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); | |
5354b2af TT |
748 | if (!grp) |
749 | return NULL; | |
c9de560d AT |
750 | list_for_each(cur, &grp->bb_prealloc_list) { |
751 | ext4_group_t groupnr; | |
752 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
753 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
754 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 755 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 756 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
757 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
758 | } | |
759 | return 0; | |
760 | } | |
761 | #undef MB_CHECK_ASSERT | |
762 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 763 | __FILE__, __func__, __LINE__) |
c9de560d AT |
764 | #else |
765 | #define mb_check_buddy(e4b) | |
766 | #endif | |
767 | ||
7c786059 CL |
768 | /* |
769 | * Divide blocks started from @first with length @len into | |
770 | * smaller chunks with power of 2 blocks. | |
771 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
772 | * then increase bb_counters[] for corresponded chunk size. | |
773 | */ | |
c9de560d | 774 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 775 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
776 | struct ext4_group_info *grp) |
777 | { | |
778 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
779 | ext4_grpblk_t min; |
780 | ext4_grpblk_t max; | |
781 | ext4_grpblk_t chunk; | |
69e43e8c | 782 | unsigned int border; |
c9de560d | 783 | |
7137d7a4 | 784 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
785 | |
786 | border = 2 << sb->s_blocksize_bits; | |
787 | ||
788 | while (len > 0) { | |
789 | /* find how many blocks can be covered since this position */ | |
790 | max = ffs(first | border) - 1; | |
791 | ||
792 | /* find how many blocks of power 2 we need to mark */ | |
793 | min = fls(len) - 1; | |
794 | ||
795 | if (max < min) | |
796 | min = max; | |
797 | chunk = 1 << min; | |
798 | ||
799 | /* mark multiblock chunks only */ | |
800 | grp->bb_counters[min]++; | |
801 | if (min > 0) | |
802 | mb_clear_bit(first >> min, | |
803 | buddy + sbi->s_mb_offsets[min]); | |
804 | ||
805 | len -= chunk; | |
806 | first += chunk; | |
807 | } | |
808 | } | |
809 | ||
83e80a6e | 810 | static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) |
196e402a | 811 | { |
83e80a6e | 812 | int order; |
196e402a | 813 | |
83e80a6e JK |
814 | /* |
815 | * We don't bother with a special lists groups with only 1 block free | |
816 | * extents and for completely empty groups. | |
817 | */ | |
818 | order = fls(len) - 2; | |
819 | if (order < 0) | |
820 | return 0; | |
821 | if (order == MB_NUM_ORDERS(sb)) | |
822 | order--; | |
823 | return order; | |
196e402a HS |
824 | } |
825 | ||
83e80a6e | 826 | /* Move group to appropriate avg_fragment_size list */ |
196e402a HS |
827 | static void |
828 | mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) | |
829 | { | |
830 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
83e80a6e | 831 | int new_order; |
196e402a HS |
832 | |
833 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) | |
834 | return; | |
835 | ||
83e80a6e JK |
836 | new_order = mb_avg_fragment_size_order(sb, |
837 | grp->bb_free / grp->bb_fragments); | |
838 | if (new_order == grp->bb_avg_fragment_size_order) | |
839 | return; | |
196e402a | 840 | |
83e80a6e JK |
841 | if (grp->bb_avg_fragment_size_order != -1) { |
842 | write_lock(&sbi->s_mb_avg_fragment_size_locks[ | |
843 | grp->bb_avg_fragment_size_order]); | |
844 | list_del(&grp->bb_avg_fragment_size_node); | |
845 | write_unlock(&sbi->s_mb_avg_fragment_size_locks[ | |
846 | grp->bb_avg_fragment_size_order]); | |
847 | } | |
848 | grp->bb_avg_fragment_size_order = new_order; | |
849 | write_lock(&sbi->s_mb_avg_fragment_size_locks[ | |
850 | grp->bb_avg_fragment_size_order]); | |
851 | list_add_tail(&grp->bb_avg_fragment_size_node, | |
852 | &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]); | |
853 | write_unlock(&sbi->s_mb_avg_fragment_size_locks[ | |
854 | grp->bb_avg_fragment_size_order]); | |
196e402a HS |
855 | } |
856 | ||
857 | /* | |
858 | * Choose next group by traversing largest_free_order lists. Updates *new_cr if | |
859 | * cr level needs an update. | |
860 | */ | |
861 | static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, | |
862 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
863 | { | |
864 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
865 | struct ext4_group_info *iter, *grp; | |
866 | int i; | |
867 | ||
868 | if (ac->ac_status == AC_STATUS_FOUND) | |
869 | return; | |
870 | ||
871 | if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED)) | |
872 | atomic_inc(&sbi->s_bal_cr0_bad_suggestions); | |
873 | ||
874 | grp = NULL; | |
875 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
876 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) | |
877 | continue; | |
878 | read_lock(&sbi->s_mb_largest_free_orders_locks[i]); | |
879 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) { | |
880 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
881 | continue; | |
882 | } | |
883 | grp = NULL; | |
884 | list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], | |
885 | bb_largest_free_order_node) { | |
886 | if (sbi->s_mb_stats) | |
887 | atomic64_inc(&sbi->s_bal_cX_groups_considered[0]); | |
888 | if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) { | |
889 | grp = iter; | |
890 | break; | |
891 | } | |
892 | } | |
893 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
894 | if (grp) | |
895 | break; | |
896 | } | |
897 | ||
898 | if (!grp) { | |
899 | /* Increment cr and search again */ | |
900 | *new_cr = 1; | |
901 | } else { | |
902 | *group = grp->bb_group; | |
196e402a HS |
903 | ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED; |
904 | } | |
905 | } | |
906 | ||
907 | /* | |
83e80a6e JK |
908 | * Choose next group by traversing average fragment size list of suitable |
909 | * order. Updates *new_cr if cr level needs an update. | |
196e402a HS |
910 | */ |
911 | static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, | |
912 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
913 | { | |
914 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
a078dff8 | 915 | struct ext4_group_info *grp = NULL, *iter; |
83e80a6e | 916 | int i; |
196e402a HS |
917 | |
918 | if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) { | |
919 | if (sbi->s_mb_stats) | |
920 | atomic_inc(&sbi->s_bal_cr1_bad_suggestions); | |
83e80a6e JK |
921 | } |
922 | ||
923 | for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); | |
924 | i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
925 | if (list_empty(&sbi->s_mb_avg_fragment_size[i])) | |
926 | continue; | |
927 | read_lock(&sbi->s_mb_avg_fragment_size_locks[i]); | |
928 | if (list_empty(&sbi->s_mb_avg_fragment_size[i])) { | |
929 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]); | |
930 | continue; | |
931 | } | |
83e80a6e JK |
932 | list_for_each_entry(iter, &sbi->s_mb_avg_fragment_size[i], |
933 | bb_avg_fragment_size_node) { | |
196e402a HS |
934 | if (sbi->s_mb_stats) |
935 | atomic64_inc(&sbi->s_bal_cX_groups_considered[1]); | |
83e80a6e JK |
936 | if (likely(ext4_mb_good_group(ac, iter->bb_group, 1))) { |
937 | grp = iter; | |
196e402a | 938 | break; |
196e402a HS |
939 | } |
940 | } | |
83e80a6e JK |
941 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]); |
942 | if (grp) | |
943 | break; | |
196e402a HS |
944 | } |
945 | ||
83e80a6e | 946 | if (grp) { |
196e402a HS |
947 | *group = grp->bb_group; |
948 | ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED; | |
949 | } else { | |
950 | *new_cr = 2; | |
951 | } | |
196e402a HS |
952 | } |
953 | ||
954 | static inline int should_optimize_scan(struct ext4_allocation_context *ac) | |
955 | { | |
956 | if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) | |
957 | return 0; | |
958 | if (ac->ac_criteria >= 2) | |
959 | return 0; | |
077d0c2c | 960 | if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) |
196e402a HS |
961 | return 0; |
962 | return 1; | |
963 | } | |
964 | ||
965 | /* | |
966 | * Return next linear group for allocation. If linear traversal should not be | |
967 | * performed, this function just returns the same group | |
968 | */ | |
969 | static int | |
970 | next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups) | |
971 | { | |
972 | if (!should_optimize_scan(ac)) | |
973 | goto inc_and_return; | |
974 | ||
975 | if (ac->ac_groups_linear_remaining) { | |
976 | ac->ac_groups_linear_remaining--; | |
977 | goto inc_and_return; | |
978 | } | |
979 | ||
196e402a HS |
980 | return group; |
981 | inc_and_return: | |
982 | /* | |
983 | * Artificially restricted ngroups for non-extent | |
984 | * files makes group > ngroups possible on first loop. | |
985 | */ | |
986 | return group + 1 >= ngroups ? 0 : group + 1; | |
987 | } | |
988 | ||
989 | /* | |
990 | * ext4_mb_choose_next_group: choose next group for allocation. | |
991 | * | |
992 | * @ac Allocation Context | |
993 | * @new_cr This is an output parameter. If the there is no good group | |
994 | * available at current CR level, this field is updated to indicate | |
995 | * the new cr level that should be used. | |
996 | * @group This is an input / output parameter. As an input it indicates the | |
997 | * next group that the allocator intends to use for allocation. As | |
998 | * output, this field indicates the next group that should be used as | |
999 | * determined by the optimization functions. | |
1000 | * @ngroups Total number of groups | |
1001 | */ | |
1002 | static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, | |
1003 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
1004 | { | |
1005 | *new_cr = ac->ac_criteria; | |
1006 | ||
4fca50d4 JK |
1007 | if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) { |
1008 | *group = next_linear_group(ac, *group, ngroups); | |
196e402a | 1009 | return; |
4fca50d4 | 1010 | } |
196e402a HS |
1011 | |
1012 | if (*new_cr == 0) { | |
1013 | ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups); | |
1014 | } else if (*new_cr == 1) { | |
1015 | ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups); | |
1016 | } else { | |
1017 | /* | |
1018 | * TODO: For CR=2, we can arrange groups in an rb tree sorted by | |
1019 | * bb_free. But until that happens, we should never come here. | |
1020 | */ | |
1021 | WARN_ON(1); | |
1022 | } | |
1023 | } | |
1024 | ||
8a57d9d6 CW |
1025 | /* |
1026 | * Cache the order of the largest free extent we have available in this block | |
1027 | * group. | |
1028 | */ | |
1029 | static void | |
1030 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
1031 | { | |
196e402a | 1032 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
8a57d9d6 | 1033 | int i; |
8a57d9d6 | 1034 | |
1940265e JK |
1035 | for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) |
1036 | if (grp->bb_counters[i] > 0) | |
1037 | break; | |
1038 | /* No need to move between order lists? */ | |
1039 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || | |
1040 | i == grp->bb_largest_free_order) { | |
1041 | grp->bb_largest_free_order = i; | |
1042 | return; | |
1043 | } | |
1044 | ||
1045 | if (grp->bb_largest_free_order >= 0) { | |
196e402a HS |
1046 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1047 | grp->bb_largest_free_order]); | |
1048 | list_del_init(&grp->bb_largest_free_order_node); | |
1049 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1050 | grp->bb_largest_free_order]); | |
1051 | } | |
1940265e JK |
1052 | grp->bb_largest_free_order = i; |
1053 | if (grp->bb_largest_free_order >= 0 && grp->bb_free) { | |
196e402a HS |
1054 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1055 | grp->bb_largest_free_order]); | |
1056 | list_add_tail(&grp->bb_largest_free_order_node, | |
1057 | &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); | |
1058 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1059 | grp->bb_largest_free_order]); | |
1060 | } | |
8a57d9d6 CW |
1061 | } |
1062 | ||
089ceecc ES |
1063 | static noinline_for_stack |
1064 | void ext4_mb_generate_buddy(struct super_block *sb, | |
5354b2af TT |
1065 | void *buddy, void *bitmap, ext4_group_t group, |
1066 | struct ext4_group_info *grp) | |
c9de560d | 1067 | { |
e43bb4e6 | 1068 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 1069 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
1070 | ext4_grpblk_t i = 0; |
1071 | ext4_grpblk_t first; | |
1072 | ext4_grpblk_t len; | |
c9de560d AT |
1073 | unsigned free = 0; |
1074 | unsigned fragments = 0; | |
1075 | unsigned long long period = get_cycles(); | |
1076 | ||
1077 | /* initialize buddy from bitmap which is aggregation | |
1078 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 1079 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
1080 | grp->bb_first_free = i; |
1081 | while (i < max) { | |
1082 | fragments++; | |
1083 | first = i; | |
ffad0a44 | 1084 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
1085 | len = i - first; |
1086 | free += len; | |
1087 | if (len > 1) | |
1088 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
