NULL; | |
882 | list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], | |
883 | bb_largest_free_order_node) { | |
884 | if (sbi->s_mb_stats) | |
885 | atomic64_inc(&sbi->s_bal_cX_groups_considered[0]); | |
886 | if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) { | |
887 | grp = iter; | |
888 | break; | |
889 | } | |
890 | } | |
891 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
892 | if (grp) | |
893 | break; | |
894 | } | |
895 | ||
896 | if (!grp) { | |
897 | /* Increment cr and search again */ | |
898 | *new_cr = 1; | |
899 | } else { | |
900 | *group = grp->bb_group; | |
196e402a HS |
901 | ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED; |
902 | } | |
903 | } | |
904 | ||
905 | /* | |
83e80a6e JK |
906 | * Choose next group by traversing average fragment size list of suitable |
907 | * order. Updates *new_cr if cr level needs an update. | |
196e402a HS |
908 | */ |
909 | static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, | |
910 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
911 | { | |
912 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
a078dff8 | 913 | struct ext4_group_info *grp = NULL, *iter; |
83e80a6e | 914 | int i; |
196e402a HS |
915 | |
916 | if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) { | |
917 | if (sbi->s_mb_stats) | |
918 | atomic_inc(&sbi->s_bal_cr1_bad_suggestions); | |
83e80a6e JK |
919 | } |
920 | ||
921 | for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); | |
922 | i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
923 | if (list_empty(&sbi->s_mb_avg_fragment_size[i])) | |
924 | continue; | |
925 | read_lock(&sbi->s_mb_avg_fragment_size_locks[i]); | |
926 | if (list_empty(&sbi->s_mb_avg_fragment_size[i])) { | |
927 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]); | |
928 | continue; | |
929 | } | |
83e80a6e JK |
930 | list_for_each_entry(iter, &sbi->s_mb_avg_fragment_size[i], |
931 | bb_avg_fragment_size_node) { | |
196e402a HS |
932 | if (sbi->s_mb_stats) |
933 | atomic64_inc(&sbi->s_bal_cX_groups_considered[1]); | |
83e80a6e JK |
934 | if (likely(ext4_mb_good_group(ac, iter->bb_group, 1))) { |
935 | grp = iter; | |
196e402a | 936 | break; |
196e402a HS |
937 | } |
938 | } | |
83e80a6e JK |
939 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]); |
940 | if (grp) | |
941 | break; | |
196e402a HS |
942 | } |
943 | ||
83e80a6e | 944 | if (grp) { |
196e402a HS |
945 | *group = grp->bb_group; |
946 | ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED; | |
947 | } else { | |
948 | *new_cr = 2; | |
949 | } | |
196e402a HS |
950 | } |
951 | ||
952 | static inline int should_optimize_scan(struct ext4_allocation_context *ac) | |
953 | { | |
954 | if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) | |
955 | return 0; | |
956 | if (ac->ac_criteria >= 2) | |
957 | return 0; | |
077d0c2c | 958 | if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) |
196e402a HS |
959 | return 0; |
960 | return 1; | |
961 | } | |
962 | ||
963 | /* | |
964 | * Return next linear group for allocation. If linear traversal should not be | |
965 | * performed, this function just returns the same group | |
966 | */ | |
967 | static int | |
968 | next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups) | |
969 | { | |
970 | if (!should_optimize_scan(ac)) | |
971 | goto inc_and_return; | |
972 | ||
973 | if (ac->ac_groups_linear_remaining) { | |
974 | ac->ac_groups_linear_remaining--; | |
975 | goto inc_and_return; | |
976 | } | |
977 | ||
196e402a HS |
978 | return group; |
979 | inc_and_return: | |
980 | /* | |
981 | * Artificially restricted ngroups for non-extent | |
982 | * files makes group > ngroups possible on first loop. | |
983 | */ | |
984 | return group + 1 >= ngroups ? 0 : group + 1; | |
985 | } | |
986 | ||
987 | /* | |
988 | * ext4_mb_choose_next_group: choose next group for allocation. | |
989 | * | |
990 | * @ac Allocation Context | |
991 | * @new_cr This is an output parameter. If the there is no good group | |
992 | * available at current CR level, this field is updated to indicate | |
993 | * the new cr level that should be used. | |
994 | * @group This is an input / output parameter. As an input it indicates the | |
995 | * next group that the allocator intends to use for allocation. As | |
996 | * output, this field indicates the next group that should be used as | |
997 | * determined by the optimization functions. | |
998 | * @ngroups Total number of groups | |
999 | */ | |
1000 | static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, | |
1001 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
1002 | { | |
1003 | *new_cr = ac->ac_criteria; | |
1004 | ||
4fca50d4 JK |
1005 | if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) { |
1006 | *group = next_linear_group(ac, *group, ngroups); | |
196e402a | 1007 | return; |
4fca50d4 | 1008 | } |
196e402a HS |
1009 | |
1010 | if (*new_cr == 0) { | |
1011 | ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups); | |
1012 | } else if (*new_cr == 1) { | |
1013 | ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups); | |
1014 | } else { | |
1015 | /* | |
1016 | * TODO: For CR=2, we can arrange groups in an rb tree sorted by | |
1017 | * bb_free. But until that happens, we should never come here. | |
1018 | */ | |
1019 | WARN_ON(1); | |
1020 | } | |
1021 | } | |
1022 | ||
8a57d9d6 CW |
1023 | /* |
1024 | * Cache the order of the largest free extent we have available in this block | |
1025 | * group. | |
1026 | */ | |
1027 | static void | |
1028 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
1029 | { | |
196e402a | 1030 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
8a57d9d6 | 1031 | int i; |
8a57d9d6 | 1032 | |
1940265e JK |
1033 | for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) |
1034 | if (grp->bb_counters[i] > 0) | |
1035 | break; | |
1036 | /* No need to move between order lists? */ | |
1037 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || | |
1038 | i == grp->bb_largest_free_order) { | |
1039 | grp->bb_largest_free_order = i; | |
1040 | return; | |
1041 | } | |
1042 | ||
1043 | if (grp->bb_largest_free_order >= 0) { | |
196e402a HS |
1044 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1045 | grp->bb_largest_free_order]); | |
1046 | list_del_init(&grp->bb_largest_free_order_node); | |
1047 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1048 | grp->bb_largest_free_order]); | |
1049 | } | |
1940265e JK |
1050 | grp->bb_largest_free_order = i; |
1051 | if (grp->bb_largest_free_order >= 0 && grp->bb_free) { | |
196e402a HS |
1052 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1053 | grp->bb_largest_free_order]); | |
1054 | list_add_tail(&grp->bb_largest_free_order_node, | |
1055 | &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); | |
1056 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1057 | grp->bb_largest_free_order]); | |
1058 | } | |
8a57d9d6 CW |
1059 | } |
1060 | ||
089ceecc ES |
1061 | static noinline_for_stack |
1062 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
1063 | void *buddy, void *bitmap, ext4_group_t group) |
1064 | { | |
1065 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 1066 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 1067 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
1068 | ext4_grpblk_t i = 0; |
1069 | ext4_grpblk_t first; | |
1070 | ext4_grpblk_t len; | |
c9de560d AT |
1071 | unsigned free = 0; |
1072 | unsigned fragments = 0; | |
1073 | unsigned long long period = get_cycles(); | |
1074 | ||
1075 | /* initialize buddy from bitmap which is aggregation | |
1076 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 1077 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
1078 | grp->bb_first_free = i; |
1079 | while (i < max) { | |
1080 | fragments++; | |
1081 | first = i; | |
ffad0a44 | 1082 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
1083 | len = i - first; |
1084 | free += len; | |
1085 | if (len > 1) | |
1086 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
1087 | else | |
1088 | grp->bb_counters[0]++; | |
1089 | if (i < max) | |
ffad0a44 | 1090 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
1091 | } |
1092 | grp->bb_fragments = fragments; | |
1093 | ||
1094 | if (free != grp->bb_free) { | |
e29136f8 | 1095 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
1096 | "block bitmap and bg descriptor " |
1097 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 1098 | free, grp->bb_free); |
e56eb659 | 1099 | /* |
163a203d | 1100 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
1101 | * corrupt and update bb_free using bitmap value |
1102 | */ | |
c9de560d | 1103 | grp->bb_free = free; |
db79e6d1 WS |
1104 | ext4_mark_group_bitmap_corrupted(sb, group, |
1105 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 1106 | } |
8a57d9d6 | 1107 | mb_set_largest_free_order(sb, grp); |
83e80a6e | 1108 | mb_update_avg_fragment_size(sb, grp); |
c9de560d AT |
1109 | |
1110 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
1111 | ||
1112 | period = get_cycles() - period; | |
67d25186 HS |
1113 | atomic_inc(&sbi->s_mb_buddies_generated); |
1114 | atomic64_add(period, &sbi->s_mb_generation_time); | |
c9de560d AT |
1115 | } |
1116 | ||
1117 | /* The buddy information is attached the buddy cache inode | |
1118 | * for convenience. The information regarding each group | |
1119 | * is loaded via ext4_mb_load_buddy. The information involve | |
1120 | * block bitmap and buddy information. The information are | |
1121 | * stored in the inode as | |
1122 | * | |
1123 | * { page } | |
c3a326a6 | 1124 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
1125 | * |
1126 | * | |
1127 | * one block each for bitmap and buddy information. | |
1128 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 1129 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
1130 | * So it can have information regarding groups_per_page which |
1131 | * is blocks_per_page/2 | |
8a57d9d6 CW |
1132 | * |
1133 | * Locking note: This routine takes the block group lock of all groups | |
1134 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
1135 | */ |
1136 | ||
adb7ef60 | 1137 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 1138 | { |
8df9675f | 1139 | ext4_group_t ngroups; |
c9de560d AT |
1140 | int blocksize; |
1141 | int blocks_per_page; | |
1142 | int groups_per_page; | |
1143 | int err = 0; | |
1144 | int i; | |
813e5727 | 1145 | ext4_group_t first_group, group; |
c9de560d AT |
1146 | int first_block; |
1147 | struct super_block *sb; | |
1148 | struct buffer_head *bhs; | |
fa77dcfa | 1149 | struct buffer_head **bh = NULL; |
c9de560d AT |
1150 | struct inode *inode; |
1151 | char *data; | |
1152 | char *bitmap; | |
9b8b7d35 | 1153 | struct ext4_group_info *grinfo; |
c9de560d | 1154 | |
c9de560d AT |
1155 | inode = page->mapping->host; |
1156 | sb = inode->i_sb; | |
8df9675f | 1157 | ngroups = ext4_get_groups_count(sb); |
93407472 | 1158 | blocksize = i_blocksize(inode); |
09cbfeaf | 1159 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d | 1160 | |
d3df1453 RH |
1161 | mb_debug(sb, "init page %lu\n", page->index); |
1162 | ||
c9de560d AT |
1163 | groups_per_page = blocks_per_page >> 1; |
1164 | if (groups_per_page == 0) | |
1165 | groups_per_page = 1; | |
1166 | ||
1167 | /* allocate buffer_heads to read bitmaps */ | |
1168 | if (groups_per_page > 1) { | |
c9de560d | 1169 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 1170 | bh = kzalloc(i, gfp); |
813e5727 TT |
1171 | if (bh == NULL) { |
1172 | err = -ENOMEM; | |
c9de560d | 1173 | goto out; |
813e5727 | 1174 | } |
c9de560d AT |
1175 | } else |
1176 | bh = &bhs; | |
1177 | ||
1178 | first_group = page->index * blocks_per_page / 2; | |
1179 | ||
1180 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
1181 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
1182 | if (group >= ngroups) | |
c9de560d AT |
1183 | break; |
1184 | ||
813e5727 | 1185 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
1186 | /* |
1187 | * If page is uptodate then we came here after online resize | |
1188 | * which added some new uninitialized group info structs, so | |
1189 | * we must skip all initialized uptodate buddies on the page, | |
1190 | * which may be currently in use by an allocating task. | |
1191 | */ | |
1192 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
1193 | bh[i] = NULL; | |
1194 | continue; | |
1195 | } | |
cfd73237 | 1196 | bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
9008a58e DW |
1197 | if (IS_ERR(bh[i])) { |
1198 | err = PTR_ERR(bh[i]); | |
1199 | bh[i] = NULL; | |
c9de560d | 1200 | goto out; |
2ccb5fb9 | 1201 | } |
d3df1453 | 1202 | mb_debug(sb, "read bitmap for group %u\n", group); |
c9de560d AT |
1203 | } |
1204 | ||
1205 | /* wait for I/O completion */ | |
813e5727 | 1206 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
1207 | int err2; |
1208 | ||
1209 | if (!bh[i]) | |
1210 | continue; | |
1211 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
1212 | if (!err) | |
1213 | err = err2; | |
813e5727 | 1214 | } |
c9de560d AT |
1215 | |
1216 | first_block = page->index * blocks_per_page; | |
1217 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 1218 | group = (first_block + i) >> 1; |
8df9675f | 1219 | if (group >= ngroups) |
c9de560d AT |
1220 | break; |
1221 | ||
9b8b7d35 AG |
1222 | if (!bh[group - first_group]) |
1223 | /* skip initialized uptodate buddy */ | |
1224 | continue; | |
1225 | ||
bbdc322f LC |
1226 | if (!buffer_verified(bh[group - first_group])) |
1227 | /* Skip faulty bitmaps */ | |
1228 | continue; | |
1229 | err = 0; | |
1230 | ||
c9de560d AT |
1231 | /* |
1232 | * data carry information regarding this | |
1233 | * particular group in the format specified | |
1234 | * above | |
1235 | * | |
1236 | */ | |
1237 | data = page_address(page) + (i * blocksize); | |
1238 | bitmap = bh[group - first_group]->b_data; | |
1239 | ||
1240 | /* | |
1241 | * We place the buddy block and bitmap block | |
1242 | * close together | |
1243 | */ | |
1244 | if ((first_block + i) & 1) { | |
1245 | /* this is block of buddy */ | |
1246 | BUG_ON(incore == NULL); | |
d3df1453 | 1247 | mb_debug(sb, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 1248 | group, page->index, i * blocksize); |
f307333e | 1249 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
1250 | grinfo = ext4_get_group_info(sb, group); |
1251 | grinfo->bb_fragments = 0; | |
1252 | memset(grinfo->bb_counters, 0, | |
1927805e | 1253 | sizeof(*grinfo->bb_counters) * |
4b68f6df | 1254 | (MB_NUM_ORDERS(sb))); |
c9de560d AT |
1255 | /* |
1256 | * incore got set to the group block bitmap below | |
1257 | */ | |
7a2fcbf7 | 1258 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
1259 | /* init the buddy */ |
1260 | memset(data, 0xff, blocksize); | |
c9de560d | 1261 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 1262 | ext4_unlock_group(sb, group); |
c9de560d AT |
1263 | incore = NULL; |
1264 | } else { | |
1265 | /* this is block of bitmap */ | |
1266 | BUG_ON(incore != NULL); | |
d3df1453 | 1267 | mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 1268 | group, page->index, i * blocksize); |
f307333e | 1269 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
1270 | |
1271 | /* see comments in ext4_mb_put_pa() */ | |
1272 | ext4_lock_group(sb, group); | |
1273 | memcpy(data, bitmap, blocksize); | |
1274 | ||
1275 | /* mark all preallocated blks used in in-core bitmap */ | |
1276 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 1277 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
1278 | ext4_unlock_group(sb, group); |
1279 | ||
1280 | /* set incore so that the buddy information can be | |
1281 | * generated using this | |
1282 | */ | |
1283 | incore = data; | |
1284 | } | |
1285 | } | |
1286 | SetPageUptodate(page); | |
1287 | ||
1288 | out: | |
1289 | if (bh) { | |
9b8b7d35 | 1290 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
1291 | brelse(bh[i]); |
1292 | if (bh != &bhs) | |
1293 | kfree(bh); | |
1294 | } | |
1295 | return err; | |
1296 | } | |
1297 | ||
eee4adc7 | 1298 | /* |
2de8807b AG |
1299 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
1300 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1301 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
1302 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 1303 | */ |
2de8807b | 1304 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1305 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1306 | { |
2de8807b AG |
1307 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1308 | int block, pnum, poff; | |
eee4adc7 | 1309 | int blocks_per_page; |
2de8807b AG |
1310 | struct page *page; |
1311 | ||
1312 | e4b->bd_buddy_page = NULL; | |
1313 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 1314 | |
09cbfeaf | 1315 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1316 | /* |
1317 | * the buddy cache inode stores the block bitmap | |
1318 | * and buddy information in consecutive blocks. | |
1319 | * So for each group we need two blocks. | |
1320 | */ | |
1321 | block = group * 2; | |
1322 | pnum = block / blocks_per_page; | |
2de8807b | 1323 | poff = block % blocks_per_page; |
adb7ef60 | 1324 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1325 | if (!page) |
c57ab39b | 1326 | return -ENOMEM; |
2de8807b AG |
1327 | BUG_ON(page->mapping != inode->i_mapping); |
1328 | e4b->bd_bitmap_page = page; | |
1329 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1330 | ||
1331 | if (blocks_per_page >= 2) { | |
1332 | /* buddy and bitmap are on the same page */ | |
1333 | return 0; | |
eee4adc7 | 1334 | } |
2de8807b AG |
1335 | |
1336 | block++; | |
1337 | pnum = block / blocks_per_page; | |
adb7ef60 | 1338 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1339 | if (!page) |
c57ab39b | 1340 | return -ENOMEM; |
2de8807b AG |
1341 | BUG_ON(page->mapping != inode->i_mapping); |
1342 | e4b->bd_buddy_page = page; | |
1343 | return 0; | |
eee4adc7 ES |
1344 | } |
1345 | ||
2de8807b | 1346 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1347 | { |
2de8807b AG |
1348 | if (e4b->bd_bitmap_page) { |
1349 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1350 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1351 | } |
1352 | if (e4b->bd_buddy_page) { | |
1353 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1354 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1355 | } |
eee4adc7 ES |
1356 | } |
1357 | ||
8a57d9d6 CW |
1358 | /* |
1359 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1360 | * block group lock of all groups for this page; do not hold the BG lock when | |
1361 | * calling this routine! | |
1362 | */ | |
b6a758ec | 1363 | static noinline_for_stack |
adb7ef60 | 1364 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1365 | { |
1366 | ||
b6a758ec | 1367 | struct ext4_group_info *this_grp; |
2de8807b AG |
1368 | struct ext4_buddy e4b; |
1369 | struct page *page; | |
1370 | int ret = 0; | |
b6a758ec | 1371 | |
b10a44c3 | 1372 | might_sleep(); |
d3df1453 | 1373 | mb_debug(sb, "init group %u\n", group); |
b6a758ec AK |
1374 | this_grp = ext4_get_group_info(sb, group); |
1375 | /* | |
08c3a813 AK |
1376 | * This ensures that we don't reinit the buddy cache |
1377 | * page which map to the group from which we are already | |
1378 | * allocating. If we are looking at the buddy cache we would | |
1379 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1380 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1381 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1382 | * page accessed. | |
b6a758ec | 1383 | */ |
adb7ef60 | 1384 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1385 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1386 | /* |
1387 | * somebody initialized the group | |
1388 | * return without doing anything | |
1389 | */ | |
b6a758ec AK |
1390 | goto err; |
1391 | } | |
2de8807b AG |
1392 | |
1393 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1394 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1395 | if (ret) |
1396 | goto err; | |
1397 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1398 | ret = -EIO; |
1399 | goto err; | |
1400 | } | |
b6a758ec | 1401 | |
2de8807b | 1402 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1403 | /* |
1404 | * If both the bitmap and buddy are in | |
1405 | * the same page we don't need to force | |
1406 | * init the buddy | |
1407 | */ | |
2de8807b AG |
1408 | ret = 0; |
1409 | goto err; | |
b6a758ec | 1410 | } |
2de8807b AG |
1411 | /* init buddy cache */ |
1412 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1413 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1414 | if (ret) |
1415 | goto err; | |
1416 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1417 | ret = -EIO; |
1418 | goto err; | |
1419 | } | |
b6a758ec | 1420 | err: |
2de8807b | 1421 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1422 | return ret; |
1423 | } | |
1424 | ||
8a57d9d6 CW |
1425 | /* |
1426 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1427 | * block group lock of all groups for this page; do not hold the BG lock when | |
1428 | * calling this routine! | |
1429 | */ | |
4ddfef7b | 1430 | static noinline_for_stack int |
adb7ef60 KK |
1431 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1432 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1433 | { |
c9de560d AT |
1434 | int blocks_per_page; |
1435 | int block; | |
1436 | int pnum; | |
1437 | int poff; | |
1438 | struct page *page; | |
fdf6c7a7 | 1439 | int ret; |
920313a7 AK |
1440 | struct ext4_group_info *grp; |
1441 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1442 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1443 | |
b10a44c3 | 1444 | might_sleep(); |
d3df1453 | 1445 | mb_debug(sb, "load group %u\n", group); |
c9de560d | 1446 | |
09cbfeaf | 1447 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1448 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1449 | |
1450 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1451 | e4b->bd_info = grp; |
c9de560d AT |
1452 | e4b->bd_sb = sb; |
1453 | e4b->bd_group = group; | |
1454 | e4b->bd_buddy_page = NULL; | |
1455 | e4b->bd_bitmap_page = NULL; | |
1456 | ||
f41c0750 | 1457 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1458 | /* |
1459 | * we need full data about the group | |
1460 | * to make a good selection | |
1461 | */ | |
adb7ef60 | 1462 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1463 | if (ret) |
1464 | return ret; | |
f41c0750 AK |
1465 | } |
1466 | ||
c9de560d AT |
1467 | /* |
1468 | * the buddy cache inode stores the block bitmap | |
1469 | * and buddy information in consecutive blocks. | |
1470 | * So for each group we need two blocks. | |
1471 | */ | |
1472 | block = group * 2; | |
1473 | pnum = block / blocks_per_page; | |
1474 | poff = block % blocks_per_page; | |
1475 | ||
1476 | /* we could use find_or_create_page(), but it locks page | |
1477 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1478 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1479 | if (page == NULL || !PageUptodate(page)) { |
1480 | if (page) | |
920313a7 AK |
1481 | /* |
1482 | * drop the page reference and try | |
1483 | * to get the page with lock. If we | |
1484 | * are not uptodate that implies | |
1485 | * somebody just created the page but | |
1486 | * is yet to initialize the same. So | |
1487 | * wait for it to initialize. | |
1488 | */ | |
09cbfeaf | 1489 | put_page(page); |
adb7ef60 | 1490 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1491 | if (page) { |
1492 | BUG_ON(page->mapping != inode->i_mapping); | |
1493 | if (!PageUptodate(page)) { | |
adb7ef60 | 1494 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1495 | if (ret) { |
1496 | unlock_page(page); | |
1497 | goto err; | |
1498 | } | |
c9de560d AT |
1499 | mb_cmp_bitmaps(e4b, page_address(page) + |
1500 | (poff * sb->s_blocksize)); | |
1501 | } | |
1502 | unlock_page(page); | |
1503 | } | |
1504 | } | |
c57ab39b YL |
1505 | if (page == NULL) { |
1506 | ret = -ENOMEM; | |
1507 | goto err; | |
1508 | } | |
1509 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1510 | ret = -EIO; |
c9de560d | 1511 | goto err; |
fdf6c7a7 | 1512 | } |
2457aec6 MG |
1513 | |
1514 | /* Pages marked accessed already */ | |
c9de560d AT |
1515 | e4b->bd_bitmap_page = page; |
1516 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1517 | |
1518 | block++; | |
1519 | pnum = block / blocks_per_page; | |
1520 | poff = block % blocks_per_page; | |
1521 | ||
2457aec6 | 1522 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1523 | if (page == NULL || !PageUptodate(page)) { |
1524 | if (page) | |
09cbfeaf | 1525 | put_page(page); |
adb7ef60 | 1526 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1527 | if (page) { |
1528 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1529 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1530 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1531 | gfp); | |
fdf6c7a7 SF |
1532 | if (ret) { |
1533 | unlock_page(page); | |
1534 | goto err; | |
1535 | } | |
1536 | } | |
c9de560d AT |
1537 | unlock_page(page); |
1538 | } | |
1539 | } | |
c57ab39b YL |
1540 | if (page == NULL) { |
1541 | ret = -ENOMEM; | |
1542 | goto err; | |
1543 | } | |
1544 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1545 | ret = -EIO; |
c9de560d | 1546 | goto err; |
fdf6c7a7 | 1547 | } |
2457aec6 MG |
1548 | |
1549 | /* Pages marked accessed already */ | |
c9de560d AT |
1550 | e4b->bd_buddy_page = page; |
1551 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d | 1552 | |
c9de560d AT |
1553 | return 0; |
1554 | ||
1555 | err: | |
26626f11 | 1556 | if (page) |
09cbfeaf | 1557 | put_page(page); |
c9de560d | 1558 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1559 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1560 | if (e4b->bd_buddy_page) |
09cbfeaf | 1561 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1562 | e4b->bd_buddy = NULL; |
1563 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1564 | return ret; |
c9de560d AT |
1565 | } |
1566 | ||
adb7ef60 KK |
1567 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1568 | struct ext4_buddy *e4b) | |
1569 | { | |
1570 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1571 | } | |
1572 | ||
e39e07fd | 1573 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1574 | { |
1575 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1576 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1577 | if (e4b->bd_buddy_page) |
09cbfeaf | 1578 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1579 | } |
1580 | ||
1581 | ||
1582 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1583 | { | |
ce3cca33 | 1584 | int order = 1, max; |
c9de560d AT |
1585 | void *bb; |
1586 | ||
c5e8f3f3 | 1587 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1588 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1589 | ||
c9de560d | 1590 | while (order <= e4b->bd_blkbits + 1) { |
ce3cca33 CX |
1591 | bb = mb_find_buddy(e4b, order, &max); |
1592 | if (!mb_test_bit(block >> order, bb)) { | |
c9de560d AT |
1593 | /* this block is part of buddy of order 'order' */ |
1594 | return order; | |
1595 | } | |
c9de560d AT |
1596 | order++; |
1597 | } | |
1598 | return 0; | |
1599 | } | |
1600 | ||
955ce5f5 | 1601 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1602 | { |
1603 | __u32 *addr; | |
1604 | ||
1605 | len = cur + len; | |
1606 | while (cur < len) { | |
1607 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1608 | /* fast path: clear whole word at once */ | |
1609 | addr = bm + (cur >> 3); | |
1610 | *addr = 0; | |
1611 | cur += 32; | |
1612 | continue; | |
1613 | } | |
955ce5f5 | 1614 | mb_clear_bit(cur, bm); |
c9de560d AT |
1615 | cur++; |
1616 | } | |
1617 | } | |
1618 | ||
eabe0444 AS |
1619 | /* clear bits in given range |
1620 | * will return first found zero bit if any, -1 otherwise | |
1621 | */ | |
1622 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1623 | { | |
1624 | __u32 *addr; | |
1625 | int zero_bit = -1; | |
1626 | ||
1627 | len = cur + len; | |
1628 | while (cur < len) { | |
1629 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1630 | /* fast path: clear whole word at once */ | |