1089 | else | |
1090 | grp->bb_counters[0]++; | |
1091 | if (i < max) | |
ffad0a44 | 1092 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
1093 | } |
1094 | grp->bb_fragments = fragments; | |
1095 | ||
1096 | if (free != grp->bb_free) { | |
e29136f8 | 1097 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
1098 | "block bitmap and bg descriptor " |
1099 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 1100 | free, grp->bb_free); |
e56eb659 | 1101 | /* |
163a203d | 1102 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
1103 | * corrupt and update bb_free using bitmap value |
1104 | */ | |
c9de560d | 1105 | grp->bb_free = free; |
db79e6d1 WS |
1106 | ext4_mark_group_bitmap_corrupted(sb, group, |
1107 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 1108 | } |
8a57d9d6 | 1109 | mb_set_largest_free_order(sb, grp); |
83e80a6e | 1110 | mb_update_avg_fragment_size(sb, grp); |
c9de560d AT |
1111 | |
1112 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
1113 | ||
1114 | period = get_cycles() - period; | |
67d25186 HS |
1115 | atomic_inc(&sbi->s_mb_buddies_generated); |
1116 | atomic64_add(period, &sbi->s_mb_generation_time); | |
c9de560d AT |
1117 | } |
1118 | ||
1119 | /* The buddy information is attached the buddy cache inode | |
1120 | * for convenience. The information regarding each group | |
1121 | * is loaded via ext4_mb_load_buddy. The information involve | |
1122 | * block bitmap and buddy information. The information are | |
1123 | * stored in the inode as | |
1124 | * | |
1125 | * { page } | |
c3a326a6 | 1126 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
1127 | * |
1128 | * | |
1129 | * one block each for bitmap and buddy information. | |
1130 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 1131 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
1132 | * So it can have information regarding groups_per_page which |
1133 | * is blocks_per_page/2 | |
8a57d9d6 CW |
1134 | * |
1135 | * Locking note: This routine takes the block group lock of all groups | |
1136 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
1137 | */ |
1138 | ||
adb7ef60 | 1139 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 1140 | { |
8df9675f | 1141 | ext4_group_t ngroups; |
c9de560d AT |
1142 | int blocksize; |
1143 | int blocks_per_page; | |
1144 | int groups_per_page; | |
1145 | int err = 0; | |
1146 | int i; | |
813e5727 | 1147 | ext4_group_t first_group, group; |
c9de560d AT |
1148 | int first_block; |
1149 | struct super_block *sb; | |
1150 | struct buffer_head *bhs; | |
fa77dcfa | 1151 | struct buffer_head **bh = NULL; |
c9de560d AT |
1152 | struct inode *inode; |
1153 | char *data; | |
1154 | char *bitmap; | |
9b8b7d35 | 1155 | struct ext4_group_info *grinfo; |
c9de560d | 1156 | |
c9de560d AT |
1157 | inode = page->mapping->host; |
1158 | sb = inode->i_sb; | |
8df9675f | 1159 | ngroups = ext4_get_groups_count(sb); |
93407472 | 1160 | blocksize = i_blocksize(inode); |
09cbfeaf | 1161 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d | 1162 | |
d3df1453 RH |
1163 | mb_debug(sb, "init page %lu\n", page->index); |
1164 | ||
c9de560d AT |
1165 | groups_per_page = blocks_per_page >> 1; |
1166 | if (groups_per_page == 0) | |
1167 | groups_per_page = 1; | |
1168 | ||
1169 | /* allocate buffer_heads to read bitmaps */ | |
1170 | if (groups_per_page > 1) { | |
c9de560d | 1171 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 1172 | bh = kzalloc(i, gfp); |
139f46d3 KS |
1173 | if (bh == NULL) |
1174 | return -ENOMEM; | |
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); |
5354b2af TT |
1186 | if (!grinfo) |
1187 | continue; | |
9b8b7d35 AG |
1188 | /* |
1189 | * If page is uptodate then we came here after online resize | |
1190 | * which added some new uninitialized group info structs, so | |
1191 | * we must skip all initialized uptodate buddies on the page, | |
1192 | * which may be currently in use by an allocating task. | |
1193 | */ | |
1194 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
1195 | bh[i] = NULL; | |
1196 | continue; | |
1197 | } | |
cfd73237 | 1198 | bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
9008a58e DW |
1199 | if (IS_ERR(bh[i])) { |
1200 | err = PTR_ERR(bh[i]); | |
1201 | bh[i] = NULL; | |
c9de560d | 1202 | goto out; |
2ccb5fb9 | 1203 | } |
d3df1453 | 1204 | mb_debug(sb, "read bitmap for group %u\n", group); |
c9de560d AT |
1205 | } |
1206 | ||
1207 | /* wait for I/O completion */ | |
813e5727 | 1208 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
1209 | int err2; |
1210 | ||
1211 | if (!bh[i]) | |
1212 | continue; | |
1213 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
1214 | if (!err) | |
1215 | err = err2; | |
813e5727 | 1216 | } |
c9de560d AT |
1217 | |
1218 | first_block = page->index * blocks_per_page; | |
1219 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 1220 | group = (first_block + i) >> 1; |
8df9675f | 1221 | if (group >= ngroups) |
c9de560d AT |
1222 | break; |
1223 | ||
9b8b7d35 AG |
1224 | if (!bh[group - first_group]) |
1225 | /* skip initialized uptodate buddy */ | |
1226 | continue; | |
1227 | ||
bbdc322f LC |
1228 | if (!buffer_verified(bh[group - first_group])) |
1229 | /* Skip faulty bitmaps */ | |
1230 | continue; | |
1231 | err = 0; | |
1232 | ||
c9de560d AT |
1233 | /* |
1234 | * data carry information regarding this | |
1235 | * particular group in the format specified | |
1236 | * above | |
1237 | * | |
1238 | */ | |
1239 | data = page_address(page) + (i * blocksize); | |
1240 | bitmap = bh[group - first_group]->b_data; | |
1241 | ||
1242 | /* | |
1243 | * We place the buddy block and bitmap block | |
1244 | * close together | |
1245 | */ | |
1246 | if ((first_block + i) & 1) { | |
1247 | /* this is block of buddy */ | |
1248 | BUG_ON(incore == NULL); | |
d3df1453 | 1249 | mb_debug(sb, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 1250 | group, page->index, i * blocksize); |
f307333e | 1251 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d | 1252 | grinfo = ext4_get_group_info(sb, group); |
5354b2af TT |
1253 | if (!grinfo) { |
1254 | err = -EFSCORRUPTED; | |
1255 | goto out; | |
1256 | } | |
c9de560d AT |
1257 | grinfo->bb_fragments = 0; |
1258 | memset(grinfo->bb_counters, 0, | |
1927805e | 1259 | sizeof(*grinfo->bb_counters) * |
4b68f6df | 1260 | (MB_NUM_ORDERS(sb))); |
c9de560d AT |
1261 | /* |
1262 | * incore got set to the group block bitmap below | |
1263 | */ | |
7a2fcbf7 | 1264 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
1265 | /* init the buddy */ |
1266 | memset(data, 0xff, blocksize); | |
5354b2af | 1267 | ext4_mb_generate_buddy(sb, data, incore, group, grinfo); |
7a2fcbf7 | 1268 | ext4_unlock_group(sb, group); |
c9de560d AT |
1269 | incore = NULL; |
1270 | } else { | |
1271 | /* this is block of bitmap */ | |
1272 | BUG_ON(incore != NULL); | |
d3df1453 | 1273 | mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 1274 | group, page->index, i * blocksize); |
f307333e | 1275 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
1276 | |
1277 | /* see comments in ext4_mb_put_pa() */ | |
1278 | ext4_lock_group(sb, group); | |
1279 | memcpy(data, bitmap, blocksize); | |
1280 | ||
1281 | /* mark all preallocated blks used in in-core bitmap */ | |
1282 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 1283 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
1284 | ext4_unlock_group(sb, group); |
1285 | ||
1286 | /* set incore so that the buddy information can be | |
1287 | * generated using this | |
1288 | */ | |
1289 | incore = data; | |
1290 | } | |
1291 | } | |
1292 | SetPageUptodate(page); | |
1293 | ||
1294 | out: | |
1295 | if (bh) { | |
9b8b7d35 | 1296 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
1297 | brelse(bh[i]); |
1298 | if (bh != &bhs) | |
1299 | kfree(bh); | |
1300 | } | |
1301 | return err; | |
1302 | } | |
1303 | ||
eee4adc7 | 1304 | /* |
2de8807b AG |
1305 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
1306 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1307 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
1308 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 1309 | */ |
2de8807b | 1310 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1311 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1312 | { |
2de8807b AG |
1313 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1314 | int block, pnum, poff; | |
eee4adc7 | 1315 | int blocks_per_page; |
2de8807b AG |
1316 | struct page *page; |
1317 | ||
1318 | e4b->bd_buddy_page = NULL; | |
1319 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 1320 | |
09cbfeaf | 1321 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1322 | /* |
1323 | * the buddy cache inode stores the block bitmap | |
1324 | * and buddy information in consecutive blocks. | |
1325 | * So for each group we need two blocks. | |
1326 | */ | |
1327 | block = group * 2; | |
1328 | pnum = block / blocks_per_page; | |
2de8807b | 1329 | poff = block % blocks_per_page; |
adb7ef60 | 1330 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1331 | if (!page) |
c57ab39b | 1332 | return -ENOMEM; |
2de8807b AG |
1333 | BUG_ON(page->mapping != inode->i_mapping); |
1334 | e4b->bd_bitmap_page = page; | |
1335 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1336 | ||
1337 | if (blocks_per_page >= 2) { | |
1338 | /* buddy and bitmap are on the same page */ | |
1339 | return 0; | |
eee4adc7 | 1340 | } |
2de8807b AG |
1341 | |
1342 | block++; | |
1343 | pnum = block / blocks_per_page; | |
adb7ef60 | 1344 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1345 | if (!page) |
c57ab39b | 1346 | return -ENOMEM; |
2de8807b AG |
1347 | BUG_ON(page->mapping != inode->i_mapping); |
1348 | e4b->bd_buddy_page = page; | |
1349 | return 0; | |
eee4adc7 ES |
1350 | } |
1351 | ||
2de8807b | 1352 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1353 | { |
2de8807b AG |
1354 | if (e4b->bd_bitmap_page) { |
1355 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1356 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1357 | } |
1358 | if (e4b->bd_buddy_page) { | |
1359 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1360 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1361 | } |
eee4adc7 ES |
1362 | } |
1363 | ||
8a57d9d6 CW |
1364 | /* |
1365 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1366 | * block group lock of all groups for this page; do not hold the BG lock when | |
1367 | * calling this routine! | |
1368 | */ | |
b6a758ec | 1369 | static noinline_for_stack |
adb7ef60 | 1370 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1371 | { |
1372 | ||
b6a758ec | 1373 | struct ext4_group_info *this_grp; |
2de8807b AG |
1374 | struct ext4_buddy e4b; |
1375 | struct page *page; | |
1376 | int ret = 0; | |
b6a758ec | 1377 | |
b10a44c3 | 1378 | might_sleep(); |
d3df1453 | 1379 | mb_debug(sb, "init group %u\n", group); |
b6a758ec | 1380 | this_grp = ext4_get_group_info(sb, group); |
5354b2af TT |
1381 | if (!this_grp) |
1382 | return -EFSCORRUPTED; | |
1383 | ||
b6a758ec | 1384 | /* |
08c3a813 AK |
1385 | * This ensures that we don't reinit the buddy cache |
1386 | * page which map to the group from which we are already | |
1387 | * allocating. If we are looking at the buddy cache we would | |
1388 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1389 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1390 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1391 | * page accessed. | |
b6a758ec | 1392 | */ |
adb7ef60 | 1393 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1394 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1395 | /* |
1396 | * somebody initialized the group | |
1397 | * return without doing anything | |
1398 | */ | |
b6a758ec AK |
1399 | goto err; |
1400 | } | |
2de8807b AG |
1401 | |
1402 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1403 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1404 | if (ret) |
1405 | goto err; | |
1406 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1407 | ret = -EIO; |
1408 | goto err; | |
1409 | } | |
b6a758ec | 1410 | |
2de8807b | 1411 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1412 | /* |
1413 | * If both the bitmap and buddy are in | |
1414 | * the same page we don't need to force | |
1415 | * init the buddy | |
1416 | */ | |
2de8807b AG |
1417 | ret = 0; |
1418 | goto err; | |
b6a758ec | 1419 | } |
2de8807b AG |
1420 | /* init buddy cache */ |
1421 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1422 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1423 | if (ret) |
1424 | goto err; | |
1425 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1426 | ret = -EIO; |
1427 | goto err; | |
1428 | } | |
b6a758ec | 1429 | err: |
2de8807b | 1430 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1431 | return ret; |
1432 | } | |
1433 | ||
8a57d9d6 CW |
1434 | /* |
1435 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1436 | * block group lock of all groups for this page; do not hold the BG lock when | |
1437 | * calling this routine! | |
1438 | */ | |
4ddfef7b | 1439 | static noinline_for_stack int |
adb7ef60 KK |
1440 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1441 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1442 | { |
c9de560d AT |
1443 | int blocks_per_page; |
1444 | int block; | |
1445 | int pnum; | |
1446 | int poff; | |
1447 | struct page *page; | |
fdf6c7a7 | 1448 | int ret; |
920313a7 AK |
1449 | struct ext4_group_info *grp; |
1450 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1451 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1452 | |
b10a44c3 | 1453 | might_sleep(); |
d3df1453 | 1454 | mb_debug(sb, "load group %u\n", group); |
c9de560d | 1455 | |
09cbfeaf | 1456 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1457 | grp = ext4_get_group_info(sb, group); |
5354b2af TT |
1458 | if (!grp) |
1459 | return -EFSCORRUPTED; | |
c9de560d AT |
1460 | |
1461 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1462 | e4b->bd_info = grp; |
c9de560d AT |
1463 | e4b->bd_sb = sb; |
1464 | e4b->bd_group = group; | |
1465 | e4b->bd_buddy_page = NULL; | |
1466 | e4b->bd_bitmap_page = NULL; | |
1467 | ||
f41c0750 | 1468 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1469 | /* |
1470 | * we need full data about the group | |
1471 | * to make a good selection | |
1472 | */ | |
adb7ef60 | 1473 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1474 | if (ret) |
1475 | return ret; | |
f41c0750 AK |
1476 | } |
1477 | ||
c9de560d AT |
1478 | /* |
1479 | * the buddy cache inode stores the block bitmap | |
1480 | * and buddy information in consecutive blocks. | |
1481 | * So for each group we need two blocks. | |
1482 | */ | |
1483 | block = group * 2; | |
1484 | pnum = block / blocks_per_page; | |
1485 | poff = block % blocks_per_page; | |
1486 | ||
1487 | /* we could use find_or_create_page(), but it locks page | |
1488 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1489 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1490 | if (page == NULL || !PageUptodate(page)) { |
1491 | if (page) | |
920313a7 AK |
1492 | /* |
1493 | * drop the page reference and try | |
1494 | * to get the page with lock. If we | |
1495 | * are not uptodate that implies | |
1496 | * somebody just created the page but | |
1497 | * is yet to initialize the same. So | |
1498 | * wait for it to initialize. | |
1499 | */ | |
09cbfeaf | 1500 | put_page(page); |
adb7ef60 | 1501 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d | 1502 | if (page) { |
19b8b035 TT |
1503 | if (WARN_RATELIMIT(page->mapping != inode->i_mapping, |
1504 | "ext4: bitmap's paging->mapping != inode->i_mapping\n")) { | |
1505 | /* should never happen */ | |
1506 | unlock_page(page); | |
1507 | ret = -EINVAL; | |
1508 | goto err; | |
1509 | } | |
c9de560d | 1510 | if (!PageUptodate(page)) { |
adb7ef60 | 1511 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1512 | if (ret) { |
1513 | unlock_page(page); | |
1514 | goto err; | |
1515 | } | |
c9de560d AT |
1516 | mb_cmp_bitmaps(e4b, page_address(page) + |
1517 | (poff * sb->s_blocksize)); | |
1518 | } | |
1519 | unlock_page(page); | |
1520 | } | |
1521 | } | |
c57ab39b YL |
1522 | if (page == NULL) { |
1523 | ret = -ENOMEM; | |
1524 | goto err; | |
1525 | } | |
1526 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1527 | ret = -EIO; |
c9de560d | 1528 | goto err; |
fdf6c7a7 | 1529 | } |
2457aec6 MG |
1530 | |
1531 | /* Pages marked accessed already */ | |
c9de560d AT |
1532 | e4b->bd_bitmap_page = page; |
1533 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1534 | |
1535 | block++; | |
1536 | pnum = block / blocks_per_page; | |
1537 | poff = block % blocks_per_page; | |
1538 | ||
2457aec6 | 1539 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1540 | if (page == NULL || !PageUptodate(page)) { |
1541 | if (page) | |
09cbfeaf | 1542 | put_page(page); |
adb7ef60 | 1543 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d | 1544 | if (page) { |
19b8b035 TT |
1545 | if (WARN_RATELIMIT(page->mapping != inode->i_mapping, |
1546 | "ext4: buddy bitmap's page->mapping != inode->i_mapping\n")) { | |
1547 | /* should never happen */ | |
1548 | unlock_page(page); | |
1549 | ret = -EINVAL; | |
1550 | goto err; | |
1551 | } | |
fdf6c7a7 | 1552 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1553 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1554 | gfp); | |
fdf6c7a7 SF |
1555 | if (ret) { |
1556 | unlock_page(page); | |
1557 | goto err; | |
1558 | } | |
1559 | } | |
c9de560d AT |
1560 | unlock_page(page); |
1561 | } | |
1562 | } | |
c57ab39b YL |
1563 | if (page == NULL) { |
1564 | ret = -ENOMEM; | |
1565 | goto err; | |
1566 | } | |
1567 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1568 | ret = -EIO; |
c9de560d | 1569 | goto err; |
fdf6c7a7 | 1570 | } |
2457aec6 MG |
1571 | |
1572 | /* Pages marked accessed already */ | |
c9de560d AT |
1573 | e4b->bd_buddy_page = page; |
1574 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d | 1575 | |
c9de560d AT |
1576 | return 0; |
1577 | ||
1578 | err: | |
26626f11 | 1579 | if (page) |
09cbfeaf | 1580 | put_page(page); |
c9de560d | 1581 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1582 | put_page(e4b->bd_bitmap_page); |
285164b8 | 1583 | |
c9de560d AT |
1584 | e4b->bd_buddy = NULL; |
1585 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1586 | return ret; |
c9de560d AT |
1587 | } |
1588 | ||
adb7ef60 KK |
1589 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1590 | struct ext4_buddy *e4b) | |
1591 | { | |
1592 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1593 | } | |
1594 | ||
e39e07fd | 1595 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1596 | { |
1597 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1598 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1599 | if (e4b->bd_buddy_page) |
09cbfeaf | 1600 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1601 | } |
1602 | ||
1603 | ||
1604 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1605 | { | |
ce3cca33 | 1606 | int order = 1, max; |
c9de560d AT |
1607 | void *bb; |
1608 | ||
c5e8f3f3 | 1609 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1610 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1611 | ||
c9de560d | 1612 | while (order <= e4b->bd_blkbits + 1) { |
ce3cca33 CX |
1613 | bb = mb_find_buddy(e4b, order, &max); |
1614 | if (!mb_test_bit(block >> order, bb)) { | |
c9de560d AT |
1615 | /* this block is part of buddy of order 'order' */ |
1616 | return order; | |
1617 | } | |
c9de560d AT |
1618 | order++; |
1619 | } | |
1620 | return 0; | |
1621 | } | |
1622 | ||
955ce5f5 | 1623 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1624 | { |
1625 | __u32 *addr; | |
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 | *addr = 0; | |
1633 | cur += 32; | |
1634 | continue; | |
1635 | } | |
955ce5f5 | 1636 | mb_clear_bit(cur, bm); |
c9de560d AT |
1637 | cur++; |
1638 | } | |
1639 | } | |
1640 | ||
eabe0444 AS |
1641 | /* clear bits in given range |
1642 | * will return first found zero bit if any, -1 otherwise | |
1643 | */ | |
1644 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1645 | { | |
1646 | __u32 *addr; | |
1647 | int zero_bit = -1; | |
1648 | ||
1649 | len = cur + len; | |
1650 | while (cur < len) { | |
1651 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1652 | /* fast path: clear whole word at once */ | |
1653 | addr = bm + (cur >> 3); | |
1654 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1655 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1656 | *addr = 0; | |
1657 | cur += 32; | |
1658 | continue; | |
1659 | } | |
1660 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1661 | zero_bit = cur; | |
1662 | cur++; | |
1663 | } | |
1664 | ||
1665 | return zero_bit; | |
1666 | } | |
1667 | ||
123e3016 | 1668 | void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1669 | { |
1670 | __u32 *addr; | |
1671 | ||
1672 | len = cur + len; | |
1673 | while (cur < len) { | |
1674 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1675 | /* fast path: set whole word at once */ | |
1676 | addr = bm + (cur >> 3); | |
1677 | *addr = 0xffffffff; | |
1678 | cur += 32; | |
1679 | continue; | |
1680 | } | |
955ce5f5 | 1681 | mb_set_bit(cur, bm); |
c9de560d AT |
1682 | cur++; |
1683 | } | |
1684 | } | |
1685 | ||
eabe0444 AS |
1686 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
1687 | { | |
1688 | if (mb_test_bit(*bit + side, bitmap)) { | |
1689 | mb_clear_bit(*bit, bitmap); | |
1690 | (*bit) -= side; | |
1691 | return 1; | |
1692 | } | |
1693 | else { | |
1694 | (*bit) += side; | |
1695 | mb_set_bit(*bit, bitmap); | |
1696 | return -1; | |
1697 | } | |
1698 | } | |
1699 | ||
1700 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1701 | { | |
1702 | int max; | |
1703 | int order = 1; | |
1704 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1705 | ||
1706 | while (buddy) { | |
1707 | void *buddy2; | |
1708 | ||
1709 | /* Bits in range [first; last] are known to be set since | |
1710 | * corresponding blocks were allocated. Bits in range | |
1711 | * (first; last) will stay set because they form buddies on | |
1712 | * upper layer. We just deal with borders if they don't | |
1713 | * align with upper layer and then go up. | |
1714 | * Releasing entire group is all about clearing | |
1715 | * single bit of highest order buddy. | |
1716 | */ | |
1717 | ||
1718 | /* Example: | |
1719 | * --------------------------------- | |
1720 | * | 1 | 1 | 1 | 1 | | |
1721 | * --------------------------------- | |
1722 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1723 | * --------------------------------- | |
1724 | * 0 1 2 3 4 5 6 7 | |
1725 | * \_____________________/ | |
1726 | * | |
1727 | * Neither [1] nor [6] is aligned to above layer. | |
1728 | * Left neighbour [0] is free, so mark it busy, | |
1729 | * decrease bb_counters and extend range to | |
1730 | * [0; 6] | |
1731 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1732 | * mark [6] free, increase bb_counters and shrink range to | |
1733 | * [0; 5]. | |
1734 | * Then shift range to [0; 2], go up and do the same. | |
1735 | */ | |
1736 | ||
1737 | ||
1738 | if (first & 1) | |
1739 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1740 | if (!(last & 1)) | |
1741 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1742 | if (first > last) | |
1743 | break; | |
1744 | order++; | |
1745 | ||
976620bd KS |
1746 | buddy2 = mb_find_buddy(e4b, order, &max); |
1747 | if (!buddy2) { | |
eabe0444 AS |
1748 | mb_clear_bits(buddy, first, last - first + 1); |
1749 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1750 | break; | |
1751 | } | |
1752 | first >>= 1; | |
1753 | last >>= 1; | |
1754 | buddy = buddy2; | |
1755 | } | |
1756 | } | |
1757 | ||
7e5a8cdd | 1758 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1759 | int first, int count) |
c9de560d | 1760 | { |
eabe0444 AS |
1761 | int left_is_free = 0; |
1762 | int right_is_free = 0; | |
1763 | int block; | |
1764 | int last = first + count - 1; | |
c9de560d AT |
1765 | struct super_block *sb = e4b->bd_sb; |
1766 | ||
c99d1e6e TT |
1767 | if (WARN_ON(count == 0)) |
1768 | return; | |
eabe0444 | 1769 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1770 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1771 | /* Don't bother if the block group is corrupt. */ |
1772 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1773 | return; | |
1774 | ||
c9de560d AT |
1775 | mb_check_buddy(e4b); |
1776 | mb_free_blocks_double(inode, e4b, first, count); | |
1777 | ||
07b5b8e1 | 1778 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1779 | e4b->bd_info->bb_free += count; |
1780 | if (first < e4b->bd_info->bb_first_free) | |
1781 | e4b->bd_info->bb_first_free = first; | |
1782 | ||
eabe0444 AS |
1783 | /* access memory sequentially: check left neighbour, |
1784 | * clear range and then check right neighbour | |
1785 | */ | |
c9de560d | 1786 | if (first != 0) |
eabe0444 AS |
1787 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1788 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1789 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1790 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1791 | ||
1792 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1793 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1794 | ext4_fsblk_t blocknr; |
1795 | ||
1796 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1797 | blocknr += EXT4_C2B(sbi, block); |
8016e29f HS |
1798 | if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { |
1799 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1800 | inode ? inode->i_ino : 0, | |
1801 | blocknr, | |
1802 | "freeing already freed block (bit %u); block bitmap corrupt.", | |
1803 | block); | |
1804 | ext4_mark_group_bitmap_corrupted( | |
1805 | sb, e4b->bd_group, | |
db79e6d1 | 1806 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
8016e29f | 1807 | } |
eabe0444 AS |
1808 | goto done; |
1809 | } | |
1810 | ||
1811 | /* let's maintain fragments counter */ | |
1812 | if (left_is_free && right_is_free) | |
c9de560d | 1813 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1814 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1815 | e4b->bd_info->bb_fragments++; |
1816 | ||
eabe0444 AS |
1817 | /* buddy[0] == bd_bitmap is a special case, so handle |