1631 | addr = bm + (cur >> 3); | |
1632 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1633 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1634 | *addr = 0; | |
1635 | cur += 32; | |
1636 | continue; | |
1637 | } | |
1638 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1639 | zero_bit = cur; | |
1640 | cur++; | |
1641 | } | |
1642 | ||
1643 | return zero_bit; | |
1644 | } | |
1645 | ||
123e3016 | 1646 | void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1647 | { |
1648 | __u32 *addr; | |
1649 | ||
1650 | len = cur + len; | |
1651 | while (cur < len) { | |
1652 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1653 | /* fast path: set whole word at once */ | |
1654 | addr = bm + (cur >> 3); | |
1655 | *addr = 0xffffffff; | |
1656 | cur += 32; | |
1657 | continue; | |
1658 | } | |
955ce5f5 | 1659 | mb_set_bit(cur, bm); |
c9de560d AT |
1660 | cur++; |
1661 | } | |
1662 | } | |
1663 | ||
eabe0444 AS |
1664 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
1665 | { | |
1666 | if (mb_test_bit(*bit + side, bitmap)) { | |
1667 | mb_clear_bit(*bit, bitmap); | |
1668 | (*bit) -= side; | |
1669 | return 1; | |
1670 | } | |
1671 | else { | |
1672 | (*bit) += side; | |
1673 | mb_set_bit(*bit, bitmap); | |
1674 | return -1; | |
1675 | } | |
1676 | } | |
1677 | ||
1678 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1679 | { | |
1680 | int max; | |
1681 | int order = 1; | |
1682 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1683 | ||
1684 | while (buddy) { | |
1685 | void *buddy2; | |
1686 | ||
1687 | /* Bits in range [first; last] are known to be set since | |
1688 | * corresponding blocks were allocated. Bits in range | |
1689 | * (first; last) will stay set because they form buddies on | |
1690 | * upper layer. We just deal with borders if they don't | |
1691 | * align with upper layer and then go up. | |
1692 | * Releasing entire group is all about clearing | |
1693 | * single bit of highest order buddy. | |
1694 | */ | |
1695 | ||
1696 | /* Example: | |
1697 | * --------------------------------- | |
1698 | * | 1 | 1 | 1 | 1 | | |
1699 | * --------------------------------- | |
1700 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1701 | * --------------------------------- | |
1702 | * 0 1 2 3 4 5 6 7 | |
1703 | * \_____________________/ | |
1704 | * | |
1705 | * Neither [1] nor [6] is aligned to above layer. | |
1706 | * Left neighbour [0] is free, so mark it busy, | |
1707 | * decrease bb_counters and extend range to | |
1708 | * [0; 6] | |
1709 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1710 | * mark [6] free, increase bb_counters and shrink range to | |
1711 | * [0; 5]. | |
1712 | * Then shift range to [0; 2], go up and do the same. | |
1713 | */ | |
1714 | ||
1715 | ||
1716 | if (first & 1) | |
1717 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1718 | if (!(last & 1)) | |
1719 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1720 | if (first > last) | |
1721 | break; | |
1722 | order++; | |
1723 | ||
1724 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1725 | mb_clear_bits(buddy, first, last - first + 1); | |
1726 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1727 | break; | |
1728 | } | |
1729 | first >>= 1; | |
1730 | last >>= 1; | |
1731 | buddy = buddy2; | |
1732 | } | |
1733 | } | |
1734 | ||
7e5a8cdd | 1735 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1736 | int first, int count) |
c9de560d | 1737 | { |
eabe0444 AS |
1738 | int left_is_free = 0; |
1739 | int right_is_free = 0; | |
1740 | int block; | |
1741 | int last = first + count - 1; | |
c9de560d AT |
1742 | struct super_block *sb = e4b->bd_sb; |
1743 | ||
c99d1e6e TT |
1744 | if (WARN_ON(count == 0)) |
1745 | return; | |
eabe0444 | 1746 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1747 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1748 | /* Don't bother if the block group is corrupt. */ |
1749 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1750 | return; | |
1751 | ||
c9de560d AT |
1752 | mb_check_buddy(e4b); |
1753 | mb_free_blocks_double(inode, e4b, first, count); | |
1754 | ||
07b5b8e1 | 1755 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1756 | e4b->bd_info->bb_free += count; |
1757 | if (first < e4b->bd_info->bb_first_free) | |
1758 | e4b->bd_info->bb_first_free = first; | |
1759 | ||
eabe0444 AS |
1760 | /* access memory sequentially: check left neighbour, |
1761 | * clear range and then check right neighbour | |
1762 | */ | |
c9de560d | 1763 | if (first != 0) |
eabe0444 AS |
1764 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1765 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1766 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1767 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1768 | ||
1769 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1770 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1771 | ext4_fsblk_t blocknr; |
1772 | ||
1773 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1774 | blocknr += EXT4_C2B(sbi, block); |
8016e29f HS |
1775 | if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { |
1776 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1777 | inode ? inode->i_ino : 0, | |
1778 | blocknr, | |
1779 | "freeing already freed block (bit %u); block bitmap corrupt.", | |
1780 | block); | |
1781 | ext4_mark_group_bitmap_corrupted( | |
1782 | sb, e4b->bd_group, | |
db79e6d1 | 1783 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
8016e29f | 1784 | } |
eabe0444 AS |
1785 | goto done; |
1786 | } | |
1787 | ||
1788 | /* let's maintain fragments counter */ | |
1789 | if (left_is_free && right_is_free) | |
c9de560d | 1790 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1791 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1792 | e4b->bd_info->bb_fragments++; |
1793 | ||
eabe0444 AS |
1794 | /* buddy[0] == bd_bitmap is a special case, so handle |
1795 | * it right away and let mb_buddy_mark_free stay free of | |
1796 | * zero order checks. | |
1797 | * Check if neighbours are to be coaleasced, | |
1798 | * adjust bitmap bb_counters and borders appropriately. | |
1799 | */ | |
1800 | if (first & 1) { | |
1801 | first += !left_is_free; | |
1802 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1803 | } | |
1804 | if (!(last & 1)) { | |
1805 | last -= !right_is_free; | |
1806 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1807 | } | |
c9de560d | 1808 | |
eabe0444 AS |
1809 | if (first <= last) |
1810 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1811 | |
eabe0444 | 1812 | done: |
8a57d9d6 | 1813 | mb_set_largest_free_order(sb, e4b->bd_info); |
196e402a | 1814 | mb_update_avg_fragment_size(sb, e4b->bd_info); |
c9de560d | 1815 | mb_check_buddy(e4b); |
c9de560d AT |
1816 | } |
1817 | ||
15c006a2 | 1818 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1819 | int needed, struct ext4_free_extent *ex) |
1820 | { | |
1821 | int next = block; | |
15c006a2 | 1822 | int max, order; |
c9de560d AT |
1823 | void *buddy; |
1824 | ||
bc8e6740 | 1825 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1826 | BUG_ON(ex == NULL); |
1827 | ||
15c006a2 | 1828 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1829 | BUG_ON(buddy == NULL); |
1830 | BUG_ON(block >= max); | |
1831 | if (mb_test_bit(block, buddy)) { | |
1832 | ex->fe_len = 0; | |
1833 | ex->fe_start = 0; | |
1834 | ex->fe_group = 0; | |
1835 | return 0; | |
1836 | } | |
1837 | ||
15c006a2 RD |
1838 | /* find actual order */ |
1839 | order = mb_find_order_for_block(e4b, block); | |
1840 | block = block >> order; | |
c9de560d AT |
1841 | |
1842 | ex->fe_len = 1 << order; | |
1843 | ex->fe_start = block << order; | |
1844 | ex->fe_group = e4b->bd_group; | |
1845 | ||
1846 | /* calc difference from given start */ | |
1847 | next = next - ex->fe_start; | |
1848 | ex->fe_len -= next; | |
1849 | ex->fe_start += next; | |
1850 | ||
1851 | while (needed > ex->fe_len && | |
d8ec0c39 | 1852 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1853 | |
1854 | if (block + 1 >= max) | |
1855 | break; | |
1856 | ||
1857 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1858 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1859 | break; |
1860 | ||
b051d8dc | 1861 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1862 | |
c9de560d AT |
1863 | block = next >> order; |
1864 | ex->fe_len += 1 << order; | |
1865 | } | |
1866 | ||
31562b95 | 1867 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1868 | /* Should never happen! (but apparently sometimes does?!?) */ |
1869 | WARN_ON(1); | |
cd84bbba SB |
1870 | ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, |
1871 | "corruption or bug in mb_find_extent " | |
1872 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1873 | block, order, needed, ex->fe_group, ex->fe_start, | |
1874 | ex->fe_len, ex->fe_logical); | |
43c73221 TT |
1875 | ex->fe_len = 0; |
1876 | ex->fe_start = 0; | |
1877 | ex->fe_group = 0; | |
1878 | } | |
c9de560d AT |
1879 | return ex->fe_len; |
1880 | } | |
1881 | ||
1882 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1883 | { | |
1884 | int ord; | |
1885 | int mlen = 0; | |
1886 | int max = 0; | |
1887 | int cur; | |
1888 | int start = ex->fe_start; | |
1889 | int len = ex->fe_len; | |
1890 | unsigned ret = 0; | |
1891 | int len0 = len; | |
1892 | void *buddy; | |
218a6944 | 1893 | bool split = false; |
c9de560d AT |
1894 | |
1895 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1896 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1897 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1898 | mb_check_buddy(e4b); |
1899 | mb_mark_used_double(e4b, start, len); | |
1900 | ||
07b5b8e1 | 1901 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1902 | e4b->bd_info->bb_free -= len; |
1903 | if (e4b->bd_info->bb_first_free == start) | |
1904 | e4b->bd_info->bb_first_free += len; | |
1905 | ||
1906 | /* let's maintain fragments counter */ | |
1907 | if (start != 0) | |
c5e8f3f3 | 1908 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1909 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1910 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1911 | if (mlen && max) |
1912 | e4b->bd_info->bb_fragments++; | |
1913 | else if (!mlen && !max) | |
1914 | e4b->bd_info->bb_fragments--; | |
1915 | ||
1916 | /* let's maintain buddy itself */ | |
1917 | while (len) { | |
218a6944 | 1918 | if (!split) |
1919 | ord = mb_find_order_for_block(e4b, start); | |
c9de560d AT |
1920 | |
1921 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1922 | /* the whole chunk may be allocated at once! */ | |
1923 | mlen = 1 << ord; | |
218a6944 | 1924 | if (!split) |
1925 | buddy = mb_find_buddy(e4b, ord, &max); | |
1926 | else | |
1927 | split = false; | |
c9de560d AT |
1928 | BUG_ON((start >> ord) >= max); |
1929 | mb_set_bit(start >> ord, buddy); | |
1930 | e4b->bd_info->bb_counters[ord]--; | |
1931 | start += mlen; | |
1932 | len -= mlen; | |
1933 | BUG_ON(len < 0); | |
1934 | continue; | |
1935 | } | |
1936 | ||
1937 | /* store for history */ | |
1938 | if (ret == 0) | |
1939 | ret = len | (ord << 16); | |
1940 | ||
1941 | /* we have to split large buddy */ | |
1942 | BUG_ON(ord <= 0); | |
1943 | buddy = mb_find_buddy(e4b, ord, &max); | |
1944 | mb_set_bit(start >> ord, buddy); | |
1945 | e4b->bd_info->bb_counters[ord]--; | |
1946 | ||
1947 | ord--; | |
1948 | cur = (start >> ord) & ~1U; | |
1949 | buddy = mb_find_buddy(e4b, ord, &max); | |
1950 | mb_clear_bit(cur, buddy); | |
1951 | mb_clear_bit(cur + 1, buddy); | |
1952 | e4b->bd_info->bb_counters[ord]++; | |
1953 | e4b->bd_info->bb_counters[ord]++; | |
218a6944 | 1954 | split = true; |
c9de560d | 1955 | } |
8a57d9d6 | 1956 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1957 | |
196e402a | 1958 | mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); |
123e3016 | 1959 | mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1960 | mb_check_buddy(e4b); |
1961 | ||
1962 | return ret; | |
1963 | } | |
1964 | ||
1965 | /* | |
1966 | * Must be called under group lock! | |
1967 | */ | |
1968 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1969 | struct ext4_buddy *e4b) | |
1970 | { | |
1971 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1972 | int ret; | |
1973 | ||
1974 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1975 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1976 | ||
1977 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1978 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1979 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1980 | ||
1981 | /* preallocation can change ac_b_ex, thus we store actually | |
1982 | * allocated blocks for history */ | |
1983 | ac->ac_f_ex = ac->ac_b_ex; | |
1984 | ||
1985 | ac->ac_status = AC_STATUS_FOUND; | |
1986 | ac->ac_tail = ret & 0xffff; | |
1987 | ac->ac_buddy = ret >> 16; | |
1988 | ||
c3a326a6 AK |
1989 | /* |
1990 | * take the page reference. We want the page to be pinned | |
1991 | * so that we don't get a ext4_mb_init_cache_call for this | |
1992 | * group until we update the bitmap. That would mean we | |
1993 | * double allocate blocks. The reference is dropped | |
1994 | * in ext4_mb_release_context | |
1995 | */ | |
c9de560d AT |
1996 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1997 | get_page(ac->ac_bitmap_page); | |
1998 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1999 | get_page(ac->ac_buddy_page); | |
c9de560d | 2000 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 2001 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
2002 | spin_lock(&sbi->s_md_lock); |
2003 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
2004 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
2005 | spin_unlock(&sbi->s_md_lock); | |
2006 | } | |
53f86b17 RH |
2007 | /* |
2008 | * As we've just preallocated more space than | |
2009 | * user requested originally, we store allocated | |
2010 | * space in a special descriptor. | |
2011 | */ | |
2012 | if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
2013 | ext4_mb_new_preallocation(ac); | |
2014 | ||
c9de560d AT |
2015 | } |
2016 | ||
c9de560d AT |
2017 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
2018 | struct ext4_buddy *e4b, | |
2019 | int finish_group) | |
2020 | { | |
2021 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2022 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2023 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2024 | struct ext4_free_extent ex; | |
2025 | int max; | |
2026 | ||
032115fc AK |
2027 | if (ac->ac_status == AC_STATUS_FOUND) |
2028 | return; | |
c9de560d AT |
2029 | /* |
2030 | * We don't want to scan for a whole year | |
2031 | */ | |
2032 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
2033 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2034 | ac->ac_status = AC_STATUS_BREAK; | |
2035 | return; | |
2036 | } | |
2037 | ||
2038 | /* | |
2039 | * Haven't found good chunk so far, let's continue | |
2040 | */ | |
2041 | if (bex->fe_len < gex->fe_len) | |
2042 | return; | |
2043 | ||
2044 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
2045 | && bex->fe_group == e4b->bd_group) { | |
2046 | /* recheck chunk's availability - we don't know | |
2047 | * when it was found (within this lock-unlock | |
2048 | * period or not) */ | |
15c006a2 | 2049 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
2050 | if (max >= gex->fe_len) { |
2051 | ext4_mb_use_best_found(ac, e4b); | |
2052 | return; | |
2053 | } | |
2054 | } | |
2055 | } | |
2056 | ||
2057 | /* | |
2058 | * The routine checks whether found extent is good enough. If it is, | |
2059 | * then the extent gets marked used and flag is set to the context | |
2060 | * to stop scanning. Otherwise, the extent is compared with the | |
2061 | * previous found extent and if new one is better, then it's stored | |
2062 | * in the context. Later, the best found extent will be used, if | |
2063 | * mballoc can't find good enough extent. | |
2064 | * | |
2065 | * FIXME: real allocation policy is to be designed yet! | |
2066 | */ | |
2067 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
2068 | struct ext4_free_extent *ex, | |
2069 | struct ext4_buddy *e4b) | |
2070 | { | |
2071 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2072 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2073 | ||
2074 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
2075 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
2076 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
2077 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
2078 | ||
2079 | ac->ac_found++; | |
2080 | ||
2081 | /* | |
2082 | * The special case - take what you catch first | |
2083 | */ | |
2084 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2085 | *bex = *ex; | |
2086 | ext4_mb_use_best_found(ac, e4b); | |
2087 | return; | |
2088 | } | |
2089 | ||
2090 | /* | |
2091 | * Let's check whether the chuck is good enough | |
2092 | */ | |
2093 | if (ex->fe_len == gex->fe_len) { | |
2094 | *bex = *ex; | |
2095 | ext4_mb_use_best_found(ac, e4b); | |
2096 | return; | |
2097 | } | |
2098 | ||
2099 | /* | |
2100 | * If this is first found extent, just store it in the context | |
2101 | */ | |
2102 | if (bex->fe_len == 0) { | |
2103 | *bex = *ex; | |
2104 | return; | |
2105 | } | |
2106 | ||
2107 | /* | |
2108 | * If new found extent is better, store it in the context | |
2109 | */ | |
2110 | if (bex->fe_len < gex->fe_len) { | |
2111 | /* if the request isn't satisfied, any found extent | |
2112 | * larger than previous best one is better */ | |
2113 | if (ex->fe_len > bex->fe_len) | |
2114 | *bex = *ex; | |
2115 | } else if (ex->fe_len > gex->fe_len) { | |
2116 | /* if the request is satisfied, then we try to find | |
2117 | * an extent that still satisfy the request, but is | |
2118 | * smaller than previous one */ | |
2119 | if (ex->fe_len < bex->fe_len) | |
2120 | *bex = *ex; | |
2121 | } | |
2122 | ||
2123 | ext4_mb_check_limits(ac, e4b, 0); | |
2124 | } | |
2125 | ||
089ceecc ES |
2126 | static noinline_for_stack |
2127 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
2128 | struct ext4_buddy *e4b) |
2129 | { | |
2130 | struct ext4_free_extent ex = ac->ac_b_ex; | |
2131 | ext4_group_t group = ex.fe_group; | |
2132 | int max; | |
2133 | int err; | |
2134 | ||
2135 | BUG_ON(ex.fe_len <= 0); | |
2136 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2137 | if (err) | |
2138 | return err; | |
2139 | ||
2140 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 2141 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
2142 | |
2143 | if (max > 0) { | |
2144 | ac->ac_b_ex = ex; | |
2145 | ext4_mb_use_best_found(ac, e4b); | |
2146 | } | |
2147 | ||
2148 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2149 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2150 | |
2151 | return 0; | |
2152 | } | |
2153 | ||
089ceecc ES |
2154 | static noinline_for_stack |
2155 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
2156 | struct ext4_buddy *e4b) |
2157 | { | |
2158 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
2159 | int max; | |
2160 | int err; | |
2161 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 2162 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
2163 | struct ext4_free_extent ex; |
2164 | ||
2165 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
2166 | return 0; | |
838cd0cf YY |
2167 | if (grp->bb_free == 0) |
2168 | return 0; | |
c9de560d AT |
2169 | |
2170 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2171 | if (err) | |
2172 | return err; | |
2173 | ||
163a203d DW |
2174 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
2175 | ext4_mb_unload_buddy(e4b); | |
2176 | return 0; | |
2177 | } | |
2178 | ||
c9de560d | 2179 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 2180 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 2181 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 2182 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
2183 | |
2184 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
2185 | ext4_fsblk_t start; | |
2186 | ||
5661bd68 AM |
2187 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
2188 | ex.fe_start; | |
c9de560d AT |
2189 | /* use do_div to get remainder (would be 64-bit modulo) */ |
2190 | if (do_div(start, sbi->s_stripe) == 0) { | |
2191 | ac->ac_found++; | |
2192 | ac->ac_b_ex = ex; | |
2193 | ext4_mb_use_best_found(ac, e4b); | |
2194 | } | |
2195 | } else if (max >= ac->ac_g_ex.fe_len) { | |
2196 | BUG_ON(ex.fe_len <= 0); | |
2197 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2198 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2199 | ac->ac_found++; | |
2200 | ac->ac_b_ex = ex; | |
2201 | ext4_mb_use_best_found(ac, e4b); | |
2202 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
2203 | /* Sometimes, caller may want to merge even small | |
2204 | * number of blocks to an existing extent */ | |
2205 | BUG_ON(ex.fe_len <= 0); | |
2206 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2207 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2208 | ac->ac_found++; | |
2209 | ac->ac_b_ex = ex; | |
2210 | ext4_mb_use_best_found(ac, e4b); | |
2211 | } | |
2212 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2213 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2214 | |
2215 | return 0; | |
2216 | } | |
2217 | ||
2218 | /* | |
2219 | * The routine scans buddy structures (not bitmap!) from given order | |
2220 | * to max order and tries to find big enough chunk to satisfy the req | |
2221 | */ | |
089ceecc ES |
2222 | static noinline_for_stack |
2223 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2224 | struct ext4_buddy *e4b) |
2225 | { | |
2226 | struct super_block *sb = ac->ac_sb; | |
2227 | struct ext4_group_info *grp = e4b->bd_info; | |
2228 | void *buddy; | |
2229 | int i; | |
2230 | int k; | |
2231 | int max; | |
2232 | ||
2233 | BUG_ON(ac->ac_2order <= 0); | |
4b68f6df | 2234 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { |
c9de560d AT |
2235 | if (grp->bb_counters[i] == 0) |
2236 | continue; | |
2237 | ||
2238 | buddy = mb_find_buddy(e4b, i, &max); | |
2239 | BUG_ON(buddy == NULL); | |
2240 | ||
ffad0a44 | 2241 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
2242 | if (k >= max) { |
2243 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
2244 | "%d free clusters of order %d. But found 0", | |
2245 | grp->bb_counters[i], i); | |
2246 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
2247 | e4b->bd_group, | |
2248 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
2249 | break; | |
2250 | } | |
c9de560d AT |
2251 | ac->ac_found++; |
2252 | ||
2253 | ac->ac_b_ex.fe_len = 1 << i; | |
2254 | ac->ac_b_ex.fe_start = k << i; | |
2255 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
2256 | ||
2257 | ext4_mb_use_best_found(ac, e4b); | |
2258 | ||
53f86b17 | 2259 | BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
c9de560d AT |
2260 | |
2261 | if (EXT4_SB(sb)->s_mb_stats) | |
2262 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
2263 | ||
2264 | break; | |
2265 | } | |
2266 | } | |
2267 | ||
2268 | /* | |
2269 | * The routine scans the group and measures all found extents. | |
2270 | * In order to optimize scanning, caller must pass number of | |
2271 | * free blocks in the group, so the routine can know upper limit. | |
2272 | */ | |
089ceecc ES |
2273 | static noinline_for_stack |
2274 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2275 | struct ext4_buddy *e4b) |
2276 | { | |
2277 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 2278 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2279 | struct ext4_free_extent ex; |
2280 | int i; | |
2281 | int free; | |
2282 | ||
2283 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
2284 | if (WARN_ON(free <= 0)) |
2285 | return; | |
c9de560d AT |
2286 | |
2287 | i = e4b->bd_info->bb_first_free; | |
2288 | ||
2289 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 2290 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
2291 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
2292 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 2293 | /* |
e56eb659 | 2294 | * IF we have corrupt bitmap, we won't find any |
26346ff6 | 2295 | * free blocks even though group info says we |
b483bb77 | 2296 | * have free blocks |
26346ff6 | 2297 | */ |
e29136f8 | 2298 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2299 | "%d free clusters as per " |
fde4d95a | 2300 | "group info. But bitmap says 0", |
26346ff6 | 2301 | free); |
736dedbb WS |
2302 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2303 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
2304 | break; |
2305 | } | |
2306 | ||
15c006a2 | 2307 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2308 | if (WARN_ON(ex.fe_len <= 0)) |
2309 | break; | |
26346ff6 | 2310 | if (free < ex.fe_len) { |
e29136f8 | 2311 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2312 | "%d free clusters as per " |
fde4d95a | 2313 | "group info. But got %d blocks", |
26346ff6 | 2314 | free, ex.fe_len); |
736dedbb WS |
2315 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2316 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
2317 | /* |
2318 | * The number of free blocks differs. This mostly | |
2319 | * indicate that the bitmap is corrupt. So exit | |
2320 | * without claiming the space. | |
2321 | */ | |
2322 | break; | |
26346ff6 | 2323 | } |
ab0c00fc | 2324 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2325 | ext4_mb_measure_extent(ac, &ex, e4b); |
2326 | ||
2327 | i += ex.fe_len; | |
2328 | free -= ex.fe_len; | |
2329 | } | |
2330 | ||
2331 | ext4_mb_check_limits(ac, e4b, 1); | |
2332 | } | |
2333 | ||
2334 | /* | |
2335 | * This is a special case for storages like raid5 | |
506bf2d8 | 2336 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2337 | */ |
089ceecc ES |
2338 | static noinline_for_stack |
2339 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2340 | struct ext4_buddy *e4b) |
2341 | { | |
2342 | struct super_block *sb = ac->ac_sb; | |
2343 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2344 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2345 | struct ext4_free_extent ex; |
2346 | ext4_fsblk_t first_group_block; | |
2347 | ext4_fsblk_t a; | |
2348 | ext4_grpblk_t i; | |
2349 | int max; | |
2350 | ||
2351 | BUG_ON(sbi->s_stripe == 0); | |
2352 | ||
2353 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2354 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2355 | ||
c9de560d AT |
2356 | a = first_group_block + sbi->s_stripe - 1; |
2357 | do_div(a, sbi->s_stripe); | |
2358 | i = (a * sbi->s_stripe) - first_group_block; | |
2359 | ||
7137d7a4 | 2360 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2361 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2362 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2363 | if (max >= sbi->s_stripe) { |
2364 | ac->ac_found++; | |
ab0c00fc | 2365 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2366 | ac->ac_b_ex = ex; |
2367 | ext4_mb_use_best_found(ac, e4b); | |
2368 | break; | |
2369 | } | |
2370 | } | |
2371 | i += sbi->s_stripe; | |
2372 | } | |
2373 | } | |
2374 | ||
42ac1848 | 2375 | /* |
8ef123fe | 2376 | * This is also called BEFORE we load the buddy bitmap. |
42ac1848 | 2377 | * Returns either 1 or 0 indicating that the group is either suitable |
8ef123fe | 2378 | * for the allocation or not. |
42ac1848 | 2379 | */ |
8ef123fe | 2380 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
c9de560d AT |
2381 | ext4_group_t group, int cr) |
2382 | { | |
8ef123fe | 2383 | ext4_grpblk_t free, fragments; |
a4912123 | 2384 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2385 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2386 | ||
2387 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2388 | |
dddcd2f9 | 2389 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
8ef123fe | 2390 | return false; |
01fc48e8 | 2391 | |
dddcd2f9 | 2392 | free = grp->bb_free; |
2393 | if (free == 0) | |
8ef123fe | 2394 | return false; |
c9de560d | 2395 | |
c9de560d | 2396 | fragments = grp->bb_fragments; |
c9de560d | 2397 | if (fragments == 0) |
8ef123fe | 2398 | return false; |
c9de560d AT |
2399 | |
2400 | switch (cr) { | |
2401 | case 0: | |
2402 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2403 | |
a4912123 TT |
2404 | /* Avoid using the first bg of a flexgroup for data files */ |
2405 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2406 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2407 | ((group % flex_size) == 0)) | |
8ef123fe | 2408 | return false; |
a4912123 | 2409 | |
dddcd2f9 | 2410 | if (free < ac->ac_g_ex.fe_len) |
2411 | return false; | |
2412 | ||
4b68f6df | 2413 | if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) |
8ef123fe | 2414 | return true; |
40ae3487 TT |
2415 | |
2416 | if (grp->bb_largest_free_order < ac->ac_2order) | |
8ef123fe | 2417 | return false; |
40ae3487 | 2418 | |
8ef123fe | 2419 | return true; |
c9de560d AT |
2420 | case 1: |
2421 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2422 | return true; |
c9de560d AT |
2423 | break; |
2424 | case 2: | |
2425 | if (free >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2426 | return true; |
c9de560d AT |
2427 | break; |
2428 | case 3: | |
8ef123fe | 2429 | return true; |
c9de560d AT |
2430 | default: |
2431 | BUG(); | |
2432 | } | |
2433 | ||
8ef123fe RH |
2434 | return false; |
2435 | } | |
2436 | ||
2437 | /* | |
2438 | * This could return negative error code if something goes wrong | |
2439 | * during ext4_mb_init_group(). This should not be called with | |
2440 | * ext4_lock_group() held. | |
a5fda113 TT |
2441 | * |
2442 | * Note: because we are conditionally operating with the group lock in | |
2443 | * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this | |
2444 | * function using __acquire and __release. This means we need to be | |
2445 | * super careful before messing with the error path handling via "goto | |
2446 | * out"! | |
8ef123fe RH |
2447 | */ |
2448 | static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, | |
2449 | ext4_group_t group, int cr) | |
2450 | { | |
2451 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
99377830 | 2452 | struct super_block *sb = ac->ac_sb; |
c1d2c7d4 | 2453 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
99377830 | 2454 | bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
8ef123fe RH |
2455 | ext4_grpblk_t free; |
2456 | int ret = 0; | |
2457 | ||
a6c75eaf HS |
2458 | if (sbi->s_mb_stats) |
2459 | atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); | |
a5fda113 | 2460 | if (should_lock) { |
99377830 | 2461 | ext4_lock_group(sb, group); |
a5fda113 TT |
2462 | __release(ext4_group_lock_ptr(sb, group)); |
2463 | } | |
8ef123fe RH |
2464 | free = grp->bb_free; |
2465 | if (free == 0) | |
2466 | goto out; | |
2467 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2468 | goto out; | |
2469 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
2470 | goto out; | |
a5fda113 TT |
2471 | if (should_lock) { |
2472 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2473 | ext4_unlock_group(sb, group); |
a5fda113 | 2474 | } |
8ef123fe RH |
2475 | |
2476 | /* We only do this if the grp has never been initialized */ | |
2477 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
c1d2c7d4 AZ |
2478 | struct ext4_group_desc *gdp = |
2479 | ext4_get_group_desc(sb, group, NULL); | |
2480 | int ret; | |
2481 | ||
2482 | /* cr=0/1 is a very optimistic search to find large | |
2483 | * good chunks almost for free. If buddy data is not | |
2484 | * ready, then this optimization makes no sense. But | |
2485 | * we never skip the first block group in a flex_bg, | |
2486 | * since this gets used for metadata block allocation, | |
2487 | * and we want to make sure we locate metadata blocks | |
2488 | * in the first block group in the flex_bg if possible. | |
2489 | */ | |
2490 | if (cr < 2 && | |
2491 | (!sbi->s_log_groups_per_flex || | |
2492 | ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && | |
2493 | !(ext4_has_group_desc_csum(sb) && | |
2494 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) | |
2495 | return 0; | |
2496 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); | |
8ef123fe RH |
2497 | if (ret) |
2498 | return ret; | |
2499 | } | |
2500 | ||
a5fda113 | 2501 | if (should_lock) { |
99377830 | 2502 | ext4_lock_group(sb, group); |
a5fda113 TT |
2503 | __release(ext4_group_lock_ptr(sb, group)); |
2504 | } | |
8ef123fe RH |
2505 | ret = ext4_mb_good_group(ac, group, cr); |
2506 | out: | |
a5fda113 TT |
2507 | if (should_lock) { |
2508 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2509 | ext4_unlock_group(sb, group); |
a5fda113 | 2510 | } |
8ef123fe | 2511 | return ret; |
c9de560d AT |
2512 | } |
2513 | ||
cfd73237 AZ |
2514 | /* |
2515 | * Start prefetching @nr block bitmaps starting at @group. | |
2516 | * Return the next group which needs to be prefetched. | |
2517 | */ | |
3d392b26 TT |
2518 | ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
2519 | unsigned int nr, int *cnt) | |
cfd73237 AZ |
2520 | { |
2521 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2522 | struct buffer_head *bh; | |
2523 | struct blk_plug plug; | |
2524 | ||
2525 | blk_start_plug(&plug); | |
2526 | while (nr-- > 0) { | |
2527 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2528 | NULL); | |
2529 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2530 | ||
2531 | /* | |
2532 | * Prefetch block groups with free blocks; but don't | |
2533 | * bother if it is marked uninitialized on disk, since | |
2534 | * it won't require I/O to read. Also only try to | |
2535 | * prefetch once, so we avoid getblk() call, which can | |
2536 | * be expensive. | |
2537 | */ | |
2538 | if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) && | |
2539 | EXT4_MB_GRP_NEED_INIT(grp) && | |
2540 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2541 | !(ext4_has_group_desc_csum(sb) && | |
2542 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2543 | bh = ext4_read_block_bitmap_nowait(sb, group, true); | |
2544 | if (bh && !IS_ERR(bh)) { | |
2545 | if (!buffer_uptodate(bh) && cnt) | |
2546 | (*cnt)++; | |
2547 | brelse(bh); | |
2548 | } | |
2549 | } | |
2550 | if (++group >= ngroups) | |
2551 | group = 0; | |
2552 | } | |
2553 | blk_finish_plug(&plug); | |
2554 | return group; | |
2555 | } | |
2556 | ||
2557 | /* | |
2558 | * Prefetching reads the block bitmap into the buffer cache; but we | |
2559 | * need to make sure that the buddy bitmap in the page cache has been | |
2560 | * initialized. Note that ext4_mb_init_group() will block if the I/O | |
2561 | * is not yet completed, or indeed if it was not initiated by | |
2562 | * ext4_mb_prefetch did not start the I/O. | |
2563 | * | |
2564 | * TODO: We should actually kick off the buddy bitmap setup in a work | |
2565 | * queue when the buffer I/O is completed, so that we don't block | |
2566 | * waiting for the block allocation bitmap read to finish when | |
2567 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). | |
2568 | */ | |
3d392b26 TT |
2569 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
2570 | unsigned int nr) | |
cfd73237 AZ |
2571 | { |
2572 | while (nr-- > 0) { | |
2573 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2574 | NULL); | |
2575 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2576 | ||
2577 | if (!group) | |
2578 | group = ext4_get_groups_count(sb); | |
2579 | group--; | |
2580 | grp = ext4_get_group_info(sb, group); | |
2581 | ||
2582 | if (EXT4_MB_GRP_NEED_INIT(grp) && | |
2583 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2584 | !(ext4_has_group_desc_csum(sb) && | |
2585 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2586 | if (ext4_mb_init_group(sb, group, GFP_NOFS)) | |
2587 | break; | |
2588 | } | |
2589 | } | |
2590 | } | |
2591 | ||
4ddfef7b ES |
2592 | static noinline_for_stack int |
2593 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2594 | { |
cfd73237 | 2595 | ext4_group_t prefetch_grp = 0, ngroups, group, i; |
4fca50d4 | 2596 | int cr = -1, new_cr; |
42ac1848 | 2597 | int err = 0, first_err = 0; |
cfd73237 | 2598 | unsigned int nr = 0, prefetch_ios = 0; |
c9de560d AT |
2599 | struct ext4_sb_info *sbi; |
2600 | struct super_block *sb; | |
2601 | struct ext4_buddy e4b; | |
66d5e027 | 2602 | int lost; |
c9de560d AT |
2603 | |
2604 | sb = ac->ac_sb; | |
2605 | sbi = EXT4_SB(sb); | |
8df9675f | 2606 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2607 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2608 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2609 | ngroups = sbi->s_blockfile_groups; |
2610 | ||
c9de560d AT |
2611 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2612 | ||
2613 | /* first, try the goal */ | |
2614 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2615 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2616 | goto out; | |
2617 | ||
2618 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2619 | goto out; | |
2620 | ||
2621 | /* | |
e9a3cd48 | 2622 | * ac->ac_2order is set only if the fe_len is a power of 2 |
2623 | * if ac->ac_2order is set we also set criteria to 0 so that we | |
c9de560d AT |
2624 | * try exact allocation using buddy. |
2625 | */ | |
2626 | i = fls(ac->ac_g_ex.fe_len); | |
2627 | ac->ac_2order = 0; | |
2628 | /* | |
2629 | * We search using buddy data only if the order of the request | |
2630 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2631 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2632 | * We also support searching for power-of-two requests only for |
2633 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2634 | */ |
4b68f6df | 2635 | if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { |
c9de560d AT |
2636 | /* |
2637 | * This should tell if fe_len is exactly power of 2 | |
2638 | */ | |
2639 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d | 2640 | ac->ac_2order = array_index_nospec(i - 1, |
4b68f6df | 2641 | MB_NUM_ORDERS(sb)); |
c9de560d AT |
2642 | } |
2643 | ||
4ba74d00 TT |
2644 | /* if stream allocation is enabled, use global goal */ |
2645 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2646 | /* TBD: may be hot point */ |
2647 | spin_lock(&sbi->s_md_lock); | |
2648 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2649 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2650 | spin_unlock(&sbi->s_md_lock); | |
2651 | } | |
4ba74d00 | 2652 | |
c9de560d AT |
2653 | /* Let's just scan groups to find more-less suitable blocks */ |
2654 | cr = ac->ac_2order ? 0 : 1; | |
2655 | /* | |
2656 | * cr == 0 try to get exact allocation, | |
2657 | * cr == 3 try to get anything | |
2658 | */ | |
2659 | repeat: | |
2660 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2661 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2662 | /* |
2663 | * searching for the right group start | |
2664 | * from the goal value specified | |
2665 | */ | |
2666 | group = ac->ac_g_ex.fe_group; | |
196e402a | 2667 | ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; |
cfd73237 | 2668 | prefetch_grp = group; |
ed8f9c75 | 2669 | |
4fca50d4 JK |
2670 | for (i = 0, new_cr = cr; i < ngroups; i++, |
2671 | ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { | |
2672 | int ret = 0; | |
196e402a | 2673 | |
2ed5724d | 2674 | cond_resched(); |
196e402a HS |
2675 | if (new_cr != cr) { |
2676 | cr = new_cr; | |
2677 | goto repeat; | |
2678 | } | |
c9de560d | 2679 | |
cfd73237 AZ |
2680 | /* |
2681 | * Batch reads of the block allocation bitmaps | |
2682 | * to get multiple READs in flight; limit | |
2683 | * prefetching at cr=0/1, otherwise mballoc can | |
2684 | * spend a lot of time loading imperfect groups | |
2685 | */ | |
2686 | if ((prefetch_grp == group) && | |
2687 | (cr > 1 || | |
2688 | prefetch_ios < sbi->s_mb_prefetch_limit)) { | |
2689 | unsigned int curr_ios = prefetch_ios; | |
2690 | ||
2691 | nr = sbi->s_mb_prefetch; | |
2692 | if (ext4_has_feature_flex_bg(sb)) { | |
82ef1370 CX |
2693 | nr = 1 << sbi->s_log_groups_per_flex; |
2694 | nr -= group & (nr - 1); | |
2695 | nr = min(nr, sbi->s_mb_prefetch); | |
cfd73237 AZ |
2696 | } |
2697 | prefetch_grp = ext4_mb_prefetch(sb, group, | |
2698 | nr, &prefetch_ios); | |
2699 | if (prefetch_ios == curr_ios) | |
2700 | nr = 0; | |
2701 | } | |
2702 | ||
8a57d9d6 | 2703 | /* This now checks without needing the buddy page */ |
8ef123fe | 2704 | ret = ext4_mb_good_group_nolock(ac, group, cr); |
42ac1848 LC |
2705 | if (ret <= 0) { |
2706 | if (!first_err) | |
2707 | first_err = ret; | |
c9de560d | 2708 | continue; |
42ac1848 | 2709 | } |
c9de560d | 2710 | |
c9de560d AT |
2711 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2712 | if (err) | |
2713 | goto out; | |
2714 | ||
2715 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2716 | |
2717 | /* | |
2718 | * We need to check again after locking the | |
2719 | * block group | |
2720 | */ | |
42ac1848 | 2721 | ret = ext4_mb_good_group(ac, group, cr); |
8ef123fe | 2722 | if (ret == 0) { |
c9de560d | 2723 | ext4_unlock_group(sb, group); |
e39e07fd | 2724 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2725 | continue; |
2726 | } | |
2727 | ||
2728 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2729 | if (cr == 0) |
c9de560d | 2730 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2731 | else if (cr == 1 && sbi->s_stripe && |
2732 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2733 | ext4_mb_scan_aligned(ac, &e4b); |
2734 | else | |
2735 | ext4_mb_complex_scan_group(ac, &e4b); | |
2736 | ||
2737 | ext4_unlock_group(sb, group); | |
e39e07fd | 2738 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2739 | |
2740 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2741 | break; | |
2742 | } | |
a6c75eaf HS |
2743 | /* Processed all groups and haven't found blocks */ |
2744 | if (sbi->s_mb_stats && i == ngroups) | |
2745 | atomic64_inc(&sbi->s_bal_cX_failed[cr]); | |
c9de560d AT |
2746 | } |
2747 | ||
2748 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2749 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2750 | /* | |
2751 | * We've been searching too long. Let's try to allocate | |
2752 | * the best chunk we've found so far | |
2753 | */ | |
c9de560d AT |
2754 | ext4_mb_try_best_found(ac, &e4b); |
2755 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2756 | /* | |
2757 | * Someone more lucky has already allocated it. | |
2758 | * The only thing we can do is just take first | |
2759 | * found block(s) | |
c9de560d | 2760 | */ |
66d5e027 | 2761 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
2762 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", | |
c55ee7d2 | 2763 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
2764 | ac->ac_b_ex.fe_len, lost); | |
2765 | ||
c9de560d AT |
2766 | ac->ac_b_ex.fe_group = 0; |
2767 | ac->ac_b_ex.fe_start = 0; | |
2768 | ac->ac_b_ex.fe_len = 0; | |
2769 | ac->ac_status = AC_STATUS_CONTINUE; | |
2770 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2771 | cr = 3; | |
c9de560d AT |
2772 | goto repeat; |
2773 | } | |
2774 | } | |
a6c75eaf HS |
2775 | |
2776 | if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) | |
2777 | atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); | |
c9de560d | 2778 | out: |
42ac1848 LC |
2779 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2780 | err = first_err; | |
bbc4ec77 | 2781 | |
d3df1453 | 2782 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
bbc4ec77 RH |
2783 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
2784 | ac->ac_flags, cr, err); | |
cfd73237 AZ |
2785 | |
2786 | if (nr) | |
2787 | ext4_mb_prefetch_fini(sb, prefetch_grp, nr); | |
2788 | ||
c9de560d AT |
2789 | return err; |
2790 | } | |
2791 | ||
c9de560d AT |
2792 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2793 | { | |
359745d7 | 2794 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
2795 | ext4_group_t group; |
2796 | ||
8df9675f | 2797 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2798 | return NULL; |
c9de560d | 2799 | group = *pos + 1; |
a9df9a49 | 2800 | return (void *) ((unsigned long) group); |
c9de560d AT |
2801 | } |
2802 | ||
2803 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2804 | { | |
359745d7 | 2805 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
2806 | ext4_group_t group; |
2807 | ||
2808 | ++*pos; | |
8df9675f | 2809 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2810 | return NULL; |
2811 | group = *pos + 1; | |
a9df9a49 | 2812 | return (void *) ((unsigned long) group); |
c9de560d AT |
2813 | } |
2814 | ||
2815 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2816 | { | |
359745d7 | 2817 | struct super_block *sb = pde_data(file_inode(seq->file)); |
a9df9a49 | 2818 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2819 | int i; |
1c8457ca | 2820 | int err, buddy_loaded = 0; |
c9de560d | 2821 | struct ext4_buddy e4b; |
1c8457ca | 2822 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2823 | unsigned char blocksize_bits = min_t(unsigned char, |
2824 | sb->s_blocksize_bits, | |
2825 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2826 | struct sg { |
2827 | struct ext4_group_info info; | |
b80b32b6 | 2828 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2829 | } sg; |
2830 | ||
2831 | group--; | |
2832 | if (group == 0) | |
97b4af2f RV |
2833 | seq_puts(seq, "#group: free frags first [" |
2834 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2835 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2836 | |
b80b32b6 TT |
2837 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2838 | sizeof(struct ext4_group_info); | |
2839 | ||
1c8457ca AK |
2840 | grinfo = ext4_get_group_info(sb, group); |
2841 | /* Load the group info in memory only if not already loaded. */ | |
2842 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2843 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2844 | if (err) { | |
2845 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2846 | return 0; | |
2847 | } | |
2848 | buddy_loaded = 1; | |
c9de560d | 2849 | } |
1c8457ca | 2850 | |
b80b32b6 | 2851 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2852 | |
2853 | if (buddy_loaded) | |
2854 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2855 | |
a9df9a49 | 2856 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2857 | sg.info.bb_fragments, sg.info.bb_first_free); |
2858 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2859 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d | 2860 | sg.info.bb_counters[i] : 0); |
e0d438c7 | 2861 | seq_puts(seq, " ]\n"); |
c9de560d AT |
2862 | |
2863 | return 0; | |
2864 | } | |
2865 | ||
2866 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2867 | { | |
2868 | } | |
2869 | ||
247dbed8 | 2870 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2871 | .start = ext4_mb_seq_groups_start, |
2872 | .next = ext4_mb_seq_groups_next, | |
2873 | .stop = ext4_mb_seq_groups_stop, | |
2874 | .show = ext4_mb_seq_groups_show, | |
2875 | }; | |
2876 | ||
a6c75eaf HS |
2877 | int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) |
2878 | { | |
c30365b9 | 2879 | struct super_block *sb = seq->private; |
a6c75eaf HS |
2880 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2881 | ||
2882 | seq_puts(seq, "mballoc:\n"); | |
2883 | if (!sbi->s_mb_stats) { | |
2884 | seq_puts(seq, "\tmb stats collection turned off.\n"); | |
2885 | seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); | |
2886 | return 0; | |
2887 | } | |
2888 | seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); | |
2889 | seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); | |
2890 | ||
2891 | seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned)); | |
2892 | ||
2893 | seq_puts(seq, "\tcr0_stats:\n"); | |
2894 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0])); | |
2895 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2896 | atomic64_read(&sbi->s_bal_cX_groups_considered[0])); | |
2897 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2898 | atomic64_read(&sbi->s_bal_cX_failed[0])); | |
196e402a HS |
2899 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2900 | atomic_read(&sbi->s_bal_cr0_bad_suggestions)); | |
a6c75eaf HS |
2901 | |
2902 | seq_puts(seq, "\tcr1_stats:\n"); | |
2903 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1])); | |
2904 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2905 | atomic64_read(&sbi->s_bal_cX_groups_considered[1])); | |
2906 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2907 | atomic64_read(&sbi->s_bal_cX_failed[1])); | |
196e402a HS |
2908 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2909 | atomic_read(&sbi->s_bal_cr1_bad_suggestions)); | |
a6c75eaf HS |
2910 | |
2911 | seq_puts(seq, "\tcr2_stats:\n"); | |
2912 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2])); | |
2913 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2914 | atomic64_read(&sbi->s_bal_cX_groups_considered[2])); | |
2915 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2916 | atomic64_read(&sbi->s_bal_cX_failed[2])); | |
2917 | ||
2918 | seq_puts(seq, "\tcr3_stats:\n"); | |
2919 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3])); | |
2920 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2921 | atomic64_read(&sbi->s_bal_cX_groups_considered[3])); | |
2922 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2923 | atomic64_read(&sbi->s_bal_cX_failed[3])); | |
2924 | seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned)); | |
2925 | seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); | |
2926 | seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); | |
2927 | seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); | |
2928 | seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); | |
2929 | ||
2930 | seq_printf(seq, "\tbuddies_generated: %u/%u\n", | |
2931 | atomic_read(&sbi->s_mb_buddies_generated), | |
2932 | ext4_get_groups_count(sb)); | |
2933 | seq_printf(seq, "\tbuddies_time_used: %llu\n", | |
2934 | atomic64_read(&sbi->s_mb_generation_time)); | |
2935 | seq_printf(seq, "\tpreallocated: %u\n", | |
2936 | atomic_read(&sbi->s_mb_preallocated)); | |
2937 | seq_printf(seq, "\tdiscarded: %u\n", | |
2938 | atomic_read(&sbi->s_mb_discarded)); | |
2939 | return 0; | |
2940 | } | |
2941 | ||
f68f4063 | 2942 | static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) |
a5fda113 | 2943 | __acquires(&EXT4_SB(sb)->s_mb_rb_lock) |
f68f4063 | 2944 | { |
359745d7 | 2945 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2946 | unsigned long position; |
2947 | ||
83e80a6e | 2948 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
f68f4063 HS |
2949 | return NULL; |
2950 | position = *pos + 1; | |
2951 | return (void *) ((unsigned long) position); | |
2952 | } | |
2953 | ||
2954 | static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) | |
2955 | { | |
359745d7 | 2956 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2957 | unsigned long position; |
2958 | ||
2959 | ++*pos; | |
83e80a6e | 2960 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
f68f4063 HS |
2961 | return NULL; |
2962 | position = *pos + 1; | |
2963 | return (void *) ((unsigned long) position); | |
2964 | } | |
2965 | ||
2966 | static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) | |
2967 | { | |
359745d7 | 2968 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2969 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2970 | unsigned long position = ((unsigned long) v); | |
2971 | struct ext4_group_info *grp; | |
83e80a6e | 2972 | unsigned int count; |
f68f4063 HS |
2973 | |
2974 | position--; | |
2975 | if (position >= MB_NUM_ORDERS(sb)) { | |
83e80a6e JK |
2976 | position -= MB_NUM_ORDERS(sb); |
2977 | if (position == 0) | |
2978 | seq_puts(seq, "avg_fragment_size_lists:\n"); | |
f68f4063 | 2979 | |
83e80a6e JK |
2980 | count = 0; |
2981 | read_lock(&sbi->s_mb_avg_fragment_size_locks[position]); | |
2982 | list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position], | |
2983 | bb_avg_fragment_size_node) | |
2984 | count++; | |
2985 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]); | |
2986 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", | |
2987 | (unsigned int)position, count); | |
f68f4063 HS |
2988 | return 0; |
2989 | } | |
2990 | ||
2991 | if (position == 0) { | |
2992 | seq_printf(seq, "optimize_scan: %d\n", | |
2993 | test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); | |
2994 | seq_puts(seq, "max_free_order_lists:\n"); | |
2995 | } | |
2996 | count = 0; | |
83e80a6e | 2997 | read_lock(&sbi->s_mb_largest_free_orders_locks[position]); |
f68f4063 HS |
2998 | list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], |
2999 | bb_largest_free_order_node) | |
3000 | count++; | |
83e80a6e | 3001 | read_unlock(&sbi->s_mb_largest_free_orders_locks[position]); |
f68f4063 HS |
3002 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", |
3003 | (unsigned int)position, count); | |
3004 | ||
3005 | return 0; | |
3006 | } | |
3007 | ||
3008 | static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) | |
3009 | { | |
f68f4063 HS |
3010 | } |
3011 | ||
3012 | const struct seq_operations ext4_mb_seq_structs_summary_ops = { | |
3013 | .start = ext4_mb_seq_structs_summary_start, | |
3014 | .next = ext4_mb_seq_structs_summary_next, | |
3015 | .stop = ext4_mb_seq_structs_summary_stop, | |
3016 | .show = ext4_mb_seq_structs_summary_show, | |
3017 | }; | |
3018 | ||
fb1813f4 CW |
3019 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
3020 | { | |
3021 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3022 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
3023 | ||
3024 | BUG_ON(!cachep); | |
3025 | return cachep; | |
3026 | } | |
5f21b0e6 | 3027 | |
28623c2f TT |
3028 | /* |
3029 | * Allocate the top-level s_group_info array for the specified number | |
3030 | * of groups | |
3031 | */ | |
3032 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
3033 | { | |
3034 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3035 | unsigned size; | |
df3da4ea | 3036 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
3037 | |
3038 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
3039 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
3040 | if (size <= sbi->s_group_info_size) | |
3041 | return 0; | |
3042 | ||
3043 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 3044 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
3045 | if (!new_groupinfo) { |
3046 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
3047 | return -ENOMEM; | |
3048 | } | |
df3da4ea SJS |
3049 | rcu_read_lock(); |
3050 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
3051 | if (old_groupinfo) | |
3052 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 3053 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
3054 | rcu_read_unlock(); |
3055 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 3056 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
3057 | if (old_groupinfo) |
3058 | ext4_kvfree_array_rcu(old_groupinfo); | |
666245d9 | 3059 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
28623c2f TT |
3060 | sbi->s_group_info_size); |
3061 | return 0; | |
3062 | } | |
3063 | ||
5f21b0e6 | 3064 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 3065 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
3066 | struct ext4_group_desc *desc) |
3067 | { | |
fb1813f4 | 3068 | int i; |
5f21b0e6 | 3069 | int metalen = 0; |
df3da4ea | 3070 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
3071 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3072 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 3073 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
3074 | |
3075 | /* | |
3076 | * First check if this group is the first of a reserved block. | |
3077 | * If it's true, we have to allocate a new table of pointers | |
3078 | * to ext4_group_info structures | |
3079 | */ | |
3080 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
3081 | metalen = sizeof(*meta_group_info) << | |
3082 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 3083 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 3084 | if (meta_group_info == NULL) { |
7f6a11e7 | 3085 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 3086 | "for a buddy group"); |
5f21b0e6 FB |
3087 | goto exit_meta_group_info; |
3088 | } | |
df3da4ea SJS |
3089 | rcu_read_lock(); |
3090 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