1818 | * it right away and let mb_buddy_mark_free stay free of | |
1819 | * zero order checks. | |
1820 | * Check if neighbours are to be coaleasced, | |
1821 | * adjust bitmap bb_counters and borders appropriately. | |
1822 | */ | |
1823 | if (first & 1) { | |
1824 | first += !left_is_free; | |
1825 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1826 | } | |
1827 | if (!(last & 1)) { | |
1828 | last -= !right_is_free; | |
1829 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1830 | } | |
c9de560d | 1831 | |
eabe0444 AS |
1832 | if (first <= last) |
1833 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1834 | |
eabe0444 | 1835 | done: |
8a57d9d6 | 1836 | mb_set_largest_free_order(sb, e4b->bd_info); |
196e402a | 1837 | mb_update_avg_fragment_size(sb, e4b->bd_info); |
c9de560d | 1838 | mb_check_buddy(e4b); |
c9de560d AT |
1839 | } |
1840 | ||
15c006a2 | 1841 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1842 | int needed, struct ext4_free_extent *ex) |
1843 | { | |
1844 | int next = block; | |
15c006a2 | 1845 | int max, order; |
c9de560d AT |
1846 | void *buddy; |
1847 | ||
bc8e6740 | 1848 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1849 | BUG_ON(ex == NULL); |
1850 | ||
15c006a2 | 1851 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1852 | BUG_ON(buddy == NULL); |
1853 | BUG_ON(block >= max); | |
1854 | if (mb_test_bit(block, buddy)) { | |
1855 | ex->fe_len = 0; | |
1856 | ex->fe_start = 0; | |
1857 | ex->fe_group = 0; | |
1858 | return 0; | |
1859 | } | |
1860 | ||
15c006a2 RD |
1861 | /* find actual order */ |
1862 | order = mb_find_order_for_block(e4b, block); | |
1863 | block = block >> order; | |
c9de560d AT |
1864 | |
1865 | ex->fe_len = 1 << order; | |
1866 | ex->fe_start = block << order; | |
1867 | ex->fe_group = e4b->bd_group; | |
1868 | ||
1869 | /* calc difference from given start */ | |
1870 | next = next - ex->fe_start; | |
1871 | ex->fe_len -= next; | |
1872 | ex->fe_start += next; | |
1873 | ||
1874 | while (needed > ex->fe_len && | |
d8ec0c39 | 1875 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1876 | |
1877 | if (block + 1 >= max) | |
1878 | break; | |
1879 | ||
1880 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1881 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1882 | break; |
1883 | ||
b051d8dc | 1884 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1885 | |
c9de560d AT |
1886 | block = next >> order; |
1887 | ex->fe_len += 1 << order; | |
1888 | } | |
1889 | ||
31562b95 | 1890 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1891 | /* Should never happen! (but apparently sometimes does?!?) */ |
1892 | WARN_ON(1); | |
cd84bbba SB |
1893 | ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, |
1894 | "corruption or bug in mb_find_extent " | |
1895 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1896 | block, order, needed, ex->fe_group, ex->fe_start, | |
1897 | ex->fe_len, ex->fe_logical); | |
43c73221 TT |
1898 | ex->fe_len = 0; |
1899 | ex->fe_start = 0; | |
1900 | ex->fe_group = 0; | |
1901 | } | |
c9de560d AT |
1902 | return ex->fe_len; |
1903 | } | |
1904 | ||
1905 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1906 | { | |
1907 | int ord; | |
1908 | int mlen = 0; | |
1909 | int max = 0; | |
1910 | int cur; | |
1911 | int start = ex->fe_start; | |
1912 | int len = ex->fe_len; | |
1913 | unsigned ret = 0; | |
1914 | int len0 = len; | |
1915 | void *buddy; | |
218a6944 | 1916 | bool split = false; |
c9de560d AT |
1917 | |
1918 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1919 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1920 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1921 | mb_check_buddy(e4b); |
1922 | mb_mark_used_double(e4b, start, len); | |
1923 | ||
07b5b8e1 | 1924 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1925 | e4b->bd_info->bb_free -= len; |
1926 | if (e4b->bd_info->bb_first_free == start) | |
1927 | e4b->bd_info->bb_first_free += len; | |
1928 | ||
1929 | /* let's maintain fragments counter */ | |
1930 | if (start != 0) | |
c5e8f3f3 | 1931 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1932 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1933 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1934 | if (mlen && max) |
1935 | e4b->bd_info->bb_fragments++; | |
1936 | else if (!mlen && !max) | |
1937 | e4b->bd_info->bb_fragments--; | |
1938 | ||
1939 | /* let's maintain buddy itself */ | |
1940 | while (len) { | |
218a6944 | 1941 | if (!split) |
1942 | ord = mb_find_order_for_block(e4b, start); | |
c9de560d AT |
1943 | |
1944 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1945 | /* the whole chunk may be allocated at once! */ | |
1946 | mlen = 1 << ord; | |
218a6944 | 1947 | if (!split) |
1948 | buddy = mb_find_buddy(e4b, ord, &max); | |
1949 | else | |
1950 | split = false; | |
c9de560d AT |
1951 | BUG_ON((start >> ord) >= max); |
1952 | mb_set_bit(start >> ord, buddy); | |
1953 | e4b->bd_info->bb_counters[ord]--; | |
1954 | start += mlen; | |
1955 | len -= mlen; | |
1956 | BUG_ON(len < 0); | |
1957 | continue; | |
1958 | } | |
1959 | ||
1960 | /* store for history */ | |
1961 | if (ret == 0) | |
1962 | ret = len | (ord << 16); | |
1963 | ||
1964 | /* we have to split large buddy */ | |
1965 | BUG_ON(ord <= 0); | |
1966 | buddy = mb_find_buddy(e4b, ord, &max); | |
1967 | mb_set_bit(start >> ord, buddy); | |
1968 | e4b->bd_info->bb_counters[ord]--; | |
1969 | ||
1970 | ord--; | |
1971 | cur = (start >> ord) & ~1U; | |
1972 | buddy = mb_find_buddy(e4b, ord, &max); | |
1973 | mb_clear_bit(cur, buddy); | |
1974 | mb_clear_bit(cur + 1, buddy); | |
1975 | e4b->bd_info->bb_counters[ord]++; | |
1976 | e4b->bd_info->bb_counters[ord]++; | |
218a6944 | 1977 | split = true; |
c9de560d | 1978 | } |
8a57d9d6 | 1979 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1980 | |
196e402a | 1981 | mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); |
123e3016 | 1982 | mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1983 | mb_check_buddy(e4b); |
1984 | ||
1985 | return ret; | |
1986 | } | |
1987 | ||
1988 | /* | |
1989 | * Must be called under group lock! | |
1990 | */ | |
1991 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1992 | struct ext4_buddy *e4b) | |
1993 | { | |
1994 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1995 | int ret; | |
1996 | ||
1997 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1998 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1999 | ||
2000 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
2001 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
2002 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
2003 | ||
2004 | /* preallocation can change ac_b_ex, thus we store actually | |
2005 | * allocated blocks for history */ | |
2006 | ac->ac_f_ex = ac->ac_b_ex; | |
2007 | ||
2008 | ac->ac_status = AC_STATUS_FOUND; | |
2009 | ac->ac_tail = ret & 0xffff; | |
2010 | ac->ac_buddy = ret >> 16; | |
2011 | ||
c3a326a6 AK |
2012 | /* |
2013 | * take the page reference. We want the page to be pinned | |
2014 | * so that we don't get a ext4_mb_init_cache_call for this | |
2015 | * group until we update the bitmap. That would mean we | |
2016 | * double allocate blocks. The reference is dropped | |
2017 | * in ext4_mb_release_context | |
2018 | */ | |
c9de560d AT |
2019 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
2020 | get_page(ac->ac_bitmap_page); | |
2021 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
2022 | get_page(ac->ac_buddy_page); | |
c9de560d | 2023 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 2024 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
2025 | spin_lock(&sbi->s_md_lock); |
2026 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
2027 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
2028 | spin_unlock(&sbi->s_md_lock); | |
2029 | } | |
53f86b17 RH |
2030 | /* |
2031 | * As we've just preallocated more space than | |
2032 | * user requested originally, we store allocated | |
2033 | * space in a special descriptor. | |
2034 | */ | |
2035 | if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
2036 | ext4_mb_new_preallocation(ac); | |
2037 | ||
c9de560d AT |
2038 | } |
2039 | ||
c9de560d AT |
2040 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
2041 | struct ext4_buddy *e4b, | |
2042 | int finish_group) | |
2043 | { | |
2044 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2045 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2046 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
c9de560d | 2047 | |
032115fc AK |
2048 | if (ac->ac_status == AC_STATUS_FOUND) |
2049 | return; | |
c9de560d AT |
2050 | /* |
2051 | * We don't want to scan for a whole year | |
2052 | */ | |
2053 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
2054 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2055 | ac->ac_status = AC_STATUS_BREAK; | |
2056 | return; | |
2057 | } | |
2058 | ||
2059 | /* | |
2060 | * Haven't found good chunk so far, let's continue | |
2061 | */ | |
2062 | if (bex->fe_len < gex->fe_len) | |
2063 | return; | |
2064 | ||
78dc9f84 KS |
2065 | if (finish_group) |
2066 | ext4_mb_use_best_found(ac, e4b); | |
c9de560d AT |
2067 | } |
2068 | ||
2069 | /* | |
2070 | * The routine checks whether found extent is good enough. If it is, | |
2071 | * then the extent gets marked used and flag is set to the context | |
2072 | * to stop scanning. Otherwise, the extent is compared with the | |
2073 | * previous found extent and if new one is better, then it's stored | |
2074 | * in the context. Later, the best found extent will be used, if | |
2075 | * mballoc can't find good enough extent. | |
2076 | * | |
2077 | * FIXME: real allocation policy is to be designed yet! | |
2078 | */ | |
2079 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
2080 | struct ext4_free_extent *ex, | |
2081 | struct ext4_buddy *e4b) | |
2082 | { | |
2083 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2084 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2085 | ||
2086 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
2087 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
2088 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
2089 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
2090 | ||
2091 | ac->ac_found++; | |
2092 | ||
2093 | /* | |
2094 | * The special case - take what you catch first | |
2095 | */ | |
2096 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2097 | *bex = *ex; | |
2098 | ext4_mb_use_best_found(ac, e4b); | |
2099 | return; | |
2100 | } | |
2101 | ||
2102 | /* | |
2103 | * Let's check whether the chuck is good enough | |
2104 | */ | |
2105 | if (ex->fe_len == gex->fe_len) { | |
2106 | *bex = *ex; | |
2107 | ext4_mb_use_best_found(ac, e4b); | |
2108 | return; | |
2109 | } | |
2110 | ||
2111 | /* | |
2112 | * If this is first found extent, just store it in the context | |
2113 | */ | |
2114 | if (bex->fe_len == 0) { | |
2115 | *bex = *ex; | |
2116 | return; | |
2117 | } | |
2118 | ||
2119 | /* | |
2120 | * If new found extent is better, store it in the context | |
2121 | */ | |
2122 | if (bex->fe_len < gex->fe_len) { | |
2123 | /* if the request isn't satisfied, any found extent | |
2124 | * larger than previous best one is better */ | |
2125 | if (ex->fe_len > bex->fe_len) | |
2126 | *bex = *ex; | |
2127 | } else if (ex->fe_len > gex->fe_len) { | |
2128 | /* if the request is satisfied, then we try to find | |
2129 | * an extent that still satisfy the request, but is | |
2130 | * smaller than previous one */ | |
2131 | if (ex->fe_len < bex->fe_len) | |
2132 | *bex = *ex; | |
2133 | } | |
2134 | ||
2135 | ext4_mb_check_limits(ac, e4b, 0); | |
2136 | } | |
2137 | ||
089ceecc | 2138 | static noinline_for_stack |
85b67ffb | 2139 | void ext4_mb_try_best_found(struct ext4_allocation_context *ac, |
c9de560d AT |
2140 | struct ext4_buddy *e4b) |
2141 | { | |
2142 | struct ext4_free_extent ex = ac->ac_b_ex; | |
2143 | ext4_group_t group = ex.fe_group; | |
2144 | int max; | |
2145 | int err; | |
2146 | ||
2147 | BUG_ON(ex.fe_len <= 0); | |
2148 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2149 | if (err) | |
85b67ffb | 2150 | return; |
c9de560d AT |
2151 | |
2152 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 2153 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
2154 | |
2155 | if (max > 0) { | |
2156 | ac->ac_b_ex = ex; | |
2157 | ext4_mb_use_best_found(ac, e4b); | |
2158 | } | |
2159 | ||
2160 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2161 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2162 | } |
2163 | ||
089ceecc ES |
2164 | static noinline_for_stack |
2165 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
2166 | struct ext4_buddy *e4b) |
2167 | { | |
2168 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
2169 | int max; | |
2170 | int err; | |
2171 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 2172 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
2173 | struct ext4_free_extent ex; |
2174 | ||
5354b2af TT |
2175 | if (!grp) |
2176 | return -EFSCORRUPTED; | |
01e4ca29 | 2177 | if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY))) |
c9de560d | 2178 | return 0; |
838cd0cf YY |