3091 | rcu_read_unlock(); | |
5f21b0e6 FB |
3092 | } |
3093 | ||
df3da4ea | 3094 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
3095 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
3096 | ||
4fdb5543 | 3097 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 3098 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 3099 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
3100 | goto exit_group_info; |
3101 | } | |
3102 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
3103 | &(meta_group_info[i]->bb_state)); | |
3104 | ||
3105 | /* | |
3106 | * initialize bb_free to be able to skip | |
3107 | * empty groups without initialization | |
3108 | */ | |
8844618d TT |
3109 | if (ext4_has_group_desc_csum(sb) && |
3110 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 3111 | meta_group_info[i]->bb_free = |
cff1dfd7 | 3112 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
3113 | } else { |
3114 | meta_group_info[i]->bb_free = | |
021b65bb | 3115 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
3116 | } |
3117 | ||
3118 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 3119 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 3120 | meta_group_info[i]->bb_free_root = RB_ROOT; |
196e402a | 3121 | INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); |
83e80a6e | 3122 | INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); |
8a57d9d6 | 3123 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
83e80a6e | 3124 | meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ |
196e402a | 3125 | meta_group_info[i]->bb_group = group; |
5f21b0e6 | 3126 | |
a3450215 | 3127 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
3128 | return 0; |
3129 | ||
3130 | exit_group_info: | |
3131 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 3132 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
3133 | struct ext4_group_info ***group_info; |
3134 | ||
3135 | rcu_read_lock(); | |
3136 | group_info = rcu_dereference(sbi->s_group_info); | |
3137 | kfree(group_info[idx]); | |
3138 | group_info[idx] = NULL; | |
3139 | rcu_read_unlock(); | |
caaf7a29 | 3140 | } |
5f21b0e6 FB |
3141 | exit_meta_group_info: |
3142 | return -ENOMEM; | |
3143 | } /* ext4_mb_add_groupinfo */ | |
3144 | ||
c9de560d AT |
3145 | static int ext4_mb_init_backend(struct super_block *sb) |
3146 | { | |
8df9675f | 3147 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 3148 | ext4_group_t i; |
c9de560d | 3149 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 3150 | int err; |
5f21b0e6 | 3151 | struct ext4_group_desc *desc; |
df3da4ea | 3152 | struct ext4_group_info ***group_info; |
fb1813f4 | 3153 | struct kmem_cache *cachep; |
5f21b0e6 | 3154 | |
28623c2f TT |
3155 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
3156 | if (err) | |
3157 | return err; | |
c9de560d | 3158 | |
c9de560d AT |
3159 | sbi->s_buddy_cache = new_inode(sb); |
3160 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 3161 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
3162 | goto err_freesgi; |
3163 | } | |
48e6061b YJ |
3164 | /* To avoid potentially colliding with an valid on-disk inode number, |
3165 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
3166 | * not in the inode hash, so it should never be found by iget(), but | |
3167 | * this will avoid confusion if it ever shows up during debugging. */ | |
3168 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 3169 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 3170 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3171 | cond_resched(); |
c9de560d AT |
3172 | desc = ext4_get_group_desc(sb, i, NULL); |
3173 | if (desc == NULL) { | |
9d8b9ec4 | 3174 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
3175 | goto err_freebuddy; |
3176 | } | |
5f21b0e6 FB |
3177 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
3178 | goto err_freebuddy; | |
c9de560d AT |
3179 | } |
3180 | ||
cfd73237 | 3181 | if (ext4_has_feature_flex_bg(sb)) { |
f91436d5 ST |
3182 | /* a single flex group is supposed to be read by a single IO. |
3183 | * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is | |
3184 | * unsigned integer, so the maximum shift is 32. | |
3185 | */ | |
3186 | if (sbi->s_es->s_log_groups_per_flex >= 32) { | |
3187 | ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); | |
a8867f4e | 3188 | goto err_freebuddy; |
f91436d5 ST |
3189 | } |
3190 | sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, | |
82ef1370 | 3191 | BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); |
cfd73237 AZ |
3192 | sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ |
3193 | } else { | |
3194 | sbi->s_mb_prefetch = 32; | |
3195 | } | |
3196 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) | |
3197 | sbi->s_mb_prefetch = ext4_get_groups_count(sb); | |
3198 | /* now many real IOs to prefetch within a single allocation at cr=0 | |
3199 | * given cr=0 is an CPU-related optimization we shouldn't try to | |
3200 | * load too many groups, at some point we should start to use what | |
3201 | * we've got in memory. | |
3202 | * with an average random access time 5ms, it'd take a second to get | |
3203 | * 200 groups (* N with flex_bg), so let's make this limit 4 | |
3204 | */ | |
3205 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; | |
3206 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) | |
3207 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); | |
3208 | ||
c9de560d AT |
3209 | return 0; |
3210 | ||
3211 | err_freebuddy: | |
fb1813f4 | 3212 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 3213 | while (i-- > 0) |
fb1813f4 | 3214 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 3215 | i = sbi->s_group_info_size; |
df3da4ea SJS |
3216 | rcu_read_lock(); |
3217 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 3218 | while (i-- > 0) |
df3da4ea SJS |
3219 | kfree(group_info[i]); |
3220 | rcu_read_unlock(); | |
c9de560d AT |
3221 | iput(sbi->s_buddy_cache); |
3222 | err_freesgi: | |
df3da4ea SJS |
3223 | rcu_read_lock(); |
3224 | kvfree(rcu_dereference(sbi->s_group_info)); | |
3225 | rcu_read_unlock(); | |
c9de560d AT |
3226 | return -ENOMEM; |
3227 | } | |
3228 | ||
2892c15d ES |
3229 | static void ext4_groupinfo_destroy_slabs(void) |
3230 | { | |
3231 | int i; | |
3232 | ||
3233 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 3234 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
3235 | ext4_groupinfo_caches[i] = NULL; |
3236 | } | |
3237 | } | |
3238 | ||
3239 | static int ext4_groupinfo_create_slab(size_t size) | |
3240 | { | |
3241 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
3242 | int slab_size; | |
3243 | int blocksize_bits = order_base_2(size); | |
3244 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3245 | struct kmem_cache *cachep; | |
3246 | ||
3247 | if (cache_index >= NR_GRPINFO_CACHES) | |
3248 | return -EINVAL; | |
3249 | ||
3250 | if (unlikely(cache_index < 0)) | |
3251 | cache_index = 0; | |
3252 | ||
3253 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
3254 | if (ext4_groupinfo_caches[cache_index]) { | |
3255 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
3256 | return 0; /* Already created */ | |
3257 | } | |
3258 | ||
3259 | slab_size = offsetof(struct ext4_group_info, | |
3260 | bb_counters[blocksize_bits + 2]); | |
3261 | ||
3262 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
3263 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
3264 | NULL); | |
3265 | ||
823ba01f TM |
3266 | ext4_groupinfo_caches[cache_index] = cachep; |
3267 | ||
2892c15d ES |
3268 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
3269 | if (!cachep) { | |
9d8b9ec4 TT |
3270 | printk(KERN_EMERG |
3271 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
3272 | return -ENOMEM; |
3273 | } | |
3274 | ||
2892c15d ES |
3275 | return 0; |
3276 | } | |
3277 | ||
55cdd0af WJ |
3278 | static void ext4_discard_work(struct work_struct *work) |
3279 | { | |
3280 | struct ext4_sb_info *sbi = container_of(work, | |
3281 | struct ext4_sb_info, s_discard_work); | |
3282 | struct super_block *sb = sbi->s_sb; | |
3283 | struct ext4_free_data *fd, *nfd; | |
3284 | struct ext4_buddy e4b; | |
3285 | struct list_head discard_list; | |
3286 | ext4_group_t grp, load_grp; | |
3287 | int err = 0; | |
3288 | ||
3289 | INIT_LIST_HEAD(&discard_list); | |
3290 | spin_lock(&sbi->s_md_lock); | |
3291 | list_splice_init(&sbi->s_discard_list, &discard_list); | |
3292 | spin_unlock(&sbi->s_md_lock); | |
3293 | ||
3294 | load_grp = UINT_MAX; | |
3295 | list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { | |
3296 | /* | |
5036ab8d WJ |
3297 | * If filesystem is umounting or no memory or suffering |
3298 | * from no space, give up the discard | |
55cdd0af | 3299 | */ |
5036ab8d WJ |
3300 | if ((sb->s_flags & SB_ACTIVE) && !err && |
3301 | !atomic_read(&sbi->s_retry_alloc_pending)) { | |
55cdd0af WJ |
3302 | grp = fd->efd_group; |
3303 | if (grp != load_grp) { | |
3304 | if (load_grp != UINT_MAX) | |
3305 | ext4_mb_unload_buddy(&e4b); | |
3306 | ||
3307 | err = ext4_mb_load_buddy(sb, grp, &e4b); | |
3308 | if (err) { | |
3309 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3310 | load_grp = UINT_MAX; | |
3311 | continue; | |
3312 | } else { | |
3313 | load_grp = grp; | |
3314 | } | |
3315 | } | |
3316 | ||
3317 | ext4_lock_group(sb, grp); | |
3318 | ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, | |
3319 | fd->efd_start_cluster + fd->efd_count - 1, 1); | |
3320 | ext4_unlock_group(sb, grp); | |
3321 | } | |
3322 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3323 | } | |
3324 | ||
3325 | if (load_grp != UINT_MAX) | |
3326 | ext4_mb_unload_buddy(&e4b); | |
3327 | } | |
3328 | ||
9d99012f | 3329 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
3330 | { |
3331 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 3332 | unsigned i, j; |
935244cd | 3333 | unsigned offset, offset_incr; |
c9de560d | 3334 | unsigned max; |
74767c5a | 3335 | int ret; |
c9de560d | 3336 | |
4b68f6df | 3337 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
3338 | |
3339 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
3340 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
3341 | ret = -ENOMEM; |
3342 | goto out; | |
c9de560d | 3343 | } |
ff7ef329 | 3344 | |
4b68f6df | 3345 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
3346 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
3347 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
3348 | ret = -ENOMEM; |
3349 | goto out; | |
3350 | } | |
3351 | ||
2892c15d ES |
3352 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
3353 | if (ret < 0) | |
3354 | goto out; | |
c9de560d AT |
3355 | |
3356 | /* order 0 is regular bitmap */ | |
3357 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
3358 | sbi->s_mb_offsets[0] = 0; | |
3359 | ||
3360 | i = 1; | |
3361 | offset = 0; | |
935244cd | 3362 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
3363 | max = sb->s_blocksize << 2; |
3364 | do { | |
3365 | sbi->s_mb_offsets[i] = offset; | |
3366 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
3367 | offset += offset_incr; |
3368 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
3369 | max = max >> 1; |
3370 | i++; | |
4b68f6df HS |
3371 | } while (i < MB_NUM_ORDERS(sb)); |
3372 | ||
83e80a6e JK |
3373 | sbi->s_mb_avg_fragment_size = |
3374 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3375 | GFP_KERNEL); | |
3376 | if (!sbi->s_mb_avg_fragment_size) { | |
3377 | ret = -ENOMEM; | |
3378 | goto out; | |
3379 | } | |
3380 | sbi->s_mb_avg_fragment_size_locks = | |
3381 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3382 | GFP_KERNEL); | |
3383 | if (!sbi->s_mb_avg_fragment_size_locks) { | |
3384 | ret = -ENOMEM; | |
3385 | goto out; | |
3386 | } | |
3387 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3388 | INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); | |
3389 | rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); | |
3390 | } | |
196e402a HS |
3391 | sbi->s_mb_largest_free_orders = |
3392 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3393 | GFP_KERNEL); | |
3394 | if (!sbi->s_mb_largest_free_orders) { | |
3395 | ret = -ENOMEM; | |
3396 | goto out; | |
3397 | } | |
3398 | sbi->s_mb_largest_free_orders_locks = | |
3399 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3400 | GFP_KERNEL); | |
3401 | if (!sbi->s_mb_largest_free_orders_locks) { | |
3402 | ret = -ENOMEM; | |
3403 | goto out; | |
3404 | } | |
3405 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3406 | INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); | |
3407 | rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); | |
3408 | } | |
c9de560d | 3409 | |
c9de560d | 3410 | spin_lock_init(&sbi->s_md_lock); |
d08854f5 | 3411 | sbi->s_mb_free_pending = 0; |
a0154344 | 3412 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
55cdd0af WJ |
3413 | INIT_LIST_HEAD(&sbi->s_discard_list); |
3414 | INIT_WORK(&sbi->s_discard_work, ext4_discard_work); | |
5036ab8d | 3415 | atomic_set(&sbi->s_retry_alloc_pending, 0); |
c9de560d AT |
3416 | |
3417 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
3418 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
3419 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
3420 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
3421 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27bc446e | 3422 | sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC; |
27baebb8 TT |
3423 | /* |
3424 | * The default group preallocation is 512, which for 4k block | |
3425 | * sizes translates to 2 megabytes. However for bigalloc file | |
3426 | * systems, this is probably too big (i.e, if the cluster size | |
3427 | * is 1 megabyte, then group preallocation size becomes half a | |
3428 | * gigabyte!). As a default, we will keep a two megabyte | |
3429 | * group pralloc size for cluster sizes up to 64k, and after | |
3430 | * that, we will force a minimum group preallocation size of | |
3431 | * 32 clusters. This translates to 8 megs when the cluster | |
3432 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
3433 | * which seems reasonable as a default. | |
3434 | */ | |
3435 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
3436 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
3437 | /* |
3438 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
3439 | * to the lowest multiple of s_stripe which is bigger than | |
3440 | * the s_mb_group_prealloc as determined above. We want | |
3441 | * the preallocation size to be an exact multiple of the | |
3442 | * RAID stripe size so that preallocations don't fragment | |
3443 | * the stripes. | |
3444 | */ | |
3445 | if (sbi->s_stripe > 1) { | |
3446 | sbi->s_mb_group_prealloc = roundup( | |
3447 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
3448 | } | |
c9de560d | 3449 | |
730c213c | 3450 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 3451 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 3452 | ret = -ENOMEM; |
029b10c5 | 3453 | goto out; |
c9de560d | 3454 | } |
730c213c | 3455 | for_each_possible_cpu(i) { |
c9de560d | 3456 | struct ext4_locality_group *lg; |
730c213c | 3457 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 3458 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
3459 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
3460 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
3461 | spin_lock_init(&lg->lg_prealloc_lock); |
3462 | } | |
3463 | ||
10f0d2a5 | 3464 | if (bdev_nonrot(sb->s_bdev)) |
196e402a HS |
3465 | sbi->s_mb_max_linear_groups = 0; |
3466 | else | |
3467 | sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; | |
79a77c5a YJ |
3468 | /* init file for buddy data */ |
3469 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
3470 | if (ret != 0) |
3471 | goto out_free_locality_groups; | |
79a77c5a | 3472 | |
7aa0baea TM |
3473 | return 0; |
3474 | ||
3475 | out_free_locality_groups: | |
3476 | free_percpu(sbi->s_locality_groups); | |
3477 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 3478 | out: |
83e80a6e JK |
3479 | kfree(sbi->s_mb_avg_fragment_size); |
3480 | kfree(sbi->s_mb_avg_fragment_size_locks); | |
196e402a HS |
3481 | kfree(sbi->s_mb_largest_free_orders); |
3482 | kfree(sbi->s_mb_largest_free_orders_locks); | |
7aa0baea TM |
3483 | kfree(sbi->s_mb_offsets); |
3484 | sbi->s_mb_offsets = NULL; | |
3485 | kfree(sbi->s_mb_maxs); | |
3486 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 3487 | return ret; |
c9de560d AT |
3488 | } |
3489 | ||
955ce5f5 | 3490 | /* need to called with the ext4 group lock held */ |
d3df1453 | 3491 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
c9de560d AT |
3492 | { |
3493 | struct ext4_prealloc_space *pa; | |
3494 | struct list_head *cur, *tmp; | |
3495 | int count = 0; | |
3496 | ||
3497 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
3498 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3499 | list_del(&pa->pa_group_list); | |
3500 | count++; | |
688f05a0 | 3501 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d | 3502 | } |
d3df1453 | 3503 | return count; |
c9de560d AT |
3504 | } |
3505 | ||
3506 | int ext4_mb_release(struct super_block *sb) | |
3507 | { | |
8df9675f | 3508 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
3509 | ext4_group_t i; |
3510 | int num_meta_group_infos; | |
df3da4ea | 3511 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 3512 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 3513 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
d3df1453 | 3514 | int count; |
c9de560d | 3515 | |
55cdd0af WJ |
3516 | if (test_opt(sb, DISCARD)) { |
3517 | /* | |
3518 | * wait the discard work to drain all of ext4_free_data | |
3519 | */ | |
3520 | flush_work(&sbi->s_discard_work); | |
3521 | WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); | |
3522 | } | |
3523 | ||
c9de560d | 3524 | if (sbi->s_group_info) { |
8df9675f | 3525 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3526 | cond_resched(); |
c9de560d | 3527 | grinfo = ext4_get_group_info(sb, i); |
a3450215 | 3528 | mb_group_bb_bitmap_free(grinfo); |
c9de560d | 3529 | ext4_lock_group(sb, i); |
d3df1453 RH |
3530 | count = ext4_mb_cleanup_pa(grinfo); |
3531 | if (count) | |
3532 | mb_debug(sb, "mballoc: %d PAs left\n", | |
3533 | count); | |
c9de560d | 3534 | ext4_unlock_group(sb, i); |
fb1813f4 | 3535 | kmem_cache_free(cachep, grinfo); |
c9de560d | 3536 | } |
8df9675f | 3537 | num_meta_group_infos = (ngroups + |
c9de560d AT |
3538 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
3539 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
3540 | rcu_read_lock(); |
3541 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 3542 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
3543 | kfree(group_info[i]); |
3544 | kvfree(group_info); | |
3545 | rcu_read_unlock(); | |
c9de560d | 3546 | } |
83e80a6e JK |
3547 | kfree(sbi->s_mb_avg_fragment_size); |
3548 | kfree(sbi->s_mb_avg_fragment_size_locks); | |
196e402a HS |
3549 | kfree(sbi->s_mb_largest_free_orders); |
3550 | kfree(sbi->s_mb_largest_free_orders_locks); | |
c9de560d AT |
3551 | kfree(sbi->s_mb_offsets); |
3552 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 3553 | iput(sbi->s_buddy_cache); |
c9de560d | 3554 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
3555 | ext4_msg(sb, KERN_INFO, |
3556 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
3557 | atomic_read(&sbi->s_bal_allocated), |
3558 | atomic_read(&sbi->s_bal_reqs), | |
3559 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 | 3560 | ext4_msg(sb, KERN_INFO, |
a6c75eaf | 3561 | "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " |
9d8b9ec4 | 3562 | "%u 2^N hits, %u breaks, %u lost", |
c9de560d | 3563 | atomic_read(&sbi->s_bal_ex_scanned), |
a6c75eaf | 3564 | atomic_read(&sbi->s_bal_groups_scanned), |
c9de560d AT |
3565 | atomic_read(&sbi->s_bal_goals), |
3566 | atomic_read(&sbi->s_bal_2orders), | |
3567 | atomic_read(&sbi->s_bal_breaks), | |
3568 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 | 3569 | ext4_msg(sb, KERN_INFO, |
67d25186 HS |
3570 | "mballoc: %u generated and it took %llu", |
3571 | atomic_read(&sbi->s_mb_buddies_generated), | |
3572 | atomic64_read(&sbi->s_mb_generation_time)); | |
9d8b9ec4 TT |
3573 | ext4_msg(sb, KERN_INFO, |
3574 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
3575 | atomic_read(&sbi->s_mb_preallocated), |
3576 | atomic_read(&sbi->s_mb_discarded)); | |
3577 | } | |
3578 | ||
730c213c | 3579 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
3580 | |
3581 | return 0; | |
3582 | } | |
3583 | ||
77ca6cdf | 3584 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
3585 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
3586 | struct bio **biop) | |
5c521830 | 3587 | { |
5c521830 JZ |
3588 | ext4_fsblk_t discard_block; |
3589 | ||
84130193 TT |
3590 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
3591 | ext4_group_first_block_no(sb, block_group)); | |
3592 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
3593 | trace_ext4_discard_blocks(sb, |
3594 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
3595 | if (biop) { |
3596 | return __blkdev_issue_discard(sb->s_bdev, | |
3597 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
3598 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
44abff2c | 3599 | GFP_NOFS, biop); |
a0154344 DJ |
3600 | } else |
3601 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
3602 | } |
3603 | ||
a0154344 DJ |
3604 | static void ext4_free_data_in_buddy(struct super_block *sb, |
3605 | struct ext4_free_data *entry) | |
c9de560d | 3606 | { |
c9de560d | 3607 | struct ext4_buddy e4b; |
c894058d | 3608 | struct ext4_group_info *db; |
d9f34504 | 3609 | int err, count = 0, count2 = 0; |
c9de560d | 3610 | |
d3df1453 | 3611 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
18aadd47 | 3612 | entry->efd_count, entry->efd_group, entry); |
c9de560d | 3613 | |
18aadd47 BJ |
3614 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
3615 | /* we expect to find existing buddy because it's pinned */ | |
3616 | BUG_ON(err != 0); | |
b90f6870 | 3617 | |
d08854f5 TT |
3618 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
3619 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
3620 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 3621 | |
18aadd47 BJ |
3622 | db = e4b.bd_info; |
3623 | /* there are blocks to put in buddy to make them really free */ | |
3624 | count += entry->efd_count; | |
3625 | count2++; | |
3626 | ext4_lock_group(sb, entry->efd_group); | |
3627 | /* Take it out of per group rb tree */ | |
3628 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
3629 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 3630 | |
18aadd47 BJ |
3631 | /* |
3632 | * Clear the trimmed flag for the group so that the next | |
3633 | * ext4_trim_fs can trim it. | |
3634 | * If the volume is mounted with -o discard, online discard | |
3635 | * is supported and the free blocks will be trimmed online. | |
3636 | */ | |
3637 | if (!test_opt(sb, DISCARD)) | |
3638 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 3639 | |
18aadd47 BJ |
3640 | if (!db->bb_free_root.rb_node) { |
3641 | /* No more items in the per group rb tree | |
3642 | * balance refcounts from ext4_mb_free_metadata() | |
3643 | */ | |
09cbfeaf KS |
3644 | put_page(e4b.bd_buddy_page); |
3645 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 3646 | } |
18aadd47 | 3647 | ext4_unlock_group(sb, entry->efd_group); |
18aadd47 | 3648 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 3649 | |
d3df1453 RH |
3650 | mb_debug(sb, "freed %d blocks in %d structures\n", count, |
3651 | count2); | |
c9de560d AT |
3652 | } |
3653 | ||
a0154344 DJ |
3654 | /* |
3655 | * This function is called by the jbd2 layer once the commit has finished, | |
3656 | * so we know we can free the blocks that were released with that commit. | |
3657 | */ | |
3658 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
3659 | { | |
3660 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3661 | struct ext4_free_data *entry, *tmp; | |
a0154344 DJ |
3662 | struct list_head freed_data_list; |
3663 | struct list_head *cut_pos = NULL; | |
55cdd0af | 3664 | bool wake; |
a0154344 DJ |
3665 | |
3666 | INIT_LIST_HEAD(&freed_data_list); | |
3667 | ||
3668 | spin_lock(&sbi->s_md_lock); | |
3669 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
3670 | if (entry->efd_tid != commit_tid) | |
3671 | break; | |
3672 | cut_pos = &entry->efd_list; | |
3673 | } | |
3674 | if (cut_pos) | |
3675 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
3676 | cut_pos); | |
3677 | spin_unlock(&sbi->s_md_lock); | |
3678 | ||
55cdd0af WJ |
3679 | list_for_each_entry(entry, &freed_data_list, efd_list) |
3680 | ext4_free_data_in_buddy(sb, entry); | |
a0154344 | 3681 | |
55cdd0af WJ |
3682 | if (test_opt(sb, DISCARD)) { |
3683 | spin_lock(&sbi->s_md_lock); | |
3684 | wake = list_empty(&sbi->s_discard_list); | |
3685 | list_splice_tail(&freed_data_list, &sbi->s_discard_list); | |
3686 | spin_unlock(&sbi->s_md_lock); | |
3687 | if (wake) | |
3688 | queue_work(system_unbound_wq, &sbi->s_discard_work); | |
3689 | } else { | |
3690 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
3691 | kmem_cache_free(ext4_free_data_cachep, entry); | |
a0154344 | 3692 | } |
a0154344 DJ |
3693 | } |
3694 | ||
5dabfc78 | 3695 | int __init ext4_init_mballoc(void) |
c9de560d | 3696 | { |
16828088 TT |
3697 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
3698 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 3699 | if (ext4_pspace_cachep == NULL) |
f283529a | 3700 | goto out; |
c9de560d | 3701 | |
16828088 TT |
3702 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
3703 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3704 | if (ext4_ac_cachep == NULL) |
3705 | goto out_pa_free; | |
c894058d | 3706 | |
18aadd47 BJ |
3707 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
3708 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3709 | if (ext4_free_data_cachep == NULL) |
3710 | goto out_ac_free; | |
3711 | ||
c9de560d | 3712 | return 0; |
f283529a RH |
3713 | |
3714 | out_ac_free: | |
3715 | kmem_cache_destroy(ext4_ac_cachep); | |
3716 | out_pa_free: | |
3717 | kmem_cache_destroy(ext4_pspace_cachep); | |
3718 | out: | |
3719 | return -ENOMEM; | |
c9de560d AT |
3720 | } |
3721 | ||
5dabfc78 | 3722 | void ext4_exit_mballoc(void) |
c9de560d | 3723 | { |
60e6679e | 3724 | /* |
3e03f9ca JDB |
3725 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
3726 | * before destroying the slab cache. | |
3727 | */ | |
3728 | rcu_barrier(); | |
c9de560d | 3729 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 3730 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 3731 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 3732 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
3733 | } |
3734 | ||
3735 | ||
3736 | /* | |
73b2c716 | 3737 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
3738 | * Returns 0 if success or error code |
3739 | */ | |
4ddfef7b ES |
3740 | static noinline_for_stack int |
3741 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 3742 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
3743 | { |
3744 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
3745 | struct ext4_group_desc *gdp; |
3746 | struct buffer_head *gdp_bh; | |
3747 | struct ext4_sb_info *sbi; | |
3748 | struct super_block *sb; | |
3749 | ext4_fsblk_t block; | |
519deca0 | 3750 | int err, len; |
c9de560d AT |
3751 | |
3752 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3753 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
3754 | ||
3755 | sb = ac->ac_sb; | |
3756 | sbi = EXT4_SB(sb); | |
c9de560d | 3757 | |
574ca174 | 3758 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
3759 | if (IS_ERR(bitmap_bh)) { |
3760 | err = PTR_ERR(bitmap_bh); | |
3761 | bitmap_bh = NULL; | |
c9de560d | 3762 | goto out_err; |
9008a58e | 3763 | } |
c9de560d | 3764 | |
5d601255 | 3765 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
3766 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
3767 | EXT4_JTR_NONE); | |
c9de560d AT |
3768 | if (err) |
3769 | goto out_err; | |
3770 | ||
3771 | err = -EIO; | |