2179 | if (grp->bb_free == 0) |
2180 | return 0; | |
c9de560d AT |
2181 | |
2182 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2183 | if (err) | |
2184 | return err; | |
2185 | ||
163a203d DW |
2186 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
2187 | ext4_mb_unload_buddy(e4b); | |
2188 | return 0; | |
2189 | } | |
2190 | ||
c9de560d | 2191 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 2192 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 2193 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 2194 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
2195 | |
2196 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
2197 | ext4_fsblk_t start; | |
2198 | ||
5661bd68 AM |
2199 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
2200 | ex.fe_start; | |
c9de560d AT |
2201 | /* use do_div to get remainder (would be 64-bit modulo) */ |
2202 | if (do_div(start, sbi->s_stripe) == 0) { | |
2203 | ac->ac_found++; | |
2204 | ac->ac_b_ex = ex; | |
2205 | ext4_mb_use_best_found(ac, e4b); | |
2206 | } | |
2207 | } else if (max >= ac->ac_g_ex.fe_len) { | |
2208 | BUG_ON(ex.fe_len <= 0); | |
2209 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2210 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2211 | ac->ac_found++; | |
2212 | ac->ac_b_ex = ex; | |
2213 | ext4_mb_use_best_found(ac, e4b); | |
2214 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
2215 | /* Sometimes, caller may want to merge even small | |
2216 | * number of blocks to an existing extent */ | |
2217 | BUG_ON(ex.fe_len <= 0); | |
2218 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2219 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2220 | ac->ac_found++; | |
2221 | ac->ac_b_ex = ex; | |
2222 | ext4_mb_use_best_found(ac, e4b); | |
2223 | } | |
2224 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2225 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2226 | |
2227 | return 0; | |
2228 | } | |
2229 | ||
2230 | /* | |
2231 | * The routine scans buddy structures (not bitmap!) from given order | |
2232 | * to max order and tries to find big enough chunk to satisfy the req | |
2233 | */ | |
089ceecc ES |
2234 | static noinline_for_stack |
2235 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2236 | struct ext4_buddy *e4b) |
2237 | { | |
2238 | struct super_block *sb = ac->ac_sb; | |
2239 | struct ext4_group_info *grp = e4b->bd_info; | |
2240 | void *buddy; | |
2241 | int i; | |
2242 | int k; | |
2243 | int max; | |
2244 | ||
2245 | BUG_ON(ac->ac_2order <= 0); | |
4b68f6df | 2246 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { |
c9de560d AT |
2247 | if (grp->bb_counters[i] == 0) |
2248 | continue; | |
2249 | ||
2250 | buddy = mb_find_buddy(e4b, i, &max); | |
19b8b035 TT |
2251 | if (WARN_RATELIMIT(buddy == NULL, |
2252 | "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i)) | |
2253 | continue; | |
c9de560d | 2254 | |
ffad0a44 | 2255 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
2256 | if (k >= max) { |
2257 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
2258 | "%d free clusters of order %d. But found 0", | |
2259 | grp->bb_counters[i], i); | |
2260 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
2261 | e4b->bd_group, | |
2262 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
2263 | break; | |
2264 | } | |
c9de560d AT |
2265 | ac->ac_found++; |
2266 | ||
2267 | ac->ac_b_ex.fe_len = 1 << i; | |
2268 | ac->ac_b_ex.fe_start = k << i; | |
2269 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
2270 | ||
2271 | ext4_mb_use_best_found(ac, e4b); | |
2272 | ||
53f86b17 | 2273 | BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
c9de560d AT |
2274 | |
2275 | if (EXT4_SB(sb)->s_mb_stats) | |
2276 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
2277 | ||
2278 | break; | |
2279 | } | |
2280 | } | |
2281 | ||
2282 | /* | |
2283 | * The routine scans the group and measures all found extents. | |
2284 | * In order to optimize scanning, caller must pass number of | |
2285 | * free blocks in the group, so the routine can know upper limit. | |
2286 | */ | |
089ceecc ES |
2287 | static noinline_for_stack |
2288 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2289 | struct ext4_buddy *e4b) |
2290 | { | |
2291 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 2292 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2293 | struct ext4_free_extent ex; |
2294 | int i; | |
2295 | int free; | |
2296 | ||
2297 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
2298 | if (WARN_ON(free <= 0)) |
2299 | return; | |
c9de560d AT |
2300 | |
2301 | i = e4b->bd_info->bb_first_free; | |
2302 | ||
2303 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 2304 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
2305 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
2306 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 2307 | /* |
e56eb659 | 2308 | * IF we have corrupt bitmap, we won't find any |
26346ff6 | 2309 | * free blocks even though group info says we |
b483bb77 | 2310 | * have free blocks |
26346ff6 | 2311 | */ |
e29136f8 | 2312 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2313 | "%d free clusters as per " |
fde4d95a | 2314 | "group info. But bitmap says 0", |
26346ff6 | 2315 | free); |
736dedbb WS |
2316 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2317 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
2318 | break; |
2319 | } | |
2320 | ||
15c006a2 | 2321 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2322 | if (WARN_ON(ex.fe_len <= 0)) |
2323 | break; | |
26346ff6 | 2324 | if (free < ex.fe_len) { |
e29136f8 | 2325 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2326 | "%d free clusters as per " |
fde4d95a | 2327 | "group info. But got %d blocks", |
26346ff6 | 2328 | free, ex.fe_len); |
736dedbb WS |
2329 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2330 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
2331 | /* |
2332 | * The number of free blocks differs. This mostly | |
2333 | * indicate that the bitmap is corrupt. So exit | |
2334 | * without claiming the space. | |
2335 | */ | |
2336 | break; | |
26346ff6 | 2337 | } |
ab0c00fc | 2338 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2339 | ext4_mb_measure_extent(ac, &ex, e4b); |
2340 | ||
2341 | i += ex.fe_len; | |
2342 | free -= ex.fe_len; | |
2343 | } | |
2344 | ||
2345 | ext4_mb_check_limits(ac, e4b, 1); | |
2346 | } | |
2347 | ||
2348 | /* | |
2349 | * This is a special case for storages like raid5 | |
506bf2d8 | 2350 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2351 | */ |
089ceecc ES |
2352 | static noinline_for_stack |
2353 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2354 | struct ext4_buddy *e4b) |
2355 | { | |
2356 | struct super_block *sb = ac->ac_sb; | |
2357 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2358 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2359 | struct ext4_free_extent ex; |
2360 | ext4_fsblk_t first_group_block; | |
2361 | ext4_fsblk_t a; | |
2362 | ext4_grpblk_t i; | |
2363 | int max; | |
2364 | ||
2365 | BUG_ON(sbi->s_stripe == 0); | |
2366 | ||
2367 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2368 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2369 | ||
c9de560d AT |
2370 | a = first_group_block + sbi->s_stripe - 1; |
2371 | do_div(a, sbi->s_stripe); | |
2372 | i = (a * sbi->s_stripe) - first_group_block; | |
2373 | ||
7137d7a4 | 2374 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2375 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2376 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2377 | if (max >= sbi->s_stripe) { |
2378 | ac->ac_found++; | |
ab0c00fc | 2379 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2380 | ac->ac_b_ex = ex; |
2381 | ext4_mb_use_best_found(ac, e4b); | |
2382 | break; | |
2383 | } | |
2384 | } | |
2385 | i += sbi->s_stripe; | |
2386 | } | |
2387 | } | |
2388 | ||
42ac1848 | 2389 | /* |
8ef123fe | 2390 | * This is also called BEFORE we load the buddy bitmap. |
42ac1848 | 2391 | * Returns either 1 or 0 indicating that the group is either suitable |
8ef123fe | 2392 | * for the allocation or not. |
42ac1848 | 2393 | */ |
8ef123fe | 2394 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
c9de560d AT |
2395 | ext4_group_t group, int cr) |
2396 | { | |
8ef123fe | 2397 | ext4_grpblk_t free, fragments; |
a4912123 | 2398 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2399 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2400 | ||
2401 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2402 | |
5354b2af | 2403 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp) || !grp)) |
8ef123fe | 2404 | return false; |
01fc48e8 | 2405 | |
dddcd2f9 | 2406 | free = grp->bb_free; |
2407 | if (free == 0) | |
8ef123fe | 2408 | return false; |
c9de560d | 2409 | |
c9de560d | 2410 | fragments = grp->bb_fragments; |
c9de560d | 2411 | if (fragments == 0) |
8ef123fe | 2412 | return false; |
c9de560d AT |
2413 | |
2414 | switch (cr) { | |
2415 | case 0: | |
2416 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2417 | |
a4912123 TT |
2418 | /* Avoid using the first bg of a flexgroup for data files */ |
2419 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2420 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2421 | ((group % flex_size) == 0)) | |
8ef123fe | 2422 | return false; |
a4912123 | 2423 | |
dddcd2f9 | 2424 | if (free < ac->ac_g_ex.fe_len) |
2425 | return false; | |
2426 | ||
4b68f6df | 2427 | if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) |
8ef123fe | 2428 | return true; |
40ae3487 TT |
2429 | |
2430 | if (grp->bb_largest_free_order < ac->ac_2order) | |
8ef123fe | 2431 | return false; |
40ae3487 | 2432 | |
8ef123fe | 2433 | return true; |
c9de560d AT |
2434 | case 1: |
2435 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2436 | return true; |
c9de560d AT |
2437 | break; |
2438 | case 2: | |
2439 | if (free >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2440 | return true; |
c9de560d AT |
2441 | break; |
2442 | case 3: | |
8ef123fe | 2443 | return true; |
c9de560d AT |
2444 | default: |
2445 | BUG(); | |
2446 | } | |
2447 | ||
8ef123fe RH |
2448 | return false; |
2449 | } | |
2450 | ||
2451 | /* | |
2452 | * This could return negative error code if something goes wrong | |
2453 | * during ext4_mb_init_group(). This should not be called with | |
2454 | * ext4_lock_group() held. | |
a5fda113 TT |
2455 | * |
2456 | * Note: because we are conditionally operating with the group lock in | |
2457 | * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this | |
2458 | * function using __acquire and __release. This means we need to be | |
2459 | * super careful before messing with the error path handling via "goto | |
2460 | * out"! | |
8ef123fe RH |
2461 | */ |
2462 | static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, | |
2463 | ext4_group_t group, int cr) | |
2464 | { | |
2465 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
99377830 | 2466 | struct super_block *sb = ac->ac_sb; |
c1d2c7d4 | 2467 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
99377830 | 2468 | bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
8ef123fe RH |
2469 | ext4_grpblk_t free; |
2470 | int ret = 0; | |
2471 | ||
5354b2af TT |
2472 | if (!grp) |
2473 | return -EFSCORRUPTED; | |
a6c75eaf HS |
2474 | if (sbi->s_mb_stats) |
2475 | atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); | |
a5fda113 | 2476 | if (should_lock) { |
99377830 | 2477 | ext4_lock_group(sb, group); |
a5fda113 TT |
2478 | __release(ext4_group_lock_ptr(sb, group)); |
2479 | } | |
8ef123fe RH |
2480 | free = grp->bb_free; |
2481 | if (free == 0) | |
2482 | goto out; | |
2483 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2484 | goto out; | |
2485 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
2486 | goto out; | |
a5fda113 TT |
2487 | if (should_lock) { |
2488 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2489 | ext4_unlock_group(sb, group); |
a5fda113 | 2490 | } |
8ef123fe RH |
2491 | |
2492 | /* We only do this if the grp has never been initialized */ | |
2493 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
c1d2c7d4 AZ |
2494 | struct ext4_group_desc *gdp = |
2495 | ext4_get_group_desc(sb, group, NULL); | |
2496 | int ret; | |
2497 | ||
2498 | /* cr=0/1 is a very optimistic search to find large | |
2499 | * good chunks almost for free. If buddy data is not | |
2500 | * ready, then this optimization makes no sense. But | |
2501 | * we never skip the first block group in a flex_bg, | |
2502 | * since this gets used for metadata block allocation, | |
2503 | * and we want to make sure we locate metadata blocks | |
2504 | * in the first block group in the flex_bg if possible. | |
2505 | */ | |
2506 | if (cr < 2 && | |
2507 | (!sbi->s_log_groups_per_flex || | |
2508 | ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && | |
2509 | !(ext4_has_group_desc_csum(sb) && | |
2510 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) | |
2511 | return 0; | |
2512 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); | |
8ef123fe RH |
2513 | if (ret) |
2514 | return ret; | |
2515 | } | |
2516 | ||
a5fda113 | 2517 | if (should_lock) { |