3772 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3773 | if (!gdp) | |
3774 | goto out_err; | |
3775 | ||
a9df9a49 | 3776 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3777 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3778 | |
5d601255 | 3779 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
188c299e | 3780 | err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE); |
c9de560d AT |
3781 | if (err) |
3782 | goto out_err; | |
3783 | ||
bda00de7 | 3784 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3785 | |
53accfa9 | 3786 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
ce9f24cc | 3787 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
12062ddd | 3788 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3789 | "fs metadata", block, block+len); |
519deca0 | 3790 | /* File system mounted not to panic on error |
554a5ccc | 3791 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3792 | * We leak some of the blocks here. |
3793 | */ | |
955ce5f5 | 3794 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
123e3016 | 3795 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
c3e94d1d | 3796 | ac->ac_b_ex.fe_len); |
955ce5f5 | 3797 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3798 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3799 | if (!err) |
554a5ccc | 3800 | err = -EFSCORRUPTED; |
519deca0 | 3801 | goto out_err; |
c9de560d | 3802 | } |
955ce5f5 AK |
3803 | |
3804 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3805 | #ifdef AGGRESSIVE_CHECK |
3806 | { | |
3807 | int i; | |
3808 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3809 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3810 | bitmap_bh->b_data)); | |
3811 | } | |
3812 | } | |
3813 | #endif | |
123e3016 | 3814 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
c3e94d1d | 3815 | ac->ac_b_ex.fe_len); |
8844618d TT |
3816 | if (ext4_has_group_desc_csum(sb) && |
3817 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3818 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3819 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3820 | ext4_free_clusters_after_init(sb, |
021b65bb | 3821 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3822 | } |
021b65bb TT |
3823 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3824 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3825 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3826 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3827 | |
3828 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3829 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3830 | /* |
6bc6e63f | 3831 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3832 | */ |
6bc6e63f AK |
3833 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3834 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3835 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3836 | reserv_clstrs); | |
c9de560d | 3837 | |
772cb7c8 JS |
3838 | if (sbi->s_log_groups_per_flex) { |
3839 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3840 | ac->ac_b_ex.fe_group); | |
90ba983f | 3841 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3842 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3843 | flex_group)->free_clusters); | |
772cb7c8 JS |
3844 | } |
3845 | ||
0390131b | 3846 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3847 | if (err) |
3848 | goto out_err; | |
0390131b | 3849 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3850 | |
3851 | out_err: | |
42a10add | 3852 | brelse(bitmap_bh); |
c9de560d AT |
3853 | return err; |
3854 | } | |
3855 | ||
8016e29f HS |
3856 | /* |
3857 | * Idempotent helper for Ext4 fast commit replay path to set the state of | |
3858 | * blocks in bitmaps and update counters. | |
3859 | */ | |
3860 | void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, | |
3861 | int len, int state) | |
3862 | { | |
3863 | struct buffer_head *bitmap_bh = NULL; | |
3864 | struct ext4_group_desc *gdp; | |
3865 | struct buffer_head *gdp_bh; | |
3866 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3867 | ext4_group_t group; | |
3868 | ext4_grpblk_t blkoff; | |
a5c0e2fd | 3869 | int i, err; |
8016e29f | 3870 | int already; |
bfdc502a | 3871 | unsigned int clen, clen_changed, thisgrp_len; |
8016e29f | 3872 | |
bfdc502a RH |
3873 | while (len > 0) { |
3874 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
8016e29f | 3875 | |
bfdc502a RH |
3876 | /* |
3877 | * Check to see if we are freeing blocks across a group | |
3878 | * boundary. | |
3879 | * In case of flex_bg, this can happen that (block, len) may | |
3880 | * span across more than one group. In that case we need to | |
3881 | * get the corresponding group metadata to work with. | |
3882 | * For this we have goto again loop. | |
3883 | */ | |
3884 | thisgrp_len = min_t(unsigned int, (unsigned int)len, | |
3885 | EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); | |
3886 | clen = EXT4_NUM_B2C(sbi, thisgrp_len); | |
3887 | ||
8c91c579 RH |
3888 | if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { |
3889 | ext4_error(sb, "Marking blocks in system zone - " | |
3890 | "Block = %llu, len = %u", | |
3891 | block, thisgrp_len); | |
3892 | bitmap_bh = NULL; | |
3893 | break; | |
3894 | } | |
3895 | ||
bfdc502a RH |
3896 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
3897 | if (IS_ERR(bitmap_bh)) { | |
3898 | err = PTR_ERR(bitmap_bh); | |
3899 | bitmap_bh = NULL; | |
3900 | break; | |
3901 | } | |
8016e29f | 3902 | |
bfdc502a RH |
3903 | err = -EIO; |
3904 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
3905 | if (!gdp) | |
3906 | break; | |
8016e29f | 3907 | |
bfdc502a RH |
3908 | ext4_lock_group(sb, group); |
3909 | already = 0; | |
3910 | for (i = 0; i < clen; i++) | |
3911 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == | |
3912 | !state) | |
3913 | already++; | |
8016e29f | 3914 | |
bfdc502a RH |
3915 | clen_changed = clen - already; |
3916 | if (state) | |
123e3016 | 3917 | mb_set_bits(bitmap_bh->b_data, blkoff, clen); |
bfdc502a | 3918 | else |
bd8247ee | 3919 | mb_clear_bits(bitmap_bh->b_data, blkoff, clen); |
bfdc502a RH |
3920 | if (ext4_has_group_desc_csum(sb) && |
3921 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
3922 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
3923 | ext4_free_group_clusters_set(sb, gdp, | |
3924 | ext4_free_clusters_after_init(sb, group, gdp)); | |
3925 | } | |
3926 | if (state) | |
3927 | clen = ext4_free_group_clusters(sb, gdp) - clen_changed; | |
3928 | else | |
3929 | clen = ext4_free_group_clusters(sb, gdp) + clen_changed; | |
8016e29f | 3930 | |
bfdc502a RH |
3931 | ext4_free_group_clusters_set(sb, gdp, clen); |
3932 | ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh); | |
3933 | ext4_group_desc_csum_set(sb, group, gdp); | |
8016e29f | 3934 | |
bfdc502a | 3935 | ext4_unlock_group(sb, group); |
8016e29f | 3936 | |
bfdc502a RH |
3937 | if (sbi->s_log_groups_per_flex) { |
3938 | ext4_group_t flex_group = ext4_flex_group(sbi, group); | |
3939 | struct flex_groups *fg = sbi_array_rcu_deref(sbi, | |
3940 | s_flex_groups, flex_group); | |
8016e29f | 3941 | |
bfdc502a RH |
3942 | if (state) |
3943 | atomic64_sub(clen_changed, &fg->free_clusters); | |
3944 | else | |
3945 | atomic64_add(clen_changed, &fg->free_clusters); | |
3946 | ||
3947 | } | |
3948 | ||
3949 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
3950 | if (err) | |
3951 | break; | |
3952 | sync_dirty_buffer(bitmap_bh); | |
3953 | err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
3954 | sync_dirty_buffer(gdp_bh); | |
3955 | if (err) | |
3956 | break; | |
3957 | ||
3958 | block += thisgrp_len; | |
3959 | len -= thisgrp_len; | |
3960 | brelse(bitmap_bh); | |
3961 | BUG_ON(len < 0); | |
8016e29f HS |
3962 | } |
3963 | ||
8016e29f | 3964 | if (err) |
bfdc502a | 3965 | brelse(bitmap_bh); |
8016e29f HS |
3966 | } |
3967 | ||
c9de560d AT |
3968 | /* |
3969 | * here we normalize request for locality group | |
d7a1fee1 DE |
3970 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3971 | * s_strip if we set the same via mount option. | |
3972 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3973 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3974 | * |
3975 | * XXX: should we try to preallocate more than the group has now? | |
3976 | */ | |
3977 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3978 | { | |
3979 | struct super_block *sb = ac->ac_sb; | |
3980 | struct ext4_locality_group *lg = ac->ac_lg; | |
3981 | ||
3982 | BUG_ON(lg == NULL); | |
d7a1fee1 | 3983 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
d3df1453 | 3984 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
c9de560d AT |
3985 | } |
3986 | ||
3987 | /* | |
3988 | * Normalization means making request better in terms of | |
3989 | * size and alignment | |
3990 | */ | |
4ddfef7b ES |
3991 | static noinline_for_stack void |
3992 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3993 | struct ext4_allocation_request *ar) |
3994 | { | |
53accfa9 | 3995 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3996 | int bsbits, max; |
3997 | ext4_lblk_t end; | |
1592d2c5 CW |
3998 | loff_t size, start_off; |
3999 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 4000 | ext4_lblk_t start; |
c9de560d | 4001 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 4002 | struct ext4_prealloc_space *pa; |
c9de560d AT |
4003 | |
4004 | /* do normalize only data requests, metadata requests | |
4005 | do not need preallocation */ | |
4006 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4007 | return; | |
4008 | ||
4009 | /* sometime caller may want exact blocks */ | |
4010 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4011 | return; | |
4012 | ||
4013 | /* caller may indicate that preallocation isn't | |
4014 | * required (it's a tail, for example) */ | |
4015 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
4016 | return; | |
4017 | ||
4018 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
4019 | ext4_mb_normalize_group_request(ac); | |
4020 | return ; | |
4021 | } | |
4022 | ||
4023 | bsbits = ac->ac_sb->s_blocksize_bits; | |
4024 | ||
4025 | /* first, let's learn actual file size | |
4026 | * given current request is allocated */ | |
53accfa9 | 4027 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
4028 | size = size << bsbits; |
4029 | if (size < i_size_read(ac->ac_inode)) | |
4030 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 4031 | orig_size = size; |
c9de560d | 4032 | |
1930479c VC |
4033 | /* max size of free chunks */ |
4034 | max = 2 << bsbits; | |
c9de560d | 4035 | |
1930479c VC |
4036 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
4037 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
4038 | |
4039 | /* first, try to predict filesize */ | |
4040 | /* XXX: should this table be tunable? */ | |
4041 | start_off = 0; | |
4042 | if (size <= 16 * 1024) { | |
4043 | size = 16 * 1024; | |
4044 | } else if (size <= 32 * 1024) { | |
4045 | size = 32 * 1024; | |
4046 | } else if (size <= 64 * 1024) { | |
4047 | size = 64 * 1024; | |
4048 | } else if (size <= 128 * 1024) { | |
4049 | size = 128 * 1024; | |
4050 | } else if (size <= 256 * 1024) { | |
4051 | size = 256 * 1024; | |
4052 | } else if (size <= 512 * 1024) { | |
4053 | size = 512 * 1024; | |
4054 | } else if (size <= 1024 * 1024) { | |
4055 | size = 1024 * 1024; | |
1930479c | 4056 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 4057 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
4058 | (21 - bsbits)) << 21; |
4059 | size = 2 * 1024 * 1024; | |
4060 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
4061 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4062 | (22 - bsbits)) << 22; | |
4063 | size = 4 * 1024 * 1024; | |
4064 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 4065 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
4066 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4067 | (23 - bsbits)) << 23; | |
4068 | size = 8 * 1024 * 1024; | |
4069 | } else { | |
b27b1535 XW |
4070 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
4071 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
4072 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 4073 | } |
5a0790c2 AK |
4074 | size = size >> bsbits; |
4075 | start = start_off >> bsbits; | |
c9de560d | 4076 | |
a08f789d BL |
4077 | /* |
4078 | * For tiny groups (smaller than 8MB) the chosen allocation | |
4079 | * alignment may be larger than group size. Make sure the | |
4080 | * alignment does not move allocation to a different group which | |
4081 | * makes mballoc fail assertions later. | |
4082 | */ | |
4083 | start = max(start, rounddown(ac->ac_o_ex.fe_logical, | |
4084 | (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); | |
4085 | ||
c9de560d AT |
4086 | /* don't cover already allocated blocks in selected range */ |
4087 | if (ar->pleft && start <= ar->lleft) { | |
4088 | size -= ar->lleft + 1 - start; | |
4089 | start = ar->lleft + 1; | |
4090 | } | |
4091 | if (ar->pright && start + size - 1 >= ar->lright) | |
4092 | size -= start + size - ar->lright; | |
4093 | ||
cd648b8a JK |
4094 | /* |
4095 | * Trim allocation request for filesystems with artificially small | |
4096 | * groups. | |
4097 | */ | |
4098 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
4099 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
4100 | ||
c9de560d AT |
4101 | end = start + size; |
4102 | ||
4103 | /* check we don't cross already preallocated blocks */ | |
4104 | rcu_read_lock(); | |
9a0762c5 | 4105 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 4106 | ext4_lblk_t pa_end; |
c9de560d | 4107 | |
c9de560d AT |
4108 | if (pa->pa_deleted) |
4109 | continue; | |
4110 | spin_lock(&pa->pa_lock); | |
4111 | if (pa->pa_deleted) { | |
4112 | spin_unlock(&pa->pa_lock); | |
4113 | continue; | |
4114 | } | |
4115 | ||
53accfa9 TT |
4116 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
4117 | pa->pa_len); | |
c9de560d AT |
4118 | |
4119 | /* PA must not overlap original request */ | |
4120 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
4121 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
4122 | ||
38877f4e ES |
4123 | /* skip PAs this normalized request doesn't overlap with */ |
4124 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
4125 | spin_unlock(&pa->pa_lock); |
4126 | continue; | |
4127 | } | |
4128 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
4129 | ||
38877f4e | 4130 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
4131 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
4132 | BUG_ON(pa_end < start); | |
4133 | start = pa_end; | |
38877f4e | 4134 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
4135 | BUG_ON(pa->pa_lstart > end); |
4136 | end = pa->pa_lstart; | |
4137 | } | |
4138 | spin_unlock(&pa->pa_lock); | |
4139 | } | |
4140 | rcu_read_unlock(); | |
4141 | size = end - start; | |
4142 | ||
4143 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
4144 | rcu_read_lock(); | |
9a0762c5 | 4145 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 4146 | ext4_lblk_t pa_end; |
53accfa9 | 4147 | |
c9de560d AT |
4148 | spin_lock(&pa->pa_lock); |
4149 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
4150 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
4151 | pa->pa_len); | |
c9de560d AT |
4152 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
4153 | } | |
4154 | spin_unlock(&pa->pa_lock); | |
4155 | } | |
4156 | rcu_read_unlock(); | |
4157 | ||
cf4ff938 BL |
4158 | /* |
4159 | * In this function "start" and "size" are normalized for better | |
4160 | * alignment and length such that we could preallocate more blocks. | |
4161 | * This normalization is done such that original request of | |
4162 | * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and | |
4163 | * "size" boundaries. | |
4164 | * (Note fe_len can be relaxed since FS block allocation API does not | |
4165 | * provide gurantee on number of contiguous blocks allocation since that | |
4166 | * depends upon free space left, etc). | |
4167 | * In case of inode pa, later we use the allocated blocks | |
4168 | * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated | |
4169 | * range of goal/best blocks [start, size] to put it at the | |
4170 | * ac_o_ex.fe_logical extent of this inode. | |
4171 | * (See ext4_mb_use_inode_pa() for more details) | |
4172 | */ | |
4173 | if (start + size <= ac->ac_o_ex.fe_logical || | |
c9de560d | 4174 | start > ac->ac_o_ex.fe_logical) { |
9d8b9ec4 TT |
4175 | ext4_msg(ac->ac_sb, KERN_ERR, |
4176 | "start %lu, size %lu, fe_logical %lu", | |
4177 | (unsigned long) start, (unsigned long) size, | |
4178 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 4179 | BUG(); |
c9de560d | 4180 | } |
b5b60778 | 4181 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
4182 | |
4183 | /* now prepare goal request */ | |
4184 | ||
4185 | /* XXX: is it better to align blocks WRT to logical | |
4186 | * placement or satisfy big request as is */ | |
4187 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 4188 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
4189 | |
4190 | /* define goal start in order to merge */ | |
4191 | if (ar->pright && (ar->lright == (start + size))) { | |
4192 | /* merge to the right */ | |
4193 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
4194 | &ac->ac_f_ex.fe_group, | |
4195 | &ac->ac_f_ex.fe_start); | |
4196 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
4197 | } | |
4198 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
4199 | /* merge to the left */ | |
4200 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
4201 | &ac->ac_f_ex.fe_group, | |
4202 | &ac->ac_f_ex.fe_start); | |
4203 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
4204 | } | |
4205 | ||
d3df1453 RH |
4206 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
4207 | orig_size, start); | |
c9de560d AT |
4208 | } |
4209 | ||
4210 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
4211 | { | |
4212 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4213 | ||
a6c75eaf | 4214 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { |
c9de560d AT |
4215 | atomic_inc(&sbi->s_bal_reqs); |
4216 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 4217 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
4218 | atomic_inc(&sbi->s_bal_success); |
4219 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
a6c75eaf | 4220 | atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); |
c9de560d AT |
4221 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && |
4222 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
4223 | atomic_inc(&sbi->s_bal_goals); | |
4224 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
4225 | atomic_inc(&sbi->s_bal_breaks); | |
4226 | } | |
4227 | ||
296c355c TT |
4228 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
4229 | trace_ext4_mballoc_alloc(ac); | |
4230 | else | |
4231 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
4232 | } |
4233 | ||
b844167e CW |
4234 | /* |
4235 | * Called on failure; free up any blocks from the inode PA for this | |
4236 | * context. We don't need this for MB_GROUP_PA because we only change | |
4237 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
4238 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
4239 | */ | |
4240 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
4241 | { | |
4242 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
4243 | struct ext4_buddy e4b; |
4244 | int err; | |
b844167e | 4245 | |
86f0afd4 | 4246 | if (pa == NULL) { |
c99d1e6e TT |
4247 | if (ac->ac_f_ex.fe_len == 0) |
4248 | return; | |
86f0afd4 TT |
4249 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
4250 | if (err) { | |
4251 | /* | |
4252 | * This should never happen since we pin the | |
4253 | * pages in the ext4_allocation_context so | |
4254 | * ext4_mb_load_buddy() should never fail. | |
4255 | */ | |
4256 | WARN(1, "mb_load_buddy failed (%d)", err); | |
4257 | return; | |
4258 | } | |
4259 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
4260 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
4261 | ac->ac_f_ex.fe_len); | |
4262 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 4263 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
4264 | return; |
4265 | } | |
4266 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 4267 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
4268 | } |
4269 | ||
c9de560d AT |
4270 | /* |
4271 | * use blocks preallocated to inode | |
4272 | */ | |
4273 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
4274 | struct ext4_prealloc_space *pa) | |
4275 | { | |
53accfa9 | 4276 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
4277 | ext4_fsblk_t start; |
4278 | ext4_fsblk_t end; | |
4279 | int len; | |
4280 | ||
4281 | /* found preallocated blocks, use them */ | |
4282 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
4283 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
4284 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
4285 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
4286 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
4287 | &ac->ac_b_ex.fe_start); | |
4288 | ac->ac_b_ex.fe_len = len; | |
4289 | ac->ac_status = AC_STATUS_FOUND; | |
4290 | ac->ac_pa = pa; | |
4291 | ||
4292 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 4293 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
4294 | BUG_ON(pa->pa_free < len); |
4295 | pa->pa_free -= len; | |
4296 | ||
d3df1453 | 4297 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
4298 | } |
4299 | ||
4300 | /* | |
4301 | * use blocks preallocated to locality group | |
4302 | */ | |
4303 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
4304 | struct ext4_prealloc_space *pa) | |
4305 | { | |
03cddb80 | 4306 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 4307 | |
c9de560d AT |
4308 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
4309 | &ac->ac_b_ex.fe_group, | |
4310 | &ac->ac_b_ex.fe_start); | |
4311 | ac->ac_b_ex.fe_len = len; | |
4312 | ac->ac_status = AC_STATUS_FOUND; | |
4313 | ac->ac_pa = pa; | |
4314 | ||
4315 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 4316 | * possible race when the group is being loaded concurrently |
c9de560d | 4317 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
4318 | * in on-disk bitmap -- see ext4_mb_release_context() |
4319 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 4320 | */ |
d3df1453 RH |
4321 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
4322 | pa->pa_lstart-len, len, pa); | |
c9de560d AT |
4323 | } |
4324 | ||
5e745b04 AK |
4325 | /* |
4326 | * Return the prealloc space that have minimal distance | |
4327 | * from the goal block. @cpa is the prealloc | |
4328 | * space that is having currently known minimal distance | |
4329 | * from the goal block. | |
4330 | */ | |
4331 | static struct ext4_prealloc_space * | |
4332 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
4333 | struct ext4_prealloc_space *pa, | |
4334 | struct ext4_prealloc_space *cpa) | |
4335 | { | |
4336 | ext4_fsblk_t cur_distance, new_distance; | |
4337 | ||
4338 | if (cpa == NULL) { | |
4339 | atomic_inc(&pa->pa_count); | |
4340 | return pa; | |
4341 | } | |
79211c8e AM |
4342 | cur_distance = abs(goal_block - cpa->pa_pstart); |
4343 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 4344 | |
5a54b2f1 | 4345 | if (cur_distance <= new_distance) |
5e745b04 AK |
4346 | return cpa; |
4347 | ||
4348 | /* drop the previous reference */ | |
4349 | atomic_dec(&cpa->pa_count); | |
4350 | atomic_inc(&pa->pa_count); | |
4351 | return pa; | |
4352 | } | |
4353 | ||
c9de560d AT |
4354 | /* |
4355 | * search goal blocks in preallocated space | |
4356 | */ | |
4fca8f07 | 4357 | static noinline_for_stack bool |
4ddfef7b | 4358 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 4359 | { |
53accfa9 | 4360 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 4361 | int order, i; |
c9de560d AT |
4362 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
4363 | struct ext4_locality_group *lg; | |
5e745b04 AK |
4364 | struct ext4_prealloc_space *pa, *cpa = NULL; |
4365 | ext4_fsblk_t goal_block; | |
c9de560d AT |
4366 | |
4367 | /* only data can be preallocated */ | |
4368 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 4369 | return false; |
c9de560d AT |
4370 | |
4371 | /* first, try per-file preallocation */ | |
4372 | rcu_read_lock(); | |
9a0762c5 | 4373 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
4374 | |
4375 | /* all fields in this condition don't change, | |
4376 | * so we can skip locking for them */ | |
4377 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
4378 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
4379 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
4380 | continue; |
4381 | ||
fb0a387d | 4382 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 4383 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
4384 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
4385 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
4386 | continue; |
4387 | ||
c9de560d AT |
4388 | /* found preallocated blocks, use them */ |
4389 | spin_lock(&pa->pa_lock); | |
4390 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
4391 | atomic_inc(&pa->pa_count); | |
4392 | ext4_mb_use_inode_pa(ac, pa); | |
4393 | spin_unlock(&pa->pa_lock); | |
4394 | ac->ac_criteria = 10; | |
4395 | rcu_read_unlock(); | |
4fca8f07 | 4396 | return true; |
c9de560d AT |
4397 | } |
4398 | spin_unlock(&pa->pa_lock); | |
4399 | } | |
4400 | rcu_read_unlock(); | |
4401 | ||
4402 | /* can we use group allocation? */ | |
4403 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 4404 | return false; |
c9de560d AT |
4405 | |
4406 | /* inode may have no locality group for some reason */ | |
4407 | lg = ac->ac_lg; | |
4408 | if (lg == NULL) | |
4fca8f07 | 4409 | return false; |
6be2ded1 AK |
4410 | order = fls(ac->ac_o_ex.fe_len) - 1; |
4411 | if (order > PREALLOC_TB_SIZE - 1) | |
4412 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4413 | order = PREALLOC_TB_SIZE - 1; | |
4414 | ||
bda00de7 | 4415 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
4416 | /* |
4417 | * search for the prealloc space that is having | |
4418 | * minimal distance from the goal block. | |
4419 | */ | |
6be2ded1 AK |
4420 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
4421 | rcu_read_lock(); | |
4422 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
4423 | pa_inode_list) { | |
4424 | spin_lock(&pa->pa_lock); | |
4425 | if (pa->pa_deleted == 0 && | |
4426 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
4427 | |
4428 | cpa = ext4_mb_check_group_pa(goal_block, | |
4429 | pa, cpa); | |
6be2ded1 | 4430 | } |
c9de560d | 4431 | spin_unlock(&pa->pa_lock); |
c9de560d | 4432 | } |
6be2ded1 | 4433 | rcu_read_unlock(); |
c9de560d | 4434 | } |
5e745b04 AK |
4435 | if (cpa) { |
4436 | ext4_mb_use_group_pa(ac, cpa); | |
4437 | ac->ac_criteria = 20; | |
4fca8f07 | 4438 | return true; |
5e745b04 | 4439 | } |
4fca8f07 | 4440 | return false; |
c9de560d AT |
4441 | } |
4442 | ||
7a2fcbf7 AK |
4443 | /* |
4444 | * the function goes through all block freed in the group | |
4445 | * but not yet committed and marks them used in in-core bitmap. | |
4446 | * buddy must be generated from this bitmap | |
955ce5f5 | 4447 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
4448 | */ |
4449 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
4450 | ext4_group_t group) | |
4451 | { | |
4452 | struct rb_node *n; | |
4453 | struct ext4_group_info *grp; | |
4454 | struct ext4_free_data *entry; | |
4455 | ||
4456 | grp = ext4_get_group_info(sb, group); | |
4457 | n = rb_first(&(grp->bb_free_root)); | |
4458 | ||
4459 | while (n) { | |
18aadd47 | 4460 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
123e3016 | 4461 | mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); |
7a2fcbf7 AK |
4462 | n = rb_next(n); |
4463 | } | |
4464 | return; | |
4465 | } | |
4466 | ||
c9de560d AT |
4467 | /* |
4468 | * the function goes through all preallocation in this group and marks them | |
4469 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 4470 | * Need to be called with ext4 group lock held |
c9de560d | 4471 | */ |
089ceecc ES |
4472 | static noinline_for_stack |
4473 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
4474 | ext4_group_t group) |
4475 | { | |
4476 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4477 | struct ext4_prealloc_space *pa; | |
4478 | struct list_head *cur; | |
4479 | ext4_group_t groupnr; | |
4480 | ext4_grpblk_t start; | |
4481 | int preallocated = 0; | |
c9de560d AT |
4482 | int len; |
4483 | ||
4484 | /* all form of preallocation discards first load group, | |
4485 | * so the only competing code is preallocation use. | |
4486 | * we don't need any locking here | |
4487 | * notice we do NOT ignore preallocations with pa_deleted | |
4488 | * otherwise we could leave used blocks available for | |
4489 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
4490 | * is dropping preallocation | |
4491 | */ | |
4492 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4493 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
4494 | spin_lock(&pa->pa_lock); | |
4495 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4496 | &groupnr, &start); | |
4497 | len = pa->pa_len; | |
4498 | spin_unlock(&pa->pa_lock); | |
4499 | if (unlikely(len == 0)) | |
4500 | continue; | |
4501 | BUG_ON(groupnr != group); | |
123e3016 | 4502 | mb_set_bits(bitmap, start, len); |
c9de560d | 4503 | preallocated += len; |
c9de560d | 4504 | } |
d3df1453 | 4505 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
4506 | } |
4507 | ||
27bc446e | 4508 | static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
4509 | struct ext4_prealloc_space *pa) | |
4510 | { | |
4511 | struct ext4_inode_info *ei; | |
4512 | ||
4513 | if (pa->pa_deleted) { | |
4514 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", | |
4515 | pa->pa_type, pa->pa_pstart, pa->pa_lstart, | |
4516 | pa->pa_len); | |
4517 | return; | |
4518 | } | |
4519 | ||
4520 | pa->pa_deleted = 1; | |
4521 | ||
4522 | if (pa->pa_type == MB_INODE_PA) { | |
4523 | ei = EXT4_I(pa->pa_inode); | |
4524 | atomic_dec(&ei->i_prealloc_active); | |
4525 | } | |
4526 | } | |
4527 | ||
c9de560d AT |
4528 | static void ext4_mb_pa_callback(struct rcu_head *head) |
4529 | { | |
4530 | struct ext4_prealloc_space *pa; | |
4531 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
4532 | |
4533 | BUG_ON(atomic_read(&pa->pa_count)); | |
4534 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
4535 | kmem_cache_free(ext4_pspace_cachep, pa); |
4536 | } | |
4537 | ||
4538 | /* | |
4539 | * drops a reference to preallocated space descriptor | |
4540 | * if this was the last reference and the space is consumed | |
4541 | */ | |
4542 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
4543 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
4544 | { | |
a9df9a49 | 4545 | ext4_group_t grp; |
d33a1976 | 4546 | ext4_fsblk_t grp_blk; |
c9de560d | 4547 | |
c9de560d AT |
4548 | /* in this short window concurrent discard can set pa_deleted */ |
4549 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
4550 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
4551 | spin_unlock(&pa->pa_lock); | |
4552 | return; | |
4553 | } | |
4554 | ||
c9de560d AT |
4555 | if (pa->pa_deleted == 1) { |
4556 | spin_unlock(&pa->pa_lock); | |
4557 | return; | |
4558 | } | |
4559 | ||
27bc446e | 4560 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4561 | spin_unlock(&pa->pa_lock); |
4562 | ||
d33a1976 | 4563 | grp_blk = pa->pa_pstart; |
60e6679e | 4564 | /* |
cc0fb9ad AK |
4565 | * If doing group-based preallocation, pa_pstart may be in the |
4566 | * next group when pa is used up | |
4567 | */ | |
4568 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
4569 | grp_blk--; |
4570 | ||
bd86298e | 4571 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
4572 | |
4573 | /* | |
4574 | * possible race: | |
4575 | * | |
4576 | * P1 (buddy init) P2 (regular allocation) | |
4577 | * find block B in PA | |
4578 | * copy on-disk bitmap to buddy | |
4579 | * mark B in on-disk bitmap | |
4580 | * drop PA from group | |
4581 | * mark all PAs in buddy | |
4582 | * | |
4583 | * thus, P1 initializes buddy with B available. to prevent this | |
4584 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
4585 | * against that pair | |
4586 | */ | |
4587 | ext4_lock_group(sb, grp); | |
4588 | list_del(&pa->pa_group_list); | |
4589 | ext4_unlock_group(sb, grp); | |
4590 | ||
4591 | spin_lock(pa->pa_obj_lock); | |
4592 | list_del_rcu(&pa->pa_inode_list); | |
4593 | spin_unlock(pa->pa_obj_lock); | |
4594 | ||
4595 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4596 | } | |
4597 | ||
4598 | /* | |
4599 | * creates new preallocated space for given inode | |
4600 | */ | |
53f86b17 | 4601 | static noinline_for_stack void |
4ddfef7b | 4602 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4603 | { |
4604 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 4605 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
4606 | struct ext4_prealloc_space *pa; |
4607 | struct ext4_group_info *grp; | |
4608 | struct ext4_inode_info *ei; | |
4609 | ||
4610 | /* preallocate only when found space is larger then requested */ | |
4611 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4612 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4613 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4614 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4615 | |
53f86b17 | 4616 | pa = ac->ac_pa; |
c9de560d AT |
4617 | |
4618 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
4619 | int winl; | |
4620 | int wins; | |
4621 | int win; | |
4622 | int offs; | |
4623 | ||
4624 | /* we can't allocate as much as normalizer wants. | |
4625 | * so, found space must get proper lstart | |
4626 | * to cover original request */ | |
4627 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
4628 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
4629 | ||
4630 | /* we're limited by original request in that | |
4631 | * logical block must be covered any way | |
4632 | * winl is window we can move our chunk within */ | |
4633 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
4634 | ||
4635 | /* also, we should cover whole original request */ | |
53accfa9 | 4636 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
4637 | |
4638 | /* the smallest one defines real window */ | |
4639 | win = min(winl, wins); | |
4640 | ||
53accfa9 TT |
4641 | offs = ac->ac_o_ex.fe_logical % |
4642 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
4643 | if (offs && offs < win) |
4644 | win = offs; | |
4645 | ||
53accfa9 | 4646 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 4647 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
4648 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
4649 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
4650 | } | |
4651 | ||
4652 | /* preallocation can change ac_b_ex, thus we store actually | |
4653 | * allocated blocks for history */ | |
4654 | ac->ac_f_ex = ac->ac_b_ex; | |
4655 | ||
4656 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
4657 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4658 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4659 | pa->pa_free = pa->pa_len; | |
c9de560d | 4660 | spin_lock_init(&pa->pa_lock); |
d794bf8e AK |
4661 | INIT_LIST_HEAD(&pa->pa_inode_list); |
4662 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 4663 | pa->pa_deleted = 0; |
cc0fb9ad | 4664 | pa->pa_type = MB_INODE_PA; |
c9de560d | 4665 | |
d3df1453 RH |
4666 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4667 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4668 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
4669 | |
4670 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 4671 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
4672 | |
4673 | ei = EXT4_I(ac->ac_inode); | |
4674 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4675 | ||
4676 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
4677 | pa->pa_inode = ac->ac_inode; | |
4678 | ||
c9de560d | 4679 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d AT |
4680 | |
4681 | spin_lock(pa->pa_obj_lock); | |
4682 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
4683 | spin_unlock(pa->pa_obj_lock); | |
27bc446e | 4684 | atomic_inc(&ei->i_prealloc_active); |
c9de560d AT |
4685 | } |
4686 | ||
4687 | /* | |
4688 | * creates new preallocated space for locality group inodes belongs to | |
4689 | */ | |
53f86b17 | 4690 | static noinline_for_stack void |
4ddfef7b | 4691 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4692 | { |
4693 | struct super_block *sb = ac->ac_sb; | |
4694 | struct ext4_locality_group *lg; | |
4695 | struct ext4_prealloc_space *pa; | |
4696 | struct ext4_group_info *grp; | |
4697 | ||
4698 | /* preallocate only when found space is larger then requested */ | |
4699 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4700 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4701 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4702 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4703 | |
53f86b17 | 4704 | pa = ac->ac_pa; |
c9de560d AT |
4705 | |
4706 | /* preallocation can change ac_b_ex, thus we store actually | |
4707 | * allocated blocks for history */ | |
4708 | ac->ac_f_ex = ac->ac_b_ex; | |
4709 | ||
4710 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4711 | pa->pa_lstart = pa->pa_pstart; | |
4712 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4713 | pa->pa_free = pa->pa_len; | |
c9de560d | 4714 | spin_lock_init(&pa->pa_lock); |
6be2ded1 | 4715 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 4716 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 4717 | pa->pa_deleted = 0; |
cc0fb9ad | 4718 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 4719 | |
d3df1453 RH |
4720 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4721 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4722 | trace_ext4_mb_new_group_pa(ac, pa); |
c9de560d AT |
4723 | |
4724 | ext4_mb_use_group_pa(ac, pa); | |
4725 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
4726 | ||
4727 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4728 | lg = ac->ac_lg; | |
4729 | BUG_ON(lg == NULL); | |
4730 | ||
4731 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
4732 | pa->pa_inode = NULL; | |
4733 | ||
c9de560d | 4734 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 4735 | |
6be2ded1 AK |
4736 | /* |
4737 | * We will later add the new pa to the right bucket | |
4738 | * after updating the pa_free in ext4_mb_release_context | |
4739 | */ | |
c9de560d AT |
4740 | } |
4741 | ||
53f86b17 | 4742 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
c9de560d | 4743 | { |
c9de560d | 4744 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
53f86b17 | 4745 | ext4_mb_new_group_pa(ac); |
c9de560d | 4746 | else |
53f86b17 | 4747 | ext4_mb_new_inode_pa(ac); |
c9de560d AT |
4748 | } |
4749 | ||
4750 | /* | |
4751 | * finds all unused blocks in on-disk bitmap, frees them in | |
4752 | * in-core bitmap and buddy. | |
4753 | * @pa must be unlinked from inode and group lists, so that | |
4754 | * nobody else can find/use it. | |
4755 | * the caller MUST hold group/inode locks. | |
4756 | * TODO: optimize the case when there are no in-core structures yet | |
4757 | */ | |
4ddfef7b ES |
4758 | static noinline_for_stack int |
4759 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 4760 | struct ext4_prealloc_space *pa) |
c9de560d | 4761 | { |
c9de560d AT |
4762 | struct super_block *sb = e4b->bd_sb; |
4763 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
4764 | unsigned int end; |
4765 | unsigned int next; | |
c9de560d AT |
4766 | ext4_group_t group; |
4767 | ext4_grpblk_t bit; | |
ba80b101 | 4768 | unsigned long long grp_blk_start; |
c9de560d AT |
4769 | int free = 0; |
4770 | ||
4771 | BUG_ON(pa->pa_deleted == 0); | |
4772 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 4773 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
4774 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
4775 | end = bit + pa->pa_len; | |
4776 | ||
c9de560d | 4777 | while (bit < end) { |
ffad0a44 | 4778 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
4779 | if (bit >= end) |
4780 | break; | |
ffad0a44 | 4781 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
d3df1453 | 4782 | mb_debug(sb, "free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
4783 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
4784 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
4785 | free += next - bit; |
4786 | ||
3e1e5f50 | 4787 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
4788 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
4789 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 4790 | next - bit); |
c9de560d AT |
4791 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
4792 | bit = next + 1; | |
4793 | } | |
4794 | if (free != pa->pa_free) { | |
9d8b9ec4 | 4795 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 4796 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
4797 | pa, (unsigned long) pa->pa_lstart, |
4798 | (unsigned long) pa->pa_pstart, | |
36bad423 | 4799 | pa->pa_len); |
e29136f8 | 4800 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 4801 | free, pa->pa_free); |
e56eb659 AK |
4802 | /* |
4803 | * pa is already deleted so we use the value obtained | |
4804 | * from the bitmap and continue. | |
4805 | */ | |
c9de560d | 4806 | } |
c9de560d AT |
4807 | atomic_add(free, &sbi->s_mb_discarded); |
4808 | ||
863c37fc | 4809 | return 0; |
c9de560d AT |
4810 | } |
4811 | ||
4ddfef7b ES |
4812 | static noinline_for_stack int |
4813 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 4814 | struct ext4_prealloc_space *pa) |
c9de560d | 4815 | { |
c9de560d AT |
4816 | struct super_block *sb = e4b->bd_sb; |
4817 | ext4_group_t group; | |
4818 | ext4_grpblk_t bit; | |
4819 | ||
60e07cf5 | 4820 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
4821 | BUG_ON(pa->pa_deleted == 0); |
4822 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
4823 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
4824 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
4825 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 4826 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
4827 | |
4828 | return 0; | |
4829 | } | |
4830 | ||
4831 | /* | |
4832 | * releases all preallocations in given group | |
4833 | * | |
4834 | * first, we need to decide discard policy: | |
4835 | * - when do we discard | |
4836 | * 1) ENOSPC | |
4837 | * - how many do we discard | |
4838 | * 1) how many requested | |
4839 | */ | |
4ddfef7b ES |
4840 | static noinline_for_stack int |
4841 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
8c80fb31 | 4842 | ext4_group_t group, int *busy) |
c9de560d AT |
4843 | { |
4844 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4845 | struct buffer_head *bitmap_bh = NULL; | |
4846 | struct ext4_prealloc_space *pa, *tmp; | |
4847 | struct list_head list; | |
4848 | struct ext4_buddy e4b; | |
4849 | int err; | |
8c80fb31 | 4850 | int free = 0; |
c9de560d | 4851 | |
d3df1453 | 4852 | mb_debug(sb, "discard preallocation for group %u\n", group); |
c9de560d | 4853 | if (list_empty(&grp->bb_prealloc_list)) |
bbc4ec77 | 4854 | goto out_dbg; |
c9de560d | 4855 | |
574ca174 | 4856 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4857 | if (IS_ERR(bitmap_bh)) { |
4858 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4859 | ext4_error_err(sb, -err, |
4860 | "Error %d reading block bitmap for %u", | |
4861 | err, group); | |
bbc4ec77 | 4862 | goto out_dbg; |
c9de560d AT |
4863 | } |
4864 | ||
4865 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 4866 | if (err) { |
9651e6b2 KK |
4867 | ext4_warning(sb, "Error %d loading buddy information for %u", |
4868 | err, group); | |
ce89f46c | 4869 | put_bh(bitmap_bh); |
bbc4ec77 | 4870 | goto out_dbg; |
ce89f46c | 4871 | } |
c9de560d | 4872 | |
c9de560d | 4873 | INIT_LIST_HEAD(&list); |
c9de560d AT |
4874 | ext4_lock_group(sb, group); |
4875 | list_for_each_entry_safe(pa, tmp, | |
4876 | &grp->bb_prealloc_list, pa_group_list) { | |
4877 | spin_lock(&pa->pa_lock); | |
4878 | if (atomic_read(&pa->pa_count)) { | |
4879 | spin_unlock(&pa->pa_lock); | |
8c80fb31 | 4880 | *busy = 1; |
c9de560d AT |
4881 | continue; |
4882 | } | |
4883 | if (pa->pa_deleted) { | |
4884 | spin_unlock(&pa->pa_lock); | |
4885 | continue; | |
4886 | } | |
4887 | ||
4888 | /* seems this one can be freed ... */ | |
27bc446e | 4889 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 4890 | |
70022da8 YB |
4891 | if (!free) |
4892 | this_cpu_inc(discard_pa_seq); | |
4893 | ||
c9de560d AT |
4894 | /* we can trust pa_free ... */ |
4895 | free += pa->pa_free; | |
4896 | ||
4897 | spin_unlock(&pa->pa_lock); | |
4898 | ||
4899 | list_del(&pa->pa_group_list); | |
4900 | list_add(&pa->u.pa_tmp_list, &list); | |
4901 | } | |
4902 | ||
c9de560d AT |
4903 | /* now free all selected PAs */ |
4904 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
4905 | ||
4906 | /* remove from object (inode or locality group) */ | |
4907 | spin_lock(pa->pa_obj_lock); | |
4908 | list_del_rcu(&pa->pa_inode_list); | |
4909 | spin_unlock(pa->pa_obj_lock); | |
4910 | ||
cc0fb9ad | 4911 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 4912 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 4913 | else |
3e1e5f50 | 4914 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4915 | |
4916 | list_del(&pa->u.pa_tmp_list); | |
4917 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4918 | } | |
4919 | ||
c9de560d | 4920 | ext4_unlock_group(sb, group); |
e39e07fd | 4921 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4922 | put_bh(bitmap_bh); |
bbc4ec77 | 4923 | out_dbg: |
d3df1453 | 4924 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
8c80fb31 CX |
4925 | free, group, grp->bb_free); |
4926 | return free; | |
c9de560d AT |
4927 | } |
4928 | ||
4929 | /* | |
4930 | * releases all non-used preallocated blocks for given inode | |
4931 | * | |
4932 | * It's important to discard preallocations under i_data_sem | |
4933 | * We don't want another block to be served from the prealloc | |
4934 | * space when we are discarding the inode prealloc space. | |
4935 | * | |
4936 | * FIXME!! Make sure it is valid at all the call sites | |
4937 | */ | |
27bc446e | 4938 | void ext4_discard_preallocations(struct inode *inode, unsigned int needed) |
c9de560d AT |
4939 | { |
4940 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4941 | struct super_block *sb = inode->i_sb; | |
4942 | struct buffer_head *bitmap_bh = NULL; | |
4943 | struct ext4_prealloc_space *pa, *tmp; | |
4944 | ext4_group_t group = 0; | |
4945 | struct list_head list; | |
4946 | struct ext4_buddy e4b; | |
4947 | int err; | |
4948 | ||
c2ea3fde | 4949 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4950 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4951 | return; | |
4952 | } | |
4953 | ||
8016e29f HS |
4954 | if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) |
4955 | return; | |
4956 | ||
d3df1453 RH |
4957 | mb_debug(sb, "discard preallocation for inode %lu\n", |
4958 | inode->i_ino); | |
27bc446e | 4959 | trace_ext4_discard_preallocations(inode, |
4960 | atomic_read(&ei->i_prealloc_active), needed); | |
c9de560d AT |
4961 | |
4962 | INIT_LIST_HEAD(&list); | |
4963 | ||
27bc446e | 4964 | if (needed == 0) |
4965 | needed = UINT_MAX; | |
4966 | ||
c9de560d AT |
4967 | repeat: |
4968 | /* first, collect all pa's in the inode */ | |
4969 | spin_lock(&ei->i_prealloc_lock); | |
27bc446e | 4970 | while (!list_empty(&ei->i_prealloc_list) && needed) { |
4971 | pa = list_entry(ei->i_prealloc_list.prev, | |
c9de560d AT |
4972 | struct ext4_prealloc_space, pa_inode_list); |
4973 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4974 | spin_lock(&pa->pa_lock); | |
4975 | if (atomic_read(&pa->pa_count)) { | |
4976 | /* this shouldn't happen often - nobody should | |
4977 | * use preallocation while we're discarding it */ | |
4978 | spin_unlock(&pa->pa_lock); | |
4979 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
4980 | ext4_msg(sb, KERN_ERR, |
4981 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
4982 | WARN_ON(1); |
4983 | schedule_timeout_uninterruptible(HZ); | |
4984 | goto repeat; | |
4985 | ||
4986 | } | |
4987 | if (pa->pa_deleted == 0) { | |
27bc446e | 4988 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4989 | spin_unlock(&pa->pa_lock); |
4990 | list_del_rcu(&pa->pa_inode_list); | |
4991 | list_add(&pa->u.pa_tmp_list, &list); | |
27bc446e | 4992 | needed--; |
c9de560d AT |
4993 | continue; |
4994 | } | |
4995 | ||
4996 | /* someone is deleting pa right now */ | |
4997 | spin_unlock(&pa->pa_lock); | |
4998 | spin_unlock(&ei->i_prealloc_lock); | |
4999 | ||
5000 | /* we have to wait here because pa_deleted | |
5001 | * doesn't mean pa is already unlinked from | |
5002 | * the list. as we might be called from | |
5003 | * ->clear_inode() the inode will get freed | |
5004 | * and concurrent thread which is unlinking | |
5005 | * pa from inode's list may access already | |
5006 | * freed memory, bad-bad-bad */ | |
5007 | ||
5008 | /* XXX: if this happens too often, we can | |
5009 | * add a flag to force wait only in case | |
5010 | * of ->clear_inode(), but not in case of | |
5011 | * regular truncate */ | |
5012 | schedule_timeout_uninterruptible(HZ); | |
5013 | goto repeat; | |
5014 | } | |
5015 | spin_unlock(&ei->i_prealloc_lock); | |
5016 | ||
5017 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 5018 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 5019 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 5020 | |
9651e6b2 KK |
5021 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5022 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 5023 | if (err) { |
54d3adbc TT |
5024 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5025 | err, group); | |
ce89f46c AK |
5026 | continue; |
5027 | } | |
c9de560d | 5028 | |
574ca174 | 5029 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5030 | if (IS_ERR(bitmap_bh)) { |
5031 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5032 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
5033 | err, group); | |
e39e07fd | 5034 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 5035 | continue; |
c9de560d AT |
5036 | } |
5037 | ||
5038 | ext4_lock_group(sb, group); | |
5039 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5040 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
5041 | ext4_unlock_group(sb, group); |
5042 | ||
e39e07fd | 5043 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
5044 | put_bh(bitmap_bh); |
5045 | ||
5046 | list_del(&pa->u.pa_tmp_list); | |
5047 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5048 | } | |
5049 | } | |
5050 | ||
53f86b17 RH |
5051 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
5052 | { | |
5053 | struct ext4_prealloc_space *pa; | |
5054 | ||
5055 | BUG_ON(ext4_pspace_cachep == NULL); | |
5056 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); | |
5057 | if (!pa) | |
5058 | return -ENOMEM; | |
5059 | atomic_set(&pa->pa_count, 1); | |
5060 | ac->ac_pa = pa; | |
5061 | return 0; | |
5062 | } | |
5063 | ||
5064 | static void ext4_mb_pa_free(struct ext4_allocation_context *ac) | |
5065 | { | |
5066 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
5067 | ||
5068 | BUG_ON(!pa); | |
5069 | ac->ac_pa = NULL; | |
5070 | WARN_ON(!atomic_dec_and_test(&pa->pa_count)); | |
5071 | kmem_cache_free(ext4_pspace_cachep, pa); | |
5072 | } | |
5073 | ||
6ba495e9 | 5074 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 5075 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 5076 | { |
e68cf40c | 5077 | ext4_group_t i, ngroups; |
c9de560d | 5078 | |
9b5f6c9b | 5079 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e3570639 ES |
5080 | return; |
5081 | ||
8df9675f | 5082 | ngroups = ext4_get_groups_count(sb); |
d3df1453 | 5083 | mb_debug(sb, "groups: "); |
8df9675f | 5084 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
5085 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
5086 | struct ext4_prealloc_space *pa; | |
5087 | ext4_grpblk_t start; | |
5088 | struct list_head *cur; | |
5089 | ext4_lock_group(sb, i); | |
5090 | list_for_each(cur, &grp->bb_prealloc_list) { | |
5091 | pa = list_entry(cur, struct ext4_prealloc_space, | |
5092 | pa_group_list); | |
5093 | spin_lock(&pa->pa_lock); | |
5094 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
5095 | NULL, &start); | |
5096 | spin_unlock(&pa->pa_lock); | |
d3df1453 RH |
5097 | mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
5098 | pa->pa_len); | |
c9de560d | 5099 | } |
60bd63d1 | 5100 | ext4_unlock_group(sb, i); |
d3df1453 RH |
5101 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
5102 | grp->bb_fragments); | |
c9de560d | 5103 | } |
c9de560d | 5104 | } |
e68cf40c RH |
5105 | |
5106 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
5107 | { | |
5108 | struct super_block *sb = ac->ac_sb; | |
5109 | ||
9b5f6c9b | 5110 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e68cf40c RH |
5111 | return; |
5112 | ||
d3df1453 | 5113 | mb_debug(sb, "Can't allocate:" |
e68cf40c | 5114 | " Allocation context details:"); |
d3df1453 | 5115 | mb_debug(sb, "status %u flags 0x%x", |
e68cf40c | 5116 | ac->ac_status, ac->ac_flags); |
d3df1453 | 5117 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
e68cf40c RH |
5118 | "goal %lu/%lu/%lu@%lu, " |
5119 | "best %lu/%lu/%lu@%lu cr %d", | |
5120 | (unsigned long)ac->ac_o_ex.fe_group, | |
5121 | (unsigned long)ac->ac_o_ex.fe_start, | |
5122 | (unsigned long)ac->ac_o_ex.fe_len, | |
5123 | (unsigned long)ac->ac_o_ex.fe_logical, | |
5124 | (unsigned long)ac->ac_g_ex.fe_group, | |
5125 | (unsigned long)ac->ac_g_ex.fe_start, | |
5126 | (unsigned long)ac->ac_g_ex.fe_len, | |
5127 | (unsigned long)ac->ac_g_ex.fe_logical, | |
5128 | (unsigned long)ac->ac_b_ex.fe_group, | |
5129 | (unsigned long)ac->ac_b_ex.fe_start, | |
5130 | (unsigned long)ac->ac_b_ex.fe_len, | |
5131 | (unsigned long)ac->ac_b_ex.fe_logical, | |
5132 | (int)ac->ac_criteria); | |
d3df1453 | 5133 | mb_debug(sb, "%u found", ac->ac_found); |
e68cf40c RH |
5134 | ext4_mb_show_pa(sb); |
5135 | } | |
c9de560d | 5136 | #else |
e68cf40c RH |
5137 | static inline void ext4_mb_show_pa(struct super_block *sb) |
5138 | { | |
5139 | return; | |
5140 | } | |
c9de560d AT |
5141 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
5142 | { | |
e68cf40c | 5143 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
5144 | return; |
5145 | } | |
5146 | #endif | |
5147 | ||
5148 | /* | |
5149 | * We use locality group preallocation for small size file. The size of the | |
5150 | * file is determined by the current size or the resulting size after | |
5151 | * allocation which ever is larger | |
5152 | * | |
b713a5ec | 5153 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
5154 | */ |
5155 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
5156 | { | |
5157 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
5158 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
5159 | loff_t size, isize; | |
a9f2a293 | 5160 | bool inode_pa_eligible, group_pa_eligible; |
c9de560d AT |
5161 | |
5162 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
5163 | return; | |
5164 | ||
4ba74d00 TT |
5165 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
5166 | return; | |
5167 | ||
a9f2a293 JK |
5168 | group_pa_eligible = sbi->s_mb_group_prealloc > 0; |
5169 | inode_pa_eligible = true; | |
53accfa9 | 5170 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
5171 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
5172 | >> bsbits; | |
c9de560d | 5173 | |
a9f2a293 | 5174 | /* No point in using inode preallocation for closed files */ |
82dd124c | 5175 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
a9f2a293 JK |
5176 | !inode_is_open_for_write(ac->ac_inode)) |
5177 | inode_pa_eligible = false; | |
50797481 | 5178 | |
71780577 | 5179 | size = max(size, isize); |
a9f2a293 JK |
5180 | /* Don't use group allocation for large files */ |
5181 | if (size > sbi->s_mb_stream_request) | |
5182 | group_pa_eligible = false; | |
5183 | ||
5184 | if (!group_pa_eligible) { | |
5185 | if (inode_pa_eligible) | |