99377830 | 2518 | ext4_lock_group(sb, group); |
a5fda113 TT |
2519 | __release(ext4_group_lock_ptr(sb, group)); |
2520 | } | |
8ef123fe RH |
2521 | ret = ext4_mb_good_group(ac, group, cr); |
2522 | out: | |
a5fda113 TT |
2523 | if (should_lock) { |
2524 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2525 | ext4_unlock_group(sb, group); |
a5fda113 | 2526 | } |
8ef123fe | 2527 | return ret; |
c9de560d AT |
2528 | } |
2529 | ||
cfd73237 AZ |
2530 | /* |
2531 | * Start prefetching @nr block bitmaps starting at @group. | |
2532 | * Return the next group which needs to be prefetched. | |
2533 | */ | |
3d392b26 TT |
2534 | ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
2535 | unsigned int nr, int *cnt) | |
cfd73237 AZ |
2536 | { |
2537 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2538 | struct buffer_head *bh; | |
2539 | struct blk_plug plug; | |
2540 | ||
2541 | blk_start_plug(&plug); | |
2542 | while (nr-- > 0) { | |
2543 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2544 | NULL); | |
2545 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2546 | ||
2547 | /* | |
2548 | * Prefetch block groups with free blocks; but don't | |
2549 | * bother if it is marked uninitialized on disk, since | |
2550 | * it won't require I/O to read. Also only try to | |
2551 | * prefetch once, so we avoid getblk() call, which can | |
2552 | * be expensive. | |
2553 | */ | |
5354b2af | 2554 | if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) && |
cfd73237 AZ |
2555 | EXT4_MB_GRP_NEED_INIT(grp) && |
2556 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2557 | !(ext4_has_group_desc_csum(sb) && | |
2558 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2559 | bh = ext4_read_block_bitmap_nowait(sb, group, true); | |
2560 | if (bh && !IS_ERR(bh)) { | |
2561 | if (!buffer_uptodate(bh) && cnt) | |
2562 | (*cnt)++; | |
2563 | brelse(bh); | |
2564 | } | |
2565 | } | |
2566 | if (++group >= ngroups) | |
2567 | group = 0; | |
2568 | } | |
2569 | blk_finish_plug(&plug); | |
2570 | return group; | |
2571 | } | |
2572 | ||
2573 | /* | |
2574 | * Prefetching reads the block bitmap into the buffer cache; but we | |
2575 | * need to make sure that the buddy bitmap in the page cache has been | |
2576 | * initialized. Note that ext4_mb_init_group() will block if the I/O | |
2577 | * is not yet completed, or indeed if it was not initiated by | |
2578 | * ext4_mb_prefetch did not start the I/O. | |
2579 | * | |
2580 | * TODO: We should actually kick off the buddy bitmap setup in a work | |
2581 | * queue when the buffer I/O is completed, so that we don't block | |
2582 | * waiting for the block allocation bitmap read to finish when | |
2583 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). | |
2584 | */ | |
3d392b26 TT |
2585 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
2586 | unsigned int nr) | |
cfd73237 | 2587 | { |
22fab984 KS |
2588 | struct ext4_group_desc *gdp; |
2589 | struct ext4_group_info *grp; | |
cfd73237 | 2590 | |
22fab984 | 2591 | while (nr-- > 0) { |
cfd73237 AZ |
2592 | if (!group) |
2593 | group = ext4_get_groups_count(sb); | |
2594 | group--; | |
22fab984 | 2595 | gdp = ext4_get_group_desc(sb, group, NULL); |
cfd73237 AZ |
2596 | grp = ext4_get_group_info(sb, group); |
2597 | ||
5354b2af | 2598 | if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) && |
cfd73237 AZ |
2599 | ext4_free_group_clusters(sb, gdp) > 0 && |
2600 | !(ext4_has_group_desc_csum(sb) && | |
2601 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2602 | if (ext4_mb_init_group(sb, group, GFP_NOFS)) | |
2603 | break; | |
2604 | } | |
2605 | } | |
2606 | } | |
2607 | ||
4ddfef7b ES |
2608 | static noinline_for_stack int |
2609 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2610 | { |
cfd73237 | 2611 | ext4_group_t prefetch_grp = 0, ngroups, group, i; |
4fca50d4 | 2612 | int cr = -1, new_cr; |
42ac1848 | 2613 | int err = 0, first_err = 0; |
cfd73237 | 2614 | unsigned int nr = 0, prefetch_ios = 0; |
c9de560d AT |
2615 | struct ext4_sb_info *sbi; |
2616 | struct super_block *sb; | |
2617 | struct ext4_buddy e4b; | |
66d5e027 | 2618 | int lost; |
c9de560d AT |
2619 | |
2620 | sb = ac->ac_sb; | |
2621 | sbi = EXT4_SB(sb); | |
8df9675f | 2622 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2623 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2624 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2625 | ngroups = sbi->s_blockfile_groups; |
2626 | ||
c9de560d AT |
2627 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2628 | ||
2629 | /* first, try the goal */ | |
2630 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2631 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2632 | goto out; | |
2633 | ||
2634 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2635 | goto out; | |
2636 | ||
2637 | /* | |
e9a3cd48 | 2638 | * ac->ac_2order is set only if the fe_len is a power of 2 |
2639 | * if ac->ac_2order is set we also set criteria to 0 so that we | |
c9de560d AT |
2640 | * try exact allocation using buddy. |
2641 | */ | |
2642 | i = fls(ac->ac_g_ex.fe_len); | |
2643 | ac->ac_2order = 0; | |
2644 | /* | |
2645 | * We search using buddy data only if the order of the request | |
2646 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2647 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2648 | * We also support searching for power-of-two requests only for |
2649 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2650 | */ |
4b68f6df | 2651 | if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { |
c9de560d AT |
2652 | /* |
2653 | * This should tell if fe_len is exactly power of 2 | |
2654 | */ | |
2655 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d | 2656 | ac->ac_2order = array_index_nospec(i - 1, |
4b68f6df | 2657 | MB_NUM_ORDERS(sb)); |
c9de560d AT |
2658 | } |
2659 | ||
4ba74d00 TT |
2660 | /* if stream allocation is enabled, use global goal */ |
2661 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2662 | /* TBD: may be hot point */ |
2663 | spin_lock(&sbi->s_md_lock); | |
2664 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2665 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2666 | spin_unlock(&sbi->s_md_lock); | |
2667 | } | |
4ba74d00 | 2668 | |
c9de560d AT |
2669 | /* Let's just scan groups to find more-less suitable blocks */ |
2670 | cr = ac->ac_2order ? 0 : 1; | |
2671 | /* | |
2672 | * cr == 0 try to get exact allocation, | |
2673 | * cr == 3 try to get anything | |
2674 | */ | |
2675 | repeat: | |
2676 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2677 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2678 | /* |
2679 | * searching for the right group start | |
2680 | * from the goal value specified | |
2681 | */ | |
2682 | group = ac->ac_g_ex.fe_group; | |
196e402a | 2683 | ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; |
cfd73237 | 2684 | prefetch_grp = group; |
ed8f9c75 | 2685 | |
4fca50d4 JK |
2686 | for (i = 0, new_cr = cr; i < ngroups; i++, |
2687 | ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { | |
2688 | int ret = 0; | |
196e402a | 2689 | |
2ed5724d | 2690 | cond_resched(); |
196e402a HS |
2691 | if (new_cr != cr) { |
2692 | cr = new_cr; | |
2693 | goto repeat; | |
2694 | } | |
c9de560d | 2695 | |
cfd73237 AZ |
2696 | /* |
2697 | * Batch reads of the block allocation bitmaps | |
2698 | * to get multiple READs in flight; limit | |
2699 | * prefetching at cr=0/1, otherwise mballoc can | |
2700 | * spend a lot of time loading imperfect groups | |
2701 | */ | |
2702 | if ((prefetch_grp == group) && | |
2703 | (cr > 1 || | |
2704 | prefetch_ios < sbi->s_mb_prefetch_limit)) { | |
2705 | unsigned int curr_ios = prefetch_ios; | |
2706 | ||
2707 | nr = sbi->s_mb_prefetch; | |
2708 | if (ext4_has_feature_flex_bg(sb)) { | |
82ef1370 CX |
2709 | nr = 1 << sbi->s_log_groups_per_flex; |
2710 | nr -= group & (nr - 1); | |
2711 | nr = min(nr, sbi->s_mb_prefetch); | |
cfd73237 AZ |
2712 | } |
2713 | prefetch_grp = ext4_mb_prefetch(sb, group, | |
2714 | nr, &prefetch_ios); | |
2715 | if (prefetch_ios == curr_ios) | |
2716 | nr = 0; | |
2717 | } | |
2718 | ||
8a57d9d6 | 2719 | /* This now checks without needing the buddy page */ |
8ef123fe | 2720 | ret = ext4_mb_good_group_nolock(ac, group, cr); |
42ac1848 LC |
2721 | if (ret <= 0) { |
2722 | if (!first_err) | |
2723 | first_err = ret; | |
c9de560d | 2724 | continue; |
42ac1848 | 2725 | } |
c9de560d | 2726 | |
c9de560d AT |
2727 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2728 | if (err) | |
2729 | goto out; | |
2730 | ||
2731 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2732 | |
2733 | /* | |
2734 | * We need to check again after locking the | |
2735 | * block group | |
2736 | */ | |
42ac1848 | 2737 | ret = ext4_mb_good_group(ac, group, cr); |
8ef123fe | 2738 | if (ret == 0) { |
c9de560d | 2739 | ext4_unlock_group(sb, group); |
e39e07fd | 2740 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2741 | continue; |
2742 | } | |
2743 | ||
2744 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2745 | if (cr == 0) |
c9de560d | 2746 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2747 | else if (cr == 1 && sbi->s_stripe && |
2748 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2749 | ext4_mb_scan_aligned(ac, &e4b); |
2750 | else | |
2751 | ext4_mb_complex_scan_group(ac, &e4b); | |
2752 | ||
2753 | ext4_unlock_group(sb, group); | |
e39e07fd | 2754 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2755 | |
2756 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2757 | break; | |
2758 | } | |
a6c75eaf HS |
2759 | /* Processed all groups and haven't found blocks */ |
2760 | if (sbi->s_mb_stats && i == ngroups) | |
2761 | atomic64_inc(&sbi->s_bal_cX_failed[cr]); | |
c9de560d AT |
2762 | } |
2763 | ||
2764 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2765 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2766 | /* | |
2767 | * We've been searching too long. Let's try to allocate | |
2768 | * the best chunk we've found so far | |
2769 | */ | |
c9de560d AT |
2770 | ext4_mb_try_best_found(ac, &e4b); |
2771 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2772 | /* | |
2773 | * Someone more lucky has already allocated it. | |
2774 | * The only thing we can do is just take first | |
2775 | * found block(s) | |
c9de560d | 2776 | */ |
66d5e027 | 2777 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
2778 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", | |
c55ee7d2 | 2779 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
2780 | ac->ac_b_ex.fe_len, lost); | |
2781 | ||
c9de560d AT |
2782 | ac->ac_b_ex.fe_group = 0; |
2783 | ac->ac_b_ex.fe_start = 0; | |
2784 | ac->ac_b_ex.fe_len = 0; | |
2785 | ac->ac_status = AC_STATUS_CONTINUE; | |
2786 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2787 | cr = 3; | |
c9de560d AT |
2788 | goto repeat; |
2789 | } | |
2790 | } | |
a6c75eaf HS |
2791 | |
2792 | if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) | |
2793 | atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); | |
c9de560d | 2794 | out: |
42ac1848 LC |
2795 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2796 | err = first_err; | |
bbc4ec77 | 2797 | |
d3df1453 | 2798 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
bbc4ec77 RH |
2799 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
2800 | ac->ac_flags, cr, err); | |
cfd73237 AZ |
2801 | |
2802 | if (nr) | |
2803 | ext4_mb_prefetch_fini(sb, prefetch_grp, nr); | |
2804 | ||
c9de560d AT |
2805 | return err; |
2806 | } | |
2807 | ||
c9de560d AT |
2808 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2809 | { | |
359745d7 | 2810 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
2811 | ext4_group_t group; |
2812 | ||
8df9675f | 2813 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2814 | return NULL; |
c9de560d | 2815 | group = *pos + 1; |
a9df9a49 | 2816 | return (void *) ((unsigned long) group); |
c9de560d AT |
2817 | } |
2818 | ||
2819 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2820 | { | |
359745d7 | 2821 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
2822 | ext4_group_t group; |
2823 | ||
2824 | ++*pos; | |
8df9675f | 2825 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2826 | return NULL; |
2827 | group = *pos + 1; | |
a9df9a49 | 2828 | return (void *) ((unsigned long) group); |
c9de560d AT |
2829 | } |
2830 | ||
2831 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2832 | { | |
359745d7 | 2833 | struct super_block *sb = pde_data(file_inode(seq->file)); |
a9df9a49 | 2834 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2835 | int i; |
1c8457ca | 2836 | int err, buddy_loaded = 0; |
c9de560d | 2837 | struct ext4_buddy e4b; |
1c8457ca | 2838 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2839 | unsigned char blocksize_bits = min_t(unsigned char, |
2840 | sb->s_blocksize_bits, | |
2841 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2842 | struct sg { |
2843 | struct ext4_group_info info; | |
b80b32b6 | 2844 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2845 | } sg; |
2846 | ||
2847 | group--; | |
2848 | if (group == 0) | |
97b4af2f RV |
2849 | seq_puts(seq, "#group: free frags first [" |