5186 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
5187 | else | |
5188 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
c9de560d | 5189 | return; |
4ba74d00 | 5190 | } |
c9de560d AT |
5191 | |
5192 | BUG_ON(ac->ac_lg != NULL); | |
5193 | /* | |
5194 | * locality group prealloc space are per cpu. The reason for having | |
5195 | * per cpu locality group is to reduce the contention between block | |
5196 | * request from multiple CPUs. | |
5197 | */ | |
a0b6bc63 | 5198 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
5199 | |
5200 | /* we're going to use group allocation */ | |
5201 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
5202 | ||
5203 | /* serialize all allocations in the group */ | |
5204 | mutex_lock(&ac->ac_lg->lg_mutex); | |
5205 | } | |
5206 | ||
4ddfef7b ES |
5207 | static noinline_for_stack int |
5208 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
5209 | struct ext4_allocation_request *ar) |
5210 | { | |
5211 | struct super_block *sb = ar->inode->i_sb; | |
5212 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
5213 | struct ext4_super_block *es = sbi->s_es; | |
5214 | ext4_group_t group; | |
498e5f24 TT |
5215 | unsigned int len; |
5216 | ext4_fsblk_t goal; | |
c9de560d AT |
5217 | ext4_grpblk_t block; |
5218 | ||
5219 | /* we can't allocate > group size */ | |
5220 | len = ar->len; | |
5221 | ||
5222 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
5223 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
5224 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
5225 | |
5226 | /* start searching from the goal */ | |
5227 | goal = ar->goal; | |
5228 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5229 | goal >= ext4_blocks_count(es)) | |
5230 | goal = le32_to_cpu(es->s_first_data_block); | |
5231 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
5232 | ||
5233 | /* set up allocation goals */ | |
f5a44db5 | 5234 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 5235 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
5236 | ac->ac_sb = sb; |
5237 | ac->ac_inode = ar->inode; | |
53accfa9 | 5238 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
5239 | ac->ac_o_ex.fe_group = group; |
5240 | ac->ac_o_ex.fe_start = block; | |
5241 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 5242 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 5243 | ac->ac_flags = ar->flags; |
c9de560d | 5244 | |
3cb77bd2 | 5245 | /* we have to define context: we'll work with a file or |
c9de560d AT |
5246 | * locality group. this is a policy, actually */ |
5247 | ext4_mb_group_or_file(ac); | |
5248 | ||
d3df1453 | 5249 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
c9de560d AT |
5250 | "left: %u/%u, right %u/%u to %swritable\n", |
5251 | (unsigned) ar->len, (unsigned) ar->logical, | |
5252 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
5253 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
5254 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 5255 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
5256 | return 0; |
5257 | ||
5258 | } | |
5259 | ||
6be2ded1 AK |
5260 | static noinline_for_stack void |
5261 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
5262 | struct ext4_locality_group *lg, | |
5263 | int order, int total_entries) | |
5264 | { | |
5265 | ext4_group_t group = 0; | |
5266 | struct ext4_buddy e4b; | |
5267 | struct list_head discard_list; | |
5268 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 5269 | |
d3df1453 | 5270 | mb_debug(sb, "discard locality group preallocation\n"); |
6be2ded1 AK |
5271 | |
5272 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
5273 | |
5274 | spin_lock(&lg->lg_prealloc_lock); | |
5275 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
92e9c58c MB |
5276 | pa_inode_list, |
5277 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
5278 | spin_lock(&pa->pa_lock); |
5279 | if (atomic_read(&pa->pa_count)) { | |
5280 | /* | |
5281 | * This is the pa that we just used | |
5282 | * for block allocation. So don't | |
5283 | * free that | |
5284 | */ | |
5285 | spin_unlock(&pa->pa_lock); | |
5286 | continue; | |
5287 | } | |
5288 | if (pa->pa_deleted) { | |
5289 | spin_unlock(&pa->pa_lock); | |
5290 | continue; | |
5291 | } | |
5292 | /* only lg prealloc space */ | |
cc0fb9ad | 5293 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
5294 | |
5295 | /* seems this one can be freed ... */ | |
27bc446e | 5296 | ext4_mb_mark_pa_deleted(sb, pa); |
6be2ded1 AK |
5297 | spin_unlock(&pa->pa_lock); |
5298 | ||
5299 | list_del_rcu(&pa->pa_inode_list); | |
5300 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
5301 | ||
5302 | total_entries--; | |
5303 | if (total_entries <= 5) { | |
5304 | /* | |
5305 | * we want to keep only 5 entries | |
5306 | * allowing it to grow to 8. This | |
5307 | * mak sure we don't call discard | |
5308 | * soon for this list. | |
5309 | */ | |
5310 | break; | |
5311 | } | |
5312 | } | |
5313 | spin_unlock(&lg->lg_prealloc_lock); | |
5314 | ||
5315 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 5316 | int err; |
6be2ded1 | 5317 | |
bd86298e | 5318 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
5319 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5320 | GFP_NOFS|__GFP_NOFAIL); | |
5321 | if (err) { | |
54d3adbc TT |
5322 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5323 | err, group); | |
6be2ded1 AK |
5324 | continue; |
5325 | } | |
5326 | ext4_lock_group(sb, group); | |
5327 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5328 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
5329 | ext4_unlock_group(sb, group); |
5330 | ||
e39e07fd | 5331 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
5332 | list_del(&pa->u.pa_tmp_list); |
5333 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5334 | } | |
6be2ded1 AK |
5335 | } |
5336 | ||
5337 | /* | |
5338 | * We have incremented pa_count. So it cannot be freed at this | |
5339 | * point. Also we hold lg_mutex. So no parallel allocation is | |
5340 | * possible from this lg. That means pa_free cannot be updated. | |
5341 | * | |
5342 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
5343 | * which can cause the lg_prealloc_list to be updated. | |
5344 | */ | |
5345 | ||
5346 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
5347 | { | |
5348 | int order, added = 0, lg_prealloc_count = 1; | |
5349 | struct super_block *sb = ac->ac_sb; | |
5350 | struct ext4_locality_group *lg = ac->ac_lg; | |
5351 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
5352 | ||
5353 | order = fls(pa->pa_free) - 1; | |
5354 | if (order > PREALLOC_TB_SIZE - 1) | |
5355 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
5356 | order = PREALLOC_TB_SIZE - 1; | |
5357 | /* Add the prealloc space to lg */ | |
f1167009 | 5358 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 5359 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
92e9c58c MB |
5360 | pa_inode_list, |
5361 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
5362 | spin_lock(&tmp_pa->pa_lock); |
5363 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 5364 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
5365 | continue; |
5366 | } | |
5367 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
5368 | /* Add to the tail of the previous entry */ | |
5369 | list_add_tail_rcu(&pa->pa_inode_list, | |
5370 | &tmp_pa->pa_inode_list); | |
5371 | added = 1; | |
5372 | /* | |
5373 | * we want to count the total | |
5374 | * number of entries in the list | |
5375 | */ | |
5376 | } | |
5377 | spin_unlock(&tmp_pa->pa_lock); | |
5378 | lg_prealloc_count++; | |
5379 | } | |
5380 | if (!added) | |
5381 | list_add_tail_rcu(&pa->pa_inode_list, | |
5382 | &lg->lg_prealloc_list[order]); | |
f1167009 | 5383 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
5384 | |
5385 | /* Now trim the list to be not more than 8 elements */ | |
5386 | if (lg_prealloc_count > 8) { | |
5387 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 5388 | order, lg_prealloc_count); |
6be2ded1 AK |
5389 | return; |
5390 | } | |
5391 | return ; | |
5392 | } | |
5393 | ||
27bc446e | 5394 | /* |
5395 | * if per-inode prealloc list is too long, trim some PA | |
5396 | */ | |
5397 | static void ext4_mb_trim_inode_pa(struct inode *inode) | |
5398 | { | |
5399 | struct ext4_inode_info *ei = EXT4_I(inode); | |
5400 | struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); | |
5401 | int count, delta; | |
5402 | ||
5403 | count = atomic_read(&ei->i_prealloc_active); | |
5404 | delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1; | |
5405 | if (count > sbi->s_mb_max_inode_prealloc + delta) { | |
5406 | count -= sbi->s_mb_max_inode_prealloc; | |
5407 | ext4_discard_preallocations(inode, count); | |
5408 | } | |
5409 | } | |
5410 | ||
c9de560d AT |
5411 | /* |
5412 | * release all resource we used in allocation | |
5413 | */ | |
5414 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
5415 | { | |
27bc446e | 5416 | struct inode *inode = ac->ac_inode; |
5417 | struct ext4_inode_info *ei = EXT4_I(inode); | |
53accfa9 | 5418 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
5419 | struct ext4_prealloc_space *pa = ac->ac_pa; |
5420 | if (pa) { | |
cc0fb9ad | 5421 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 5422 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 5423 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
5424 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
5425 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
5426 | pa->pa_free -= ac->ac_b_ex.fe_len; |
5427 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
5428 | spin_unlock(&pa->pa_lock); | |
66d5e027 | 5429 | |
5430 | /* | |
5431 | * We want to add the pa to the right bucket. | |
5432 | * Remove it from the list and while adding | |
5433 | * make sure the list to which we are adding | |
5434 | * doesn't grow big. | |
5435 | */ | |
5436 | if (likely(pa->pa_free)) { | |
5437 | spin_lock(pa->pa_obj_lock); | |
5438 | list_del_rcu(&pa->pa_inode_list); | |
5439 | spin_unlock(pa->pa_obj_lock); | |
5440 | ext4_mb_add_n_trim(ac); | |
5441 | } | |
ba443916 | 5442 | } |
27bc446e | 5443 | |
5444 | if (pa->pa_type == MB_INODE_PA) { | |
5445 | /* | |
5446 | * treat per-inode prealloc list as a lru list, then try | |
5447 | * to trim the least recently used PA. | |
5448 | */ | |
5449 | spin_lock(pa->pa_obj_lock); | |
5450 | list_move(&pa->pa_inode_list, &ei->i_prealloc_list); | |
5451 | spin_unlock(pa->pa_obj_lock); | |
5452 | } | |
5453 | ||
ba443916 AK |
5454 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
5455 | } | |
c9de560d | 5456 | if (ac->ac_bitmap_page) |
09cbfeaf | 5457 | put_page(ac->ac_bitmap_page); |
c9de560d | 5458 | if (ac->ac_buddy_page) |
09cbfeaf | 5459 | put_page(ac->ac_buddy_page); |
c9de560d AT |
5460 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
5461 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
5462 | ext4_mb_collect_stats(ac); | |
27bc446e | 5463 | ext4_mb_trim_inode_pa(inode); |
c9de560d AT |
5464 | return 0; |
5465 | } | |
5466 | ||
5467 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
5468 | { | |
8df9675f | 5469 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d | 5470 | int ret; |
8c80fb31 CX |
5471 | int freed = 0, busy = 0; |
5472 | int retry = 0; | |
c9de560d | 5473 | |
9bffad1e | 5474 | trace_ext4_mb_discard_preallocations(sb, needed); |
8c80fb31 CX |
5475 | |
5476 | if (needed == 0) | |
5477 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; | |
5478 | repeat: | |
8df9675f | 5479 | for (i = 0; i < ngroups && needed > 0; i++) { |
8c80fb31 | 5480 | ret = ext4_mb_discard_group_preallocations(sb, i, &busy); |
c9de560d AT |
5481 | freed += ret; |
5482 | needed -= ret; | |
8c80fb31 CX |
5483 | cond_resched(); |
5484 | } | |
5485 | ||
5486 | if (needed > 0 && busy && ++retry < 3) { | |
5487 | busy = 0; | |
5488 | goto repeat; | |
c9de560d AT |
5489 | } |
5490 | ||
5491 | return freed; | |
5492 | } | |
5493 | ||
cf5e2ca6 | 5494 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
07b5b8e1 | 5495 | struct ext4_allocation_context *ac, u64 *seq) |
cf5e2ca6 RH |
5496 | { |
5497 | int freed; | |
07b5b8e1 RH |
5498 | u64 seq_retry = 0; |
5499 | bool ret = false; | |
cf5e2ca6 RH |
5500 | |
5501 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
07b5b8e1 RH |
5502 | if (freed) { |
5503 | ret = true; | |
5504 | goto out_dbg; | |
5505 | } | |
5506 | seq_retry = ext4_get_discard_pa_seq_sum(); | |
99377830 RH |
5507 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
5508 | ac->ac_flags |= EXT4_MB_STRICT_CHECK; | |
07b5b8e1 RH |
5509 | *seq = seq_retry; |
5510 | ret = true; | |
5511 | } | |
5512 | ||
5513 | out_dbg: | |
5514 | mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); | |
5515 | return ret; | |
cf5e2ca6 RH |
5516 | } |
5517 | ||
8016e29f HS |
5518 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, |
5519 | struct ext4_allocation_request *ar, int *errp); | |
5520 | ||
c9de560d AT |
5521 | /* |
5522 | * Main entry point into mballoc to allocate blocks | |
5523 | * it tries to use preallocation first, then falls back | |
5524 | * to usual allocation | |
5525 | */ | |
5526 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 5527 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 5528 | { |
256bdb49 | 5529 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
5530 | struct ext4_sb_info *sbi; |
5531 | struct super_block *sb; | |
5532 | ext4_fsblk_t block = 0; | |
60e58e0f | 5533 | unsigned int inquota = 0; |
53accfa9 | 5534 | unsigned int reserv_clstrs = 0; |
80fa46d6 | 5535 | int retries = 0; |
07b5b8e1 | 5536 | u64 seq; |
c9de560d | 5537 | |
b10a44c3 | 5538 | might_sleep(); |
c9de560d AT |
5539 | sb = ar->inode->i_sb; |
5540 | sbi = EXT4_SB(sb); | |
5541 | ||
9bffad1e | 5542 | trace_ext4_request_blocks(ar); |
8016e29f HS |
5543 | if (sbi->s_mount_state & EXT4_FC_REPLAY) |
5544 | return ext4_mb_new_blocks_simple(handle, ar, errp); | |
ba80b101 | 5545 | |
45dc63e7 | 5546 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 5547 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
5548 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
5549 | ||
e3cf5d5d | 5550 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
5551 | /* Without delayed allocation we need to verify |
5552 | * there is enough free blocks to do block allocation | |
5553 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 5554 | */ |
55f020db | 5555 | while (ar->len && |
e7d5f315 | 5556 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 5557 | |
030ba6bc | 5558 | /* let others to free the space */ |
bb8b20ed | 5559 | cond_resched(); |
030ba6bc AK |
5560 | ar->len = ar->len >> 1; |
5561 | } | |
5562 | if (!ar->len) { | |
bbc4ec77 | 5563 | ext4_mb_show_pa(sb); |
a30d542a AK |
5564 | *errp = -ENOSPC; |
5565 | return 0; | |
5566 | } | |
53accfa9 | 5567 | reserv_clstrs = ar->len; |
55f020db | 5568 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
5569 | dquot_alloc_block_nofail(ar->inode, |
5570 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
5571 | } else { |
5572 | while (ar->len && | |
53accfa9 TT |
5573 | dquot_alloc_block(ar->inode, |
5574 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
5575 | |
5576 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
5577 | ar->len--; | |
5578 | } | |
60e58e0f MC |
5579 | } |
5580 | inquota = ar->len; | |
5581 | if (ar->len == 0) { | |
5582 | *errp = -EDQUOT; | |
6c7a120a | 5583 | goto out; |
60e58e0f | 5584 | } |
07031431 | 5585 | } |
d2a17637 | 5586 | |
85556c9a | 5587 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 5588 | if (!ac) { |
363d4251 | 5589 | ar->len = 0; |
256bdb49 | 5590 | *errp = -ENOMEM; |
6c7a120a | 5591 | goto out; |
256bdb49 ES |
5592 | } |
5593 | ||
256bdb49 | 5594 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
5595 | if (*errp) { |
5596 | ar->len = 0; | |
6c7a120a | 5597 | goto out; |
c9de560d AT |
5598 | } |
5599 | ||
256bdb49 | 5600 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
81198536 | 5601 | seq = this_cpu_read(discard_pa_seq); |
256bdb49 | 5602 | if (!ext4_mb_use_preallocated(ac)) { |
256bdb49 ES |
5603 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
5604 | ext4_mb_normalize_request(ac, ar); | |
53f86b17 RH |
5605 | |
5606 | *errp = ext4_mb_pa_alloc(ac); | |
5607 | if (*errp) | |
5608 | goto errout; | |
c9de560d AT |
5609 | repeat: |
5610 | /* allocate space in core */ | |
6c7a120a | 5611 | *errp = ext4_mb_regular_allocator(ac); |
53f86b17 RH |
5612 | /* |
5613 | * pa allocated above is added to grp->bb_prealloc_list only | |
5614 | * when we were able to allocate some block i.e. when | |
5615 | * ac->ac_status == AC_STATUS_FOUND. | |
5616 | * And error from above mean ac->ac_status != AC_STATUS_FOUND | |
5617 | * So we have to free this pa here itself. | |
5618 | */ | |
2c00ef3e | 5619 | if (*errp) { |
53f86b17 | 5620 | ext4_mb_pa_free(ac); |
2c00ef3e AK |
5621 | ext4_discard_allocated_blocks(ac); |
5622 | goto errout; | |
5623 | } | |
53f86b17 RH |
5624 | if (ac->ac_status == AC_STATUS_FOUND && |
5625 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) | |
5626 | ext4_mb_pa_free(ac); | |
c9de560d | 5627 | } |
256bdb49 | 5628 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 5629 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 5630 | if (*errp) { |
b844167e | 5631 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
5632 | goto errout; |
5633 | } else { | |
519deca0 AK |
5634 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
5635 | ar->len = ac->ac_b_ex.fe_len; | |
5636 | } | |
c9de560d | 5637 | } else { |
80fa46d6 TT |
5638 | if (++retries < 3 && |
5639 | ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) | |
c9de560d | 5640 | goto repeat; |
53f86b17 RH |
5641 | /* |
5642 | * If block allocation fails then the pa allocated above | |
5643 | * needs to be freed here itself. | |
5644 | */ | |
5645 | ext4_mb_pa_free(ac); | |
c9de560d | 5646 | *errp = -ENOSPC; |
6c7a120a AK |
5647 | } |
5648 | ||
6d138ced | 5649 | errout: |
6c7a120a | 5650 | if (*errp) { |
256bdb49 | 5651 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 5652 | ar->len = 0; |
256bdb49 | 5653 | ext4_mb_show_ac(ac); |
c9de560d | 5654 | } |
256bdb49 | 5655 | ext4_mb_release_context(ac); |
6c7a120a AK |
5656 | out: |
5657 | if (ac) | |
5658 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 5659 | if (inquota && ar->len < inquota) |
53accfa9 | 5660 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 5661 | if (!ar->len) { |
e3cf5d5d | 5662 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 5663 | /* release all the reserved blocks if non delalloc */ |
57042651 | 5664 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 5665 | reserv_clstrs); |
0087d9fb | 5666 | } |
c9de560d | 5667 | |
9bffad1e | 5668 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 5669 | |
c9de560d AT |
5670 | return block; |
5671 | } | |
c9de560d | 5672 | |
c894058d AK |
5673 | /* |
5674 | * We can merge two free data extents only if the physical blocks | |
5675 | * are contiguous, AND the extents were freed by the same transaction, | |
5676 | * AND the blocks are associated with the same group. | |
5677 | */ | |
a0154344 DJ |
5678 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
5679 | struct ext4_free_data *entry, | |
5680 | struct ext4_free_data *new_entry, | |
5681 | struct rb_root *entry_rb_root) | |
c894058d | 5682 | { |
a0154344 DJ |
5683 | if ((entry->efd_tid != new_entry->efd_tid) || |
5684 | (entry->efd_group != new_entry->efd_group)) | |
5685 | return; | |
5686 | if (entry->efd_start_cluster + entry->efd_count == | |
5687 | new_entry->efd_start_cluster) { | |
5688 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
5689 | new_entry->efd_count += entry->efd_count; | |
5690 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
5691 | entry->efd_start_cluster) { | |
5692 | new_entry->efd_count += entry->efd_count; | |
5693 | } else | |
5694 | return; | |
5695 | spin_lock(&sbi->s_md_lock); | |
5696 | list_del(&entry->efd_list); | |
5697 | spin_unlock(&sbi->s_md_lock); | |
5698 | rb_erase(&entry->efd_node, entry_rb_root); | |
5699 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
5700 | } |
5701 | ||
4ddfef7b ES |
5702 | static noinline_for_stack int |
5703 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 5704 | struct ext4_free_data *new_entry) |
c9de560d | 5705 | { |
e29136f8 | 5706 | ext4_group_t group = e4b->bd_group; |
84130193 | 5707 | ext4_grpblk_t cluster; |
d08854f5 | 5708 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 5709 | struct ext4_free_data *entry; |
c9de560d AT |
5710 | struct ext4_group_info *db = e4b->bd_info; |
5711 | struct super_block *sb = e4b->bd_sb; | |
5712 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
5713 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
5714 | struct rb_node *parent = NULL, *new_node; | |
5715 | ||
0390131b | 5716 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
5717 | BUG_ON(e4b->bd_bitmap_page == NULL); |
5718 | BUG_ON(e4b->bd_buddy_page == NULL); | |
5719 | ||
18aadd47 BJ |
5720 | new_node = &new_entry->efd_node; |
5721 | cluster = new_entry->efd_start_cluster; | |
c894058d | 5722 | |
c894058d AK |
5723 | if (!*n) { |
5724 | /* first free block exent. We need to | |
5725 | protect buddy cache from being freed, | |
5726 | * otherwise we'll refresh it from | |
5727 | * on-disk bitmap and lose not-yet-available | |
5728 | * blocks */ | |
09cbfeaf KS |
5729 | get_page(e4b->bd_buddy_page); |
5730 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
5731 | } |
5732 | while (*n) { | |
5733 | parent = *n; | |
18aadd47 BJ |
5734 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
5735 | if (cluster < entry->efd_start_cluster) | |
c894058d | 5736 | n = &(*n)->rb_left; |
18aadd47 | 5737 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
5738 | n = &(*n)->rb_right; |
5739 | else { | |
e29136f8 | 5740 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
5741 | ext4_group_first_block_no(sb, group) + |
5742 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 5743 | "Block already on to-be-freed list"); |
cca41553 | 5744 | kmem_cache_free(ext4_free_data_cachep, new_entry); |
c894058d | 5745 | return 0; |
c9de560d | 5746 | } |
c894058d | 5747 | } |
c9de560d | 5748 | |
c894058d AK |
5749 | rb_link_node(new_node, parent, n); |
5750 | rb_insert_color(new_node, &db->bb_free_root); | |
5751 | ||
5752 | /* Now try to see the extent can be merged to left and right */ | |
5753 | node = rb_prev(new_node); | |
5754 | if (node) { | |
18aadd47 | 5755 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5756 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5757 | &(db->bb_free_root)); | |
c894058d | 5758 | } |
c9de560d | 5759 | |
c894058d AK |
5760 | node = rb_next(new_node); |
5761 | if (node) { | |
18aadd47 | 5762 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5763 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5764 | &(db->bb_free_root)); | |
c9de560d | 5765 | } |
a0154344 | 5766 | |
d08854f5 | 5767 | spin_lock(&sbi->s_md_lock); |
a0154344 | 5768 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
5769 | sbi->s_mb_free_pending += clusters; |
5770 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
5771 | return 0; |
5772 | } | |
5773 | ||
8016e29f HS |
5774 | /* |
5775 | * Simple allocator for Ext4 fast commit replay path. It searches for blocks | |
5776 | * linearly starting at the goal block and also excludes the blocks which | |
5777 | * are going to be in use after fast commit replay. | |
5778 | */ | |
5779 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, | |
5780 | struct ext4_allocation_request *ar, int *errp) | |
5781 | { | |
5782 | struct buffer_head *bitmap_bh; | |
5783 | struct super_block *sb = ar->inode->i_sb; | |
5784 | ext4_group_t group; | |
5785 | ext4_grpblk_t blkoff; | |
31a074a0 XY |
5786 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
5787 | ext4_grpblk_t i = 0; | |
8016e29f HS |
5788 | ext4_fsblk_t goal, block; |
5789 | struct ext4_super_block *es = EXT4_SB(sb)->s_es; | |
5790 | ||
5791 | goal = ar->goal; | |
5792 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5793 | goal >= ext4_blocks_count(es)) | |
5794 | goal = le32_to_cpu(es->s_first_data_block); | |
5795 | ||
5796 | ar->len = 0; | |
5797 | ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); | |
5798 | for (; group < ext4_get_groups_count(sb); group++) { | |
5799 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
5800 | if (IS_ERR(bitmap_bh)) { | |
5801 | *errp = PTR_ERR(bitmap_bh); | |
5802 | pr_warn("Failed to read block bitmap\n"); | |
5803 | return 0; | |
5804 | } | |
5805 | ||
5806 | ext4_get_group_no_and_offset(sb, | |
5807 | max(ext4_group_first_block_no(sb, group), goal), | |
5808 | NULL, &blkoff); | |
31a074a0 XY |
5809 | while (1) { |
5810 | i = mb_find_next_zero_bit(bitmap_bh->b_data, max, | |
8016e29f | 5811 | blkoff); |
31a074a0 XY |
5812 | if (i >= max) |
5813 | break; | |
5814 | if (ext4_fc_replay_check_excluded(sb, | |
5815 | ext4_group_first_block_no(sb, group) + i)) { | |
5816 | blkoff = i + 1; | |
5817 | } else | |
5818 | break; | |
5819 | } | |
8016e29f | 5820 | brelse(bitmap_bh); |
31a074a0 XY |
5821 | if (i < max) |
5822 | break; | |
8016e29f HS |
5823 | } |
5824 | ||
31a074a0 XY |
5825 | if (group >= ext4_get_groups_count(sb) || i >= max) { |
5826 | *errp = -ENOSPC; | |
8016e29f | 5827 | return 0; |
31a074a0 | 5828 | } |
8016e29f HS |
5829 | |
5830 | block = ext4_group_first_block_no(sb, group) + i; | |
5831 | ext4_mb_mark_bb(sb, block, 1, 1); | |
5832 | ar->len = 1; | |
5833 | ||
5834 | return block; | |
5835 | } | |
5836 | ||
5837 | static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, | |
5838 | unsigned long count) | |
5839 | { | |
5840 | struct buffer_head *bitmap_bh; | |
5841 | struct super_block *sb = inode->i_sb; | |
5842 | struct ext4_group_desc *gdp; | |
5843 | struct buffer_head *gdp_bh; | |
5844 | ext4_group_t group; | |
5845 | ext4_grpblk_t blkoff; | |
5846 | int already_freed = 0, err, i; | |
5847 | ||
5848 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
5849 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
5850 | if (IS_ERR(bitmap_bh)) { | |
5851 | err = PTR_ERR(bitmap_bh); | |
5852 | pr_warn("Failed to read block bitmap\n"); | |
5853 | return; | |
5854 | } | |
5855 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
5856 | if (!gdp) | |
5857 | return; | |
5858 | ||
5859 | for (i = 0; i < count; i++) { | |
5860 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data)) | |
5861 | already_freed++; | |
5862 | } | |
5863 | mb_clear_bits(bitmap_bh->b_data, blkoff, count); | |
5864 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
5865 | if (err) | |
5866 | return; | |
5867 | ext4_free_group_clusters_set( | |
5868 | sb, gdp, ext4_free_group_clusters(sb, gdp) + | |
5869 | count - already_freed); | |
5870 | ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh); | |
5871 | ext4_group_desc_csum_set(sb, group, gdp); | |
5872 | ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
5873 | sync_dirty_buffer(bitmap_bh); | |
5874 | sync_dirty_buffer(gdp_bh); | |
5875 | brelse(bitmap_bh); | |
5876 | } | |
5877 | ||
44338711 | 5878 | /** |
8ac3939d RH |
5879 | * ext4_mb_clear_bb() -- helper function for freeing blocks. |
5880 | * Used by ext4_free_blocks() | |
44338711 TT |
5881 | * @handle: handle for this transaction |
5882 | * @inode: inode | |
c60990b3 TT |
5883 | * @block: starting physical block to be freed |
5884 | * @count: number of blocks to be freed | |
5def1360 | 5885 | * @flags: flags used by ext4_free_blocks |
c9de560d | 5886 | */ |
8ac3939d RH |
5887 | static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, |
5888 | ext4_fsblk_t block, unsigned long count, | |
5889 | int flags) | |
c9de560d | 5890 | { |
26346ff6 | 5891 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 5892 | struct super_block *sb = inode->i_sb; |
c9de560d | 5893 | struct ext4_group_desc *gdp; |
498e5f24 | 5894 | unsigned int overflow; |
c9de560d AT |
5895 | ext4_grpblk_t bit; |
5896 | struct buffer_head *gd_bh; | |
5897 | ext4_group_t block_group; | |
5898 | struct ext4_sb_info *sbi; | |
5899 | struct ext4_buddy e4b; | |