2850 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2851 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2852 | |
b80b32b6 TT |
2853 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2854 | sizeof(struct ext4_group_info); | |
2855 | ||
1c8457ca | 2856 | grinfo = ext4_get_group_info(sb, group); |
5354b2af TT |
2857 | if (!grinfo) |
2858 | return 0; | |
1c8457ca AK |
2859 | /* Load the group info in memory only if not already loaded. */ |
2860 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2861 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2862 | if (err) { | |
2863 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2864 | return 0; | |
2865 | } | |
2866 | buddy_loaded = 1; | |
c9de560d | 2867 | } |
1c8457ca | 2868 | |
5354b2af | 2869 | memcpy(&sg, grinfo, i); |
1c8457ca AK |
2870 | |
2871 | if (buddy_loaded) | |
2872 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2873 | |
a9df9a49 | 2874 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2875 | sg.info.bb_fragments, sg.info.bb_first_free); |
2876 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2877 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d | 2878 | sg.info.bb_counters[i] : 0); |
e0d438c7 | 2879 | seq_puts(seq, " ]\n"); |
c9de560d AT |
2880 | |
2881 | return 0; | |
2882 | } | |
2883 | ||
2884 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2885 | { | |
2886 | } | |
2887 | ||
247dbed8 | 2888 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2889 | .start = ext4_mb_seq_groups_start, |
2890 | .next = ext4_mb_seq_groups_next, | |
2891 | .stop = ext4_mb_seq_groups_stop, | |
2892 | .show = ext4_mb_seq_groups_show, | |
2893 | }; | |
2894 | ||
a6c75eaf HS |
2895 | int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) |
2896 | { | |
c30365b9 | 2897 | struct super_block *sb = seq->private; |
a6c75eaf HS |
2898 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2899 | ||
2900 | seq_puts(seq, "mballoc:\n"); | |
2901 | if (!sbi->s_mb_stats) { | |
2902 | seq_puts(seq, "\tmb stats collection turned off.\n"); | |
2903 | seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); | |
2904 | return 0; | |
2905 | } | |
2906 | seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); | |
2907 | seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); | |
2908 | ||
2909 | seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned)); | |
2910 | ||
2911 | seq_puts(seq, "\tcr0_stats:\n"); | |
2912 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0])); | |
2913 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2914 | atomic64_read(&sbi->s_bal_cX_groups_considered[0])); | |
2915 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2916 | atomic64_read(&sbi->s_bal_cX_failed[0])); | |
196e402a HS |
2917 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2918 | atomic_read(&sbi->s_bal_cr0_bad_suggestions)); | |
a6c75eaf HS |
2919 | |
2920 | seq_puts(seq, "\tcr1_stats:\n"); | |
2921 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1])); | |
2922 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2923 | atomic64_read(&sbi->s_bal_cX_groups_considered[1])); | |
2924 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2925 | atomic64_read(&sbi->s_bal_cX_failed[1])); | |
196e402a HS |
2926 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2927 | atomic_read(&sbi->s_bal_cr1_bad_suggestions)); | |
a6c75eaf HS |
2928 | |
2929 | seq_puts(seq, "\tcr2_stats:\n"); | |
2930 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2])); | |
2931 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2932 | atomic64_read(&sbi->s_bal_cX_groups_considered[2])); | |
2933 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2934 | atomic64_read(&sbi->s_bal_cX_failed[2])); | |
2935 | ||
2936 | seq_puts(seq, "\tcr3_stats:\n"); | |
2937 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3])); | |
2938 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2939 | atomic64_read(&sbi->s_bal_cX_groups_considered[3])); | |
2940 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2941 | atomic64_read(&sbi->s_bal_cX_failed[3])); | |
2942 | seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned)); | |
2943 | seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); | |
2944 | seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); | |
2945 | seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); | |
2946 | seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); | |
2947 | ||
2948 | seq_printf(seq, "\tbuddies_generated: %u/%u\n", | |
2949 | atomic_read(&sbi->s_mb_buddies_generated), | |
2950 | ext4_get_groups_count(sb)); | |
2951 | seq_printf(seq, "\tbuddies_time_used: %llu\n", | |
2952 | atomic64_read(&sbi->s_mb_generation_time)); | |
2953 | seq_printf(seq, "\tpreallocated: %u\n", | |
2954 | atomic_read(&sbi->s_mb_preallocated)); | |
2955 | seq_printf(seq, "\tdiscarded: %u\n", | |
2956 | atomic_read(&sbi->s_mb_discarded)); | |
2957 | return 0; | |
2958 | } | |
2959 | ||
f68f4063 | 2960 | static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) |
a5fda113 | 2961 | __acquires(&EXT4_SB(sb)->s_mb_rb_lock) |
f68f4063 | 2962 | { |
359745d7 | 2963 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2964 | unsigned long position; |
2965 | ||
83e80a6e | 2966 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
f68f4063 HS |
2967 | return NULL; |
2968 | position = *pos + 1; | |
2969 | return (void *) ((unsigned long) position); | |
2970 | } | |
2971 | ||
2972 | static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) | |
2973 | { | |
359745d7 | 2974 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2975 | unsigned long position; |
2976 | ||
2977 | ++*pos; | |
83e80a6e | 2978 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
f68f4063 HS |
2979 | return NULL; |
2980 | position = *pos + 1; | |
2981 | return (void *) ((unsigned long) position); | |
2982 | } | |
2983 | ||
2984 | static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) | |
2985 | { | |
359745d7 | 2986 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2987 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2988 | unsigned long position = ((unsigned long) v); | |
2989 | struct ext4_group_info *grp; | |
83e80a6e | 2990 | unsigned int count; |
f68f4063 HS |
2991 | |
2992 | position--; | |
2993 | if (position >= MB_NUM_ORDERS(sb)) { | |
83e80a6e JK |
2994 | position -= MB_NUM_ORDERS(sb); |
2995 | if (position == 0) | |
2996 | seq_puts(seq, "avg_fragment_size_lists:\n"); | |
f68f4063 | 2997 | |
83e80a6e JK |
2998 | count = 0; |
2999 | read_lock(&sbi->s_mb_avg_fragment_size_locks[position]); | |
3000 | list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position], | |
3001 | bb_avg_fragment_size_node) | |
3002 | count++; | |
3003 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]); | |
3004 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", | |
3005 | (unsigned int)position, count); | |
f68f4063 HS |
3006 | return 0; |
3007 | } | |
3008 | ||
3009 | if (position == 0) { | |
3010 | seq_printf(seq, "optimize_scan: %d\n", | |
3011 | test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); | |
3012 | seq_puts(seq, "max_free_order_lists:\n"); | |
3013 | } | |
3014 | count = 0; | |
83e80a6e | 3015 | read_lock(&sbi->s_mb_largest_free_orders_locks[position]); |
f68f4063 HS |
3016 | list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], |
3017 | bb_largest_free_order_node) | |
3018 | count++; | |
83e80a6e | 3019 | read_unlock(&sbi->s_mb_largest_free_orders_locks[position]); |
f68f4063 HS |
3020 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", |
3021 | (unsigned int)position, count); | |
3022 | ||
3023 | return 0; | |
3024 | } | |
3025 | ||
3026 | static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) | |
3027 | { | |
f68f4063 HS |
3028 | } |
3029 | ||
3030 | const struct seq_operations ext4_mb_seq_structs_summary_ops = { | |
3031 | .start = ext4_mb_seq_structs_summary_start, | |
3032 | .next = ext4_mb_seq_structs_summary_next, | |
3033 | .stop = ext4_mb_seq_structs_summary_stop, | |
3034 | .show = ext4_mb_seq_structs_summary_show, | |
3035 | }; | |
3036 | ||
fb1813f4 CW |
3037 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
3038 | { | |
3039 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3040 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
3041 | ||
3042 | BUG_ON(!cachep); | |
3043 | return cachep; | |
3044 | } | |
5f21b0e6 | 3045 | |
28623c2f TT |
3046 | /* |
3047 | * Allocate the top-level s_group_info array for the specified number | |
3048 | * of groups | |
3049 | */ | |
3050 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
3051 | { | |
3052 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3053 | unsigned size; | |
df3da4ea | 3054 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
3055 | |
3056 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
3057 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
3058 | if (size <= sbi->s_group_info_size) | |
3059 | return 0; | |
3060 | ||
3061 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 3062 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
3063 | if (!new_groupinfo) { |
3064 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
3065 | return -ENOMEM; | |
3066 | } | |
df3da4ea SJS |
3067 | rcu_read_lock(); |
3068 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
3069 | if (old_groupinfo) | |
3070 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 3071 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
3072 | rcu_read_unlock(); |
3073 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 3074 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
3075 | if (old_groupinfo) |
3076 | ext4_kvfree_array_rcu(old_groupinfo); | |
666245d9 | 3077 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
28623c2f TT |
3078 | sbi->s_group_info_size); |
3079 | return 0; | |
3080 | } | |
3081 | ||
5f21b0e6 | 3082 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 3083 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
3084 | struct ext4_group_desc *desc) |
3085 | { | |
fb1813f4 | 3086 | int i; |
5f21b0e6 | 3087 | int metalen = 0; |
df3da4ea | 3088 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
3089 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3090 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 3091 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
3092 | |
3093 | /* | |
3094 | * First check if this group is the first of a reserved block. | |
3095 | * If it's true, we have to allocate a new table of pointers | |
3096 | * to ext4_group_info structures | |
3097 | */ | |
3098 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
3099 | metalen = sizeof(*meta_group_info) << | |
3100 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 3101 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 3102 | if (meta_group_info == NULL) { |
7f6a11e7 | 3103 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 3104 | "for a buddy group"); |
df119095 | 3105 | return -ENOMEM; |
5f21b0e6 | 3106 | } |
df3da4ea SJS |
3107 | rcu_read_lock(); |
3108 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
3109 | rcu_read_unlock(); | |
5f21b0e6 FB |
3110 | } |
3111 | ||
df3da4ea | 3112 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
3113 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
3114 | ||
4fdb5543 | 3115 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 3116 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 3117 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
3118 | goto exit_group_info; |
3119 | } | |
3120 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
3121 | &(meta_group_info[i]->bb_state)); | |
3122 | ||
3123 | /* | |
3124 | * initialize bb_free to be able to skip | |
3125 | * empty groups without initialization | |
3126 | */ | |
8844618d TT |
3127 | if (ext4_has_group_desc_csum(sb) && |
3128 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 3129 | meta_group_info[i]->bb_free = |
cff1dfd7 | 3130 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
3131 | } else { |
3132 | meta_group_info[i]->bb_free = | |
021b65bb | 3133 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
3134 | } |
3135 | ||
3136 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 3137 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 3138 | meta_group_info[i]->bb_free_root = RB_ROOT; |
196e402a | 3139 | INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); |
83e80a6e | 3140 | INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); |
8a57d9d6 | 3141 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
83e80a6e | 3142 | meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ |
196e402a | 3143 | meta_group_info[i]->bb_group = group; |
5f21b0e6 | 3144 | |
a3450215 | 3145 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
3146 | return 0; |
3147 | ||
3148 | exit_group_info: | |
3149 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 3150 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
3151 | struct ext4_group_info ***group_info; |
3152 | ||
3153 | rcu_read_lock(); | |
3154 | group_info = rcu_dereference(sbi->s_group_info); | |
3155 | kfree(group_info[idx]); | |
3156 | group_info[idx] = NULL; | |