84130193 | 5900 | unsigned int count_clusters; |
c9de560d AT |
5901 | int err = 0; |
5902 | int ret; | |
5903 | ||
8016e29f HS |
5904 | sbi = EXT4_SB(sb); |
5905 | ||
1e1c2b86 LC |
5906 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
5907 | !ext4_inode_block_valid(inode, block, count)) { | |
5908 | ext4_error(sb, "Freeing blocks in system zone - " | |
5909 | "Block = %llu, count = %lu", block, count); | |
5910 | /* err = 0. ext4_std_error should be a no op */ | |
5911 | goto error_return; | |
5912 | } | |
5913 | flags |= EXT4_FREE_BLOCKS_VALIDATED; | |
5914 | ||
c9de560d AT |
5915 | do_more: |
5916 | overflow = 0; | |
5917 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
5918 | ||
163a203d DW |
5919 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
5920 | ext4_get_group_info(sb, block_group)))) | |
5921 | return; | |
5922 | ||
c9de560d AT |
5923 | /* |
5924 | * Check to see if we are freeing blocks across a group | |
5925 | * boundary. | |
5926 | */ | |
84130193 TT |
5927 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
5928 | overflow = EXT4_C2B(sbi, bit) + count - | |
5929 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d | 5930 | count -= overflow; |
1e1c2b86 LC |
5931 | /* The range changed so it's no longer validated */ |
5932 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d | 5933 | } |
810da240 | 5934 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 5935 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5936 | if (IS_ERR(bitmap_bh)) { |
5937 | err = PTR_ERR(bitmap_bh); | |
5938 | bitmap_bh = NULL; | |
c9de560d | 5939 | goto error_return; |
ce89f46c | 5940 | } |
c9de560d | 5941 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
5942 | if (!gdp) { |
5943 | err = -EIO; | |
c9de560d | 5944 | goto error_return; |
ce89f46c | 5945 | } |
c9de560d | 5946 | |
1e1c2b86 LC |
5947 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
5948 | !ext4_inode_block_valid(inode, block, count)) { | |
12062ddd | 5949 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 5950 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
5951 | /* err = 0. ext4_std_error should be a no op */ |
5952 | goto error_return; | |
c9de560d AT |
5953 | } |
5954 | ||
5955 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
188c299e JK |
5956 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
5957 | EXT4_JTR_NONE); | |
c9de560d AT |
5958 | if (err) |
5959 | goto error_return; | |
5960 | ||
5961 | /* | |
5962 | * We are about to modify some metadata. Call the journal APIs | |
5963 | * to unshare ->b_data if a currently-committing transaction is | |
5964 | * using it | |
5965 | */ | |
5966 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 5967 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
c9de560d AT |
5968 | if (err) |
5969 | goto error_return; | |
c9de560d AT |
5970 | #ifdef AGGRESSIVE_CHECK |
5971 | { | |
5972 | int i; | |
84130193 | 5973 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
5974 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
5975 | } | |
5976 | #endif | |
84130193 | 5977 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 5978 | |
adb7ef60 KK |
5979 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
5980 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
5981 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
5982 | if (err) |
5983 | goto error_return; | |
e6362609 | 5984 | |
f96c450d DJ |
5985 | /* |
5986 | * We need to make sure we don't reuse the freed block until after the | |
5987 | * transaction is committed. We make an exception if the inode is to be | |
5988 | * written in writeback mode since writeback mode has weak data | |
5989 | * consistency guarantees. | |
5990 | */ | |
5991 | if (ext4_handle_valid(handle) && | |
5992 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
5993 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
5994 | struct ext4_free_data *new_entry; |
5995 | /* | |
7444a072 MH |
5996 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
5997 | * to fail. | |
7a2fcbf7 | 5998 | */ |
7444a072 MH |
5999 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
6000 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
6001 | new_entry->efd_start_cluster = bit; |
6002 | new_entry->efd_group = block_group; | |
6003 | new_entry->efd_count = count_clusters; | |
6004 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 6005 | |
7a2fcbf7 | 6006 | ext4_lock_group(sb, block_group); |
84130193 | 6007 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 6008 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 6009 | } else { |
7a2fcbf7 AK |
6010 | /* need to update group_info->bb_free and bitmap |
6011 | * with group lock held. generate_buddy look at | |
6012 | * them with group lock_held | |
6013 | */ | |
d71c1ae2 | 6014 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
6015 | err = ext4_issue_discard(sb, block_group, bit, count, |
6016 | NULL); | |
d71c1ae2 LC |
6017 | if (err && err != -EOPNOTSUPP) |
6018 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
a00b482b | 6019 | " group:%u block:%d count:%lu failed" |
d71c1ae2 LC |
6020 | " with %d", block_group, bit, count, |
6021 | err); | |
8f9ff189 LC |
6022 | } else |
6023 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 6024 | |
955ce5f5 | 6025 | ext4_lock_group(sb, block_group); |
84130193 TT |
6026 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
6027 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
6028 | } |
6029 | ||
021b65bb TT |
6030 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
6031 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 6032 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 6033 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 6034 | ext4_unlock_group(sb, block_group); |
c9de560d | 6035 | |
772cb7c8 JS |
6036 | if (sbi->s_log_groups_per_flex) { |
6037 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 6038 | atomic64_add(count_clusters, |
7c990728 SJS |
6039 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6040 | flex_group)->free_clusters); | |
772cb7c8 JS |
6041 | } |
6042 | ||
9fe67149 EW |
6043 | /* |
6044 | * on a bigalloc file system, defer the s_freeclusters_counter | |
6045 | * update to the caller (ext4_remove_space and friends) so they | |
6046 | * can determine if a cluster freed here should be rereserved | |
6047 | */ | |
6048 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
6049 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
6050 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
6051 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
6052 | count_clusters); | |
6053 | } | |
7d734532 JK |
6054 | |
6055 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 6056 | |
7a2fcbf7 AK |
6057 | /* We dirtied the bitmap block */ |
6058 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6059 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6060 | ||
c9de560d AT |
6061 | /* And the group descriptor block */ |
6062 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 6063 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
6064 | if (!err) |
6065 | err = ret; | |
6066 | ||
6067 | if (overflow && !err) { | |
6068 | block += count; | |
6069 | count = overflow; | |
6070 | put_bh(bitmap_bh); | |
1e1c2b86 LC |
6071 | /* The range changed so it's no longer validated */ |
6072 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d AT |
6073 | goto do_more; |
6074 | } | |
c9de560d AT |
6075 | error_return: |
6076 | brelse(bitmap_bh); | |
6077 | ext4_std_error(sb, err); | |
6078 | return; | |
6079 | } | |
7360d173 | 6080 | |
8ac3939d RH |
6081 | /** |
6082 | * ext4_free_blocks() -- Free given blocks and update quota | |
6083 | * @handle: handle for this transaction | |
6084 | * @inode: inode | |
6085 | * @bh: optional buffer of the block to be freed | |
6086 | * @block: starting physical block to be freed | |
6087 | * @count: number of blocks to be freed | |
6088 | * @flags: flags used by ext4_free_blocks | |
6089 | */ | |
6090 | void ext4_free_blocks(handle_t *handle, struct inode *inode, | |
6091 | struct buffer_head *bh, ext4_fsblk_t block, | |
6092 | unsigned long count, int flags) | |
6093 | { | |
6094 | struct super_block *sb = inode->i_sb; | |
6095 | unsigned int overflow; | |
6096 | struct ext4_sb_info *sbi; | |
6097 | ||
6098 | sbi = EXT4_SB(sb); | |
6099 | ||
6100 | if (sbi->s_mount_state & EXT4_FC_REPLAY) { | |
6101 | ext4_free_blocks_simple(inode, block, count); | |
6102 | return; | |
6103 | } | |
6104 | ||
6105 | might_sleep(); | |
6106 | if (bh) { | |
6107 | if (block) | |
6108 | BUG_ON(block != bh->b_blocknr); | |
6109 | else | |
6110 | block = bh->b_blocknr; | |
6111 | } | |
6112 | ||
6113 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && | |
6114 | !ext4_inode_block_valid(inode, block, count)) { | |
6115 | ext4_error(sb, "Freeing blocks not in datazone - " | |
6116 | "block = %llu, count = %lu", block, count); | |
6117 | return; | |
6118 | } | |
1e1c2b86 | 6119 | flags |= EXT4_FREE_BLOCKS_VALIDATED; |
8ac3939d RH |
6120 | |
6121 | ext4_debug("freeing block %llu\n", block); | |
6122 | trace_ext4_free_blocks(inode, block, count, flags); | |
6123 | ||
6124 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6125 | BUG_ON(count > 1); | |
6126 | ||
6127 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
6128 | inode, bh, block); | |
6129 | } | |
6130 | ||
6131 | /* | |
6132 | * If the extent to be freed does not begin on a cluster | |
6133 | * boundary, we need to deal with partial clusters at the | |
6134 | * beginning and end of the extent. Normally we will free | |
6135 | * blocks at the beginning or the end unless we are explicitly | |
6136 | * requested to avoid doing so. | |
6137 | */ | |
6138 | overflow = EXT4_PBLK_COFF(sbi, block); | |
6139 | if (overflow) { | |
6140 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
6141 | overflow = sbi->s_cluster_ratio - overflow; | |
6142 | block += overflow; | |
6143 | if (count > overflow) | |
6144 | count -= overflow; | |
6145 | else | |
6146 | return; | |
6147 | } else { | |
6148 | block -= overflow; | |
6149 | count += overflow; | |
6150 | } | |
1e1c2b86 LC |
6151 | /* The range changed so it's no longer validated */ |
6152 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6153 | } |
6154 | overflow = EXT4_LBLK_COFF(sbi, count); | |
6155 | if (overflow) { | |
6156 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
6157 | if (count > overflow) | |
6158 | count -= overflow; | |
6159 | else | |
6160 | return; | |
6161 | } else | |
6162 | count += sbi->s_cluster_ratio - overflow; | |
1e1c2b86 LC |
6163 | /* The range changed so it's no longer validated */ |
6164 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6165 | } |
6166 | ||
6167 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6168 | int i; | |
6169 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; | |
6170 | ||
6171 | for (i = 0; i < count; i++) { | |
6172 | cond_resched(); | |
6173 | if (is_metadata) | |
6174 | bh = sb_find_get_block(inode->i_sb, block + i); | |
6175 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
6176 | } | |
6177 | } | |
6178 | ||
6179 | ext4_mb_clear_bb(handle, inode, block, count, flags); | |
6180 | return; | |
6181 | } | |
6182 | ||
2846e820 | 6183 | /** |
0529155e | 6184 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
6185 | * @handle: handle to this transaction |
6186 | * @sb: super block | |
4907cb7b | 6187 | * @block: start physical block to add to the block group |
2846e820 AG |
6188 | * @count: number of blocks to free |
6189 | * | |
e73a347b | 6190 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 6191 | */ |
cc7365df | 6192 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
6193 | ext4_fsblk_t block, unsigned long count) |
6194 | { | |
6195 | struct buffer_head *bitmap_bh = NULL; | |
6196 | struct buffer_head *gd_bh; | |
6197 | ext4_group_t block_group; | |
6198 | ext4_grpblk_t bit; | |
6199 | unsigned int i; | |
6200 | struct ext4_group_desc *desc; | |
6201 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 6202 | struct ext4_buddy e4b; |
d77147ff | 6203 | int err = 0, ret, free_clusters_count; |
6204 | ext4_grpblk_t clusters_freed; | |
6205 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
6206 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
6207 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
6208 | |
6209 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
6210 | ||
4740b830 YY |
6211 | if (count == 0) |
6212 | return 0; | |
6213 | ||
2846e820 | 6214 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
6215 | /* |
6216 | * Check to see if we are freeing blocks across a group | |
6217 | * boundary. | |
6218 | */ | |
d77147ff | 6219 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
6220 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
6221 | block_group); |
6222 | err = -EINVAL; | |
2846e820 | 6223 | goto error_return; |
cc7365df | 6224 | } |
2cd05cc3 | 6225 | |
2846e820 | 6226 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
6227 | if (IS_ERR(bitmap_bh)) { |
6228 | err = PTR_ERR(bitmap_bh); | |
6229 | bitmap_bh = NULL; | |
2846e820 | 6230 | goto error_return; |
cc7365df YY |
6231 | } |
6232 | ||
2846e820 | 6233 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
6234 | if (!desc) { |
6235 | err = -EIO; | |
2846e820 | 6236 | goto error_return; |
cc7365df | 6237 | } |
2846e820 | 6238 | |
a00b482b | 6239 | if (!ext4_sb_block_valid(sb, NULL, block, count)) { |
2846e820 AG |
6240 | ext4_error(sb, "Adding blocks in system zones - " |
6241 | "Block = %llu, count = %lu", | |
6242 | block, count); | |
cc7365df | 6243 | err = -EINVAL; |
2846e820 AG |
6244 | goto error_return; |
6245 | } | |
6246 | ||
2cd05cc3 | 6247 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
6248 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
6249 | EXT4_JTR_NONE); | |
2846e820 AG |
6250 | if (err) |
6251 | goto error_return; | |
6252 | ||
6253 | /* | |
6254 | * We are about to modify some metadata. Call the journal APIs | |
6255 | * to unshare ->b_data if a currently-committing transaction is | |
6256 | * using it | |
6257 | */ | |
6258 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 6259 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
2846e820 AG |
6260 | if (err) |
6261 | goto error_return; | |
e73a347b | 6262 | |
d77147ff | 6263 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 6264 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 6265 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
6266 | ext4_error(sb, "bit already cleared for block %llu", |
6267 | (ext4_fsblk_t)(block + i)); | |
6268 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
6269 | } else { | |
d77147ff | 6270 | clusters_freed++; |
2846e820 AG |
6271 | } |
6272 | } | |
e73a347b AG |
6273 | |
6274 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
6275 | if (err) | |
6276 | goto error_return; | |
6277 | ||
6278 | /* | |
6279 | * need to update group_info->bb_free and bitmap | |
6280 | * with group lock held. generate_buddy look at | |
6281 | * them with group lock_held | |
6282 | */ | |
2846e820 | 6283 | ext4_lock_group(sb, block_group); |
d77147ff | 6284 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
6285 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
6286 | free_clusters_count = clusters_freed + | |
6287 | ext4_free_group_clusters(sb, desc); | |
6288 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
79f1ba49 | 6289 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 6290 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 6291 | ext4_unlock_group(sb, block_group); |
57042651 | 6292 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 6293 | clusters_freed); |
2846e820 AG |
6294 | |
6295 | if (sbi->s_log_groups_per_flex) { | |
6296 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 6297 | atomic64_add(clusters_freed, |
7c990728 SJS |
6298 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6299 | flex_group)->free_clusters); | |
2846e820 | 6300 | } |
e73a347b AG |
6301 | |
6302 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
6303 | |
6304 | /* We dirtied the bitmap block */ | |
6305 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6306 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6307 | ||
6308 | /* And the group descriptor block */ | |
6309 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
6310 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
6311 | if (!err) | |
6312 | err = ret; | |
6313 | ||
6314 | error_return: | |
6315 | brelse(bitmap_bh); | |
6316 | ext4_std_error(sb, err); | |
cc7365df | 6317 | return err; |
2846e820 AG |
6318 | } |
6319 | ||
7360d173 LC |
6320 | /** |
6321 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
6322 | * @sb: super block for the file system | |
6323 | * @start: starting block of the free extent in the alloc. group | |
6324 | * @count: number of blocks to TRIM | |
7360d173 LC |
6325 | * @e4b: ext4 buddy for the group |
6326 | * | |
6327 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
6328 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
6329 | * be called with under the group lock. | |
6330 | */ | |
bd2eea8d WJ |
6331 | static int ext4_trim_extent(struct super_block *sb, |
6332 | int start, int count, struct ext4_buddy *e4b) | |
e2cbd587 | 6333 | __releases(bitlock) |
6334 | __acquires(bitlock) | |
7360d173 LC |
6335 | { |
6336 | struct ext4_free_extent ex; | |
bd2eea8d | 6337 | ext4_group_t group = e4b->bd_group; |
d71c1ae2 | 6338 | int ret = 0; |
7360d173 | 6339 | |
b3d4c2b1 TM |
6340 | trace_ext4_trim_extent(sb, group, start, count); |
6341 | ||
7360d173 LC |
6342 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
6343 | ||
6344 | ex.fe_start = start; | |
6345 | ex.fe_group = group; | |
6346 | ex.fe_len = count; | |
6347 | ||
6348 | /* | |
6349 | * Mark blocks used, so no one can reuse them while | |
6350 | * being trimmed. | |
6351 | */ | |
6352 | mb_mark_used(e4b, &ex); | |
6353 | ext4_unlock_group(sb, group); | |
a0154344 | 6354 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
6355 | ext4_lock_group(sb, group); |
6356 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 6357 | return ret; |
7360d173 LC |
6358 | } |
6359 | ||
6920b391 WJ |
6360 | static int ext4_try_to_trim_range(struct super_block *sb, |
6361 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
6362 | ext4_grpblk_t max, ext4_grpblk_t minblocks) | |
a5fda113 TT |
6363 | __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) |
6364 | __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) | |
6920b391 WJ |
6365 | { |
6366 | ext4_grpblk_t next, count, free_count; | |
6367 | void *bitmap; | |
6920b391 WJ |
6368 | |
6369 | bitmap = e4b->bd_bitmap; | |
6370 | start = (e4b->bd_info->bb_first_free > start) ? | |
6371 | e4b->bd_info->bb_first_free : start; | |
6372 | count = 0; | |
6373 | free_count = 0; | |
6374 | ||
6375 | while (start <= max) { | |
6376 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
6377 | if (start > max) | |
6378 | break; | |
6379 | next = mb_find_next_bit(bitmap, max + 1, start); | |
6380 | ||
6381 | if ((next - start) >= minblocks) { | |
afcc4e32 LB |
6382 | int ret = ext4_trim_extent(sb, start, next - start, e4b); |
6383 | ||
6920b391 WJ |
6384 | if (ret && ret != -EOPNOTSUPP) |
6385 | break; | |
6920b391 WJ |
6386 | count += next - start; |
6387 | } | |
6388 | free_count += next - start; | |
6389 | start = next + 1; | |
6390 | ||
6391 | if (fatal_signal_pending(current)) { | |
6392 | count = -ERESTARTSYS; | |
6393 | break; | |
6394 | } | |
6395 | ||
6396 | if (need_resched()) { | |
6397 | ext4_unlock_group(sb, e4b->bd_group); | |
6398 | cond_resched(); | |
6399 | ext4_lock_group(sb, e4b->bd_group); | |
6400 | } | |
6401 | ||
6402 | if ((e4b->bd_info->bb_free - free_count) < minblocks) | |
6403 | break; | |
6404 | } | |
6405 | ||
6406 | return count; | |
6407 | } | |
6408 | ||
7360d173 LC |
6409 | /** |
6410 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
6411 | * @sb: super block for file system | |
22612283 | 6412 | * @group: group to be trimmed |
7360d173 LC |
6413 | * @start: first group block to examine |
6414 | * @max: last group block to examine | |
6415 | * @minblocks: minimum extent block count | |
d63c00ea | 6416 | * @set_trimmed: set the trimmed flag if at least one block is trimmed |
7360d173 | 6417 | * |
7360d173 LC |
6418 | * ext4_trim_all_free walks through group's block bitmap searching for free |
6419 | * extents. When the free extent is found, mark it as used in group buddy | |
6420 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
b6f5558c | 6421 | * the group buddy bitmap. |
7360d173 | 6422 | */ |
0b75a840 | 6423 | static ext4_grpblk_t |
78944086 LC |
6424 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
6425 | ext4_grpblk_t start, ext4_grpblk_t max, | |
d63c00ea | 6426 | ext4_grpblk_t minblocks, bool set_trimmed) |
7360d173 | 6427 | { |
78944086 | 6428 | struct ext4_buddy e4b; |
6920b391 | 6429 | int ret; |
7360d173 | 6430 | |
b3d4c2b1 TM |
6431 | trace_ext4_trim_all_free(sb, group, start, max); |
6432 | ||
78944086 LC |
6433 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
6434 | if (ret) { | |
9651e6b2 KK |
6435 | ext4_warning(sb, "Error %d loading buddy information for %u", |
6436 | ret, group); | |
78944086 LC |
6437 | return ret; |
6438 | } | |
28739eea LC |
6439 | |
6440 | ext4_lock_group(sb, group); | |
7360d173 | 6441 | |
6920b391 | 6442 | if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || |
2327fb2e | 6443 | minblocks < EXT4_SB(sb)->s_last_trim_minblks) { |
6920b391 | 6444 | ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); |
d63c00ea | 6445 | if (ret >= 0 && set_trimmed) |
6920b391 WJ |
6446 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
6447 | } else { | |
6448 | ret = 0; | |
7360d173 | 6449 | } |
3d56b8d2 | 6450 | |
7360d173 | 6451 | ext4_unlock_group(sb, group); |
78944086 | 6452 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
6453 | |
6454 | ext4_debug("trimmed %d blocks in the group %d\n", | |
6920b391 | 6455 | ret, group); |
7360d173 | 6456 | |
d71c1ae2 | 6457 | return ret; |
7360d173 LC |
6458 | } |
6459 | ||
6460 | /** | |
6461 | * ext4_trim_fs() -- trim ioctl handle function | |
6462 | * @sb: superblock for filesystem | |
6463 | * @range: fstrim_range structure | |
6464 | * | |
6465 | * start: First Byte to trim | |
6466 | * len: number of Bytes to trim from start | |
6467 | * minlen: minimum extent length in Bytes | |
6468 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
6469 | * start to start+len. For each such a group ext4_trim_all_free function | |
6470 | * is invoked to trim all free space. | |
6471 | */ | |
6472 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
6473 | { | |
7b47ef52 | 6474 | unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); |
78944086 | 6475 | struct ext4_group_info *grp; |
913eed83 | 6476 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 6477 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 6478 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
6479 | ext4_fsblk_t first_data_blk = |
6480 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 6481 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
d63c00ea | 6482 | bool whole_group, eof = false; |
7360d173 LC |
6483 | int ret = 0; |
6484 | ||
6485 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 6486 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
6487 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
6488 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 6489 | |
5de35e8d LC |
6490 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
6491 | start >= max_blks || | |
6492 | range->len < sb->s_blocksize) | |
7360d173 | 6493 | return -EINVAL; |
173b6e38 | 6494 | /* No point to try to trim less than discard granularity */ |
7b47ef52 | 6495 | if (range->minlen < discard_granularity) { |
173b6e38 | 6496 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
7b47ef52 | 6497 | discard_granularity >> sb->s_blocksize_bits); |
173b6e38 JK |
6498 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) |
6499 | goto out; | |
6500 | } | |
d63c00ea | 6501 | if (end >= max_blks - 1) { |
913eed83 | 6502 | end = max_blks - 1; |
d63c00ea DM |
6503 | eof = true; |
6504 | } | |
913eed83 | 6505 | if (end <= first_data_blk) |
22f10457 | 6506 | goto out; |
913eed83 | 6507 | if (start < first_data_blk) |
0f0a25bf | 6508 | start = first_data_blk; |
7360d173 | 6509 | |
913eed83 | 6510 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 6511 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 6512 | &first_group, &first_cluster); |
913eed83 | 6513 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 6514 | &last_group, &last_cluster); |
7360d173 | 6515 | |
913eed83 LC |
6516 | /* end now represents the last cluster to discard in this group */ |
6517 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
d63c00ea | 6518 | whole_group = true; |
7360d173 LC |
6519 | |
6520 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
6521 | grp = ext4_get_group_info(sb, group); |
6522 | /* We only do this if the grp has never been initialized */ | |
6523 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 6524 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
6525 | if (ret) |
6526 | break; | |
7360d173 LC |
6527 | } |
6528 | ||
0ba08517 | 6529 | /* |
913eed83 LC |
6530 | * For all the groups except the last one, last cluster will |
6531 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
6532 | * change it for the last group, note that last_cluster is | |
6533 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 6534 | */ |
d63c00ea | 6535 | if (group == last_group) { |
913eed83 | 6536 | end = last_cluster; |
d63c00ea DM |
6537 | whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1; |
6538 | } | |
78944086 | 6539 | if (grp->bb_free >= minlen) { |
7137d7a4 | 6540 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
d63c00ea | 6541 | end, minlen, whole_group); |
7360d173 LC |
6542 | if (cnt < 0) { |
6543 | ret = cnt; | |
7360d173 LC |
6544 | break; |
6545 | } | |
21e7fd22 | 6546 | trimmed += cnt; |
7360d173 | 6547 | } |
913eed83 LC |
6548 | |
6549 | /* | |
6550 | * For every group except the first one, we are sure | |
6551 | * that the first cluster to discard will be cluster #0. | |
6552 | */ | |
7137d7a4 | 6553 | first_cluster = 0; |
7360d173 | 6554 | } |
7360d173 | 6555 | |
3d56b8d2 | 6556 | if (!ret) |
2327fb2e | 6557 | EXT4_SB(sb)->s_last_trim_minblks = minlen; |
3d56b8d2 | 6558 | |
22f10457 | 6559 | out: |
aaf7d73e | 6560 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
6561 | return ret; |
6562 | } | |
0c9ec4be DW |
6563 | |
6564 | /* Iterate all the free extents in the group. */ | |
6565 | int | |
6566 | ext4_mballoc_query_range( | |
6567 | struct super_block *sb, | |
6568 | ext4_group_t group, | |
6569 | ext4_grpblk_t start, | |
6570 | ext4_grpblk_t end, | |
6571 | ext4_mballoc_query_range_fn formatter, | |
6572 | void *priv) | |
6573 | { | |
6574 | void *bitmap; | |
6575 | ext4_grpblk_t next; | |
6576 | struct ext4_buddy e4b; | |
6577 | int error; | |
6578 | ||
6579 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
6580 | if (error) | |
6581 | return error; | |
6582 | bitmap = e4b.bd_bitmap; | |
6583 | ||
6584 | ext4_lock_group(sb, group); | |
6585 | ||
6586 | start = (e4b.bd_info->bb_first_free > start) ? | |
6587 | e4b.bd_info->bb_first_free : start; | |
6588 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
6589 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
6590 | ||
6591 | while (start <= end) { | |
6592 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
6593 | if (start > end) | |
6594 | break; | |
6595 | next = mb_find_next_bit(bitmap, end + 1, start); | |
6596 | ||
6597 | ext4_unlock_group(sb, group); | |
6598 | error = formatter(sb, group, start, next - start, priv); | |
6599 | if (error) | |
6600 | goto out_unload; | |
6601 | ext4_lock_group(sb, group); | |
6602 | ||
6603 | start = next + 1; | |
6604 | } | |
6605 | ||
6606 | ext4_unlock_group(sb, group); | |
6607 | out_unload: | |
6608 | ext4_mb_unload_buddy(&e4b); | |
6609 | ||
6610 | return error; | |
6611 | } |