3157 | rcu_read_unlock(); | |
caaf7a29 | 3158 | } |
5f21b0e6 FB |
3159 | return -ENOMEM; |
3160 | } /* ext4_mb_add_groupinfo */ | |
3161 | ||
c9de560d AT |
3162 | static int ext4_mb_init_backend(struct super_block *sb) |
3163 | { | |
8df9675f | 3164 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 3165 | ext4_group_t i; |
c9de560d | 3166 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 3167 | int err; |
5f21b0e6 | 3168 | struct ext4_group_desc *desc; |
df3da4ea | 3169 | struct ext4_group_info ***group_info; |
fb1813f4 | 3170 | struct kmem_cache *cachep; |
5f21b0e6 | 3171 | |
28623c2f TT |
3172 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
3173 | if (err) | |
3174 | return err; | |
c9de560d | 3175 | |
c9de560d AT |
3176 | sbi->s_buddy_cache = new_inode(sb); |
3177 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 3178 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
3179 | goto err_freesgi; |
3180 | } | |
48e6061b YJ |
3181 | /* To avoid potentially colliding with an valid on-disk inode number, |
3182 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
3183 | * not in the inode hash, so it should never be found by iget(), but | |
3184 | * this will avoid confusion if it ever shows up during debugging. */ | |
3185 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 3186 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 3187 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3188 | cond_resched(); |
c9de560d AT |
3189 | desc = ext4_get_group_desc(sb, i, NULL); |
3190 | if (desc == NULL) { | |
9d8b9ec4 | 3191 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
3192 | goto err_freebuddy; |
3193 | } | |
5f21b0e6 FB |
3194 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
3195 | goto err_freebuddy; | |
c9de560d AT |
3196 | } |
3197 | ||
cfd73237 | 3198 | if (ext4_has_feature_flex_bg(sb)) { |
f91436d5 ST |
3199 | /* a single flex group is supposed to be read by a single IO. |
3200 | * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is | |
3201 | * unsigned integer, so the maximum shift is 32. | |
3202 | */ | |
3203 | if (sbi->s_es->s_log_groups_per_flex >= 32) { | |
3204 | ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); | |
a8867f4e | 3205 | goto err_freebuddy; |
f91436d5 ST |
3206 | } |
3207 | sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, | |
82ef1370 | 3208 | BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); |
cfd73237 AZ |
3209 | sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ |
3210 | } else { | |
3211 | sbi->s_mb_prefetch = 32; | |
3212 | } | |
3213 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) | |
3214 | sbi->s_mb_prefetch = ext4_get_groups_count(sb); | |
3215 | /* now many real IOs to prefetch within a single allocation at cr=0 | |
3216 | * given cr=0 is an CPU-related optimization we shouldn't try to | |
3217 | * load too many groups, at some point we should start to use what | |
3218 | * we've got in memory. | |
3219 | * with an average random access time 5ms, it'd take a second to get | |
3220 | * 200 groups (* N with flex_bg), so let's make this limit 4 | |
3221 | */ | |
3222 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; | |
3223 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) | |
3224 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); | |
3225 | ||
c9de560d AT |
3226 | return 0; |
3227 | ||
3228 | err_freebuddy: | |
fb1813f4 | 3229 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5354b2af TT |
3230 | while (i-- > 0) { |
3231 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); | |
3232 | ||
3233 | if (grp) | |
3234 | kmem_cache_free(cachep, grp); | |
3235 | } | |
28623c2f | 3236 | i = sbi->s_group_info_size; |
df3da4ea SJS |
3237 | rcu_read_lock(); |
3238 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 3239 | while (i-- > 0) |
df3da4ea SJS |
3240 | kfree(group_info[i]); |
3241 | rcu_read_unlock(); | |
c9de560d AT |
3242 | iput(sbi->s_buddy_cache); |
3243 | err_freesgi: | |
df3da4ea SJS |
3244 | rcu_read_lock(); |
3245 | kvfree(rcu_dereference(sbi->s_group_info)); | |
3246 | rcu_read_unlock(); | |
c9de560d AT |
3247 | return -ENOMEM; |
3248 | } | |
3249 | ||
2892c15d ES |
3250 | static void ext4_groupinfo_destroy_slabs(void) |
3251 | { | |
3252 | int i; | |
3253 | ||
3254 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 3255 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
3256 | ext4_groupinfo_caches[i] = NULL; |
3257 | } | |
3258 | } | |
3259 | ||
3260 | static int ext4_groupinfo_create_slab(size_t size) | |
3261 | { | |
3262 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
3263 | int slab_size; | |
3264 | int blocksize_bits = order_base_2(size); | |
3265 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3266 | struct kmem_cache *cachep; | |
3267 | ||
3268 | if (cache_index >= NR_GRPINFO_CACHES) | |
3269 | return -EINVAL; | |
3270 | ||
3271 | if (unlikely(cache_index < 0)) | |
3272 | cache_index = 0; | |
3273 | ||
3274 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
3275 | if (ext4_groupinfo_caches[cache_index]) { | |
3276 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
3277 | return 0; /* Already created */ | |
3278 | } | |
3279 | ||
3280 | slab_size = offsetof(struct ext4_group_info, | |
3281 | bb_counters[blocksize_bits + 2]); | |
3282 | ||
3283 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
3284 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
3285 | NULL); | |
3286 | ||
823ba01f TM |
3287 | ext4_groupinfo_caches[cache_index] = cachep; |
3288 | ||
2892c15d ES |
3289 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
3290 | if (!cachep) { | |
9d8b9ec4 TT |
3291 | printk(KERN_EMERG |
3292 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
3293 | return -ENOMEM; |
3294 | } | |
3295 | ||
2892c15d ES |
3296 | return 0; |
3297 | } | |
3298 | ||
55cdd0af WJ |
3299 | static void ext4_discard_work(struct work_struct *work) |
3300 | { | |
3301 | struct ext4_sb_info *sbi = container_of(work, | |
3302 | struct ext4_sb_info, s_discard_work); | |
3303 | struct super_block *sb = sbi->s_sb; | |
3304 | struct ext4_free_data *fd, *nfd; | |
3305 | struct ext4_buddy e4b; | |
3306 | struct list_head discard_list; | |
3307 | ext4_group_t grp, load_grp; | |
3308 | int err = 0; | |
3309 | ||
3310 | INIT_LIST_HEAD(&discard_list); | |
3311 | spin_lock(&sbi->s_md_lock); | |
3312 | list_splice_init(&sbi->s_discard_list, &discard_list); | |
3313 | spin_unlock(&sbi->s_md_lock); | |
3314 | ||
3315 | load_grp = UINT_MAX; | |
3316 | list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { | |
3317 | /* | |
5036ab8d WJ |
3318 | * If filesystem is umounting or no memory or suffering |
3319 | * from no space, give up the discard | |
55cdd0af | 3320 | */ |
5036ab8d WJ |
3321 | if ((sb->s_flags & SB_ACTIVE) && !err && |
3322 | !atomic_read(&sbi->s_retry_alloc_pending)) { | |
55cdd0af WJ |
3323 | grp = fd->efd_group; |
3324 | if (grp != load_grp) { | |
3325 | if (load_grp != UINT_MAX) | |
3326 | ext4_mb_unload_buddy(&e4b); | |
3327 | ||
3328 | err = ext4_mb_load_buddy(sb, grp, &e4b); | |
3329 | if (err) { | |
3330 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3331 | load_grp = UINT_MAX; | |
3332 | continue; | |
3333 | } else { | |
3334 | load_grp = grp; | |
3335 | } | |
3336 | } | |
3337 | ||
3338 | ext4_lock_group(sb, grp); | |
3339 | ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, | |
3340 | fd->efd_start_cluster + fd->efd_count - 1, 1); | |
3341 | ext4_unlock_group(sb, grp); | |
3342 | } | |
3343 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3344 | } | |
3345 | ||
3346 | if (load_grp != UINT_MAX) | |
3347 | ext4_mb_unload_buddy(&e4b); | |
3348 | } | |
3349 | ||
9d99012f | 3350 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
3351 | { |
3352 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 3353 | unsigned i, j; |
935244cd | 3354 | unsigned offset, offset_incr; |
c9de560d | 3355 | unsigned max; |
74767c5a | 3356 | int ret; |
c9de560d | 3357 | |
4b68f6df | 3358 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
3359 | |
3360 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
3361 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
3362 | ret = -ENOMEM; |
3363 | goto out; | |
c9de560d | 3364 | } |
ff7ef329 | 3365 | |
4b68f6df | 3366 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
3367 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
3368 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
3369 | ret = -ENOMEM; |
3370 | goto out; | |
3371 | } | |
3372 | ||
2892c15d ES |
3373 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
3374 | if (ret < 0) | |
3375 | goto out; | |
c9de560d AT |
3376 | |
3377 | /* order 0 is regular bitmap */ | |
3378 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
3379 | sbi->s_mb_offsets[0] = 0; | |
3380 | ||
3381 | i = 1; | |
3382 | offset = 0; | |
935244cd | 3383 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
3384 | max = sb->s_blocksize << 2; |
3385 | do { | |
3386 | sbi->s_mb_offsets[i] = offset; | |
3387 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
3388 | offset += offset_incr; |
3389 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
3390 | max = max >> 1; |
3391 | i++; | |
4b68f6df HS |
3392 | } while (i < MB_NUM_ORDERS(sb)); |
3393 | ||
83e80a6e JK |
3394 | sbi->s_mb_avg_fragment_size = |
3395 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3396 | GFP_KERNEL); | |
3397 | if (!sbi->s_mb_avg_fragment_size) { | |
3398 | ret = -ENOMEM; | |
3399 | goto out; | |
3400 | } | |
3401 | sbi->s_mb_avg_fragment_size_locks = | |
3402 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3403 | GFP_KERNEL); | |
3404 | if (!sbi->s_mb_avg_fragment_size_locks) { | |
3405 | ret = -ENOMEM; | |
3406 | goto out; | |
3407 | } | |
3408 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3409 | INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); | |
3410 | rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); | |
3411 | } | |
196e402a HS |
3412 | sbi->s_mb_largest_free_orders = |
3413 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3414 | GFP_KERNEL); | |
3415 | if (!sbi->s_mb_largest_free_orders) { | |
3416 | ret = -ENOMEM; | |
3417 | goto out; | |
3418 | } | |
3419 | sbi->s_mb_largest_free_orders_locks = | |
3420 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3421 | GFP_KERNEL); | |
3422 | if (!sbi->s_mb_largest_free_orders_locks) { | |
3423 | ret = -ENOMEM; | |
3424 | goto out; | |
3425 | } | |
3426 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3427 | INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); | |
3428 | rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); | |
3429 | } | |
c9de560d | 3430 | |
c9de560d | 3431 | spin_lock_init(&sbi->s_md_lock); |
d08854f5 | 3432 | sbi->s_mb_free_pending = 0; |
a0154344 | 3433 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
55cdd0af WJ |
3434 | INIT_LIST_HEAD(&sbi->s_discard_list); |
3435 | INIT_WORK(&sbi->s_discard_work, ext4_discard_work); | |
5036ab8d | 3436 | atomic_set(&sbi->s_retry_alloc_pending, 0); |
c9de560d AT |
3437 | |
3438 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
3439 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
3440 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
3441 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
3442 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
3443 | /* |
3444 | * The default group preallocation is 512, which for 4k block | |
3445 | * sizes translates to 2 megabytes. However for bigalloc file | |
3446 | * systems, this is probably too big (i.e, if the cluster size | |
3447 | * is 1 megabyte, then group preallocation size becomes half a | |
3448 | * gigabyte!). As a default, we will keep a two megabyte | |
3449 | * group pralloc size for cluster sizes up to 64k, and after | |
3450 | * that, we will force a minimum group preallocation size of | |
3451 | * 32 clusters. This translates to 8 megs when the cluster | |
3452 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
3453 | * which seems reasonable as a default. | |
3454 | */ | |
3455 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
3456 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
3457 | /* |
3458 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
3459 | * to the lowest multiple of s_stripe which is bigger than | |
3460 | * the s_mb_group_prealloc as determined above. We want | |
3461 | * the preallocation size to be an exact multiple of the | |
3462 | * RAID stripe size so that preallocations don't fragment | |
3463 | * the stripes. | |
3464 | */ | |
3465 | if (sbi->s_stripe > 1) { | |
3466 | sbi->s_mb_group_prealloc = roundup( | |
3467 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
3468 | } | |
c9de560d | 3469 | |
730c213c | 3470 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 3471 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 3472 | ret = -ENOMEM; |
029b10c5 | 3473 | goto out; |
c9de560d | 3474 | } |
730c213c | 3475 | for_each_possible_cpu(i) { |
c9de560d | 3476 | struct ext4_locality_group *lg; |
730c213c | 3477 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 3478 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
3479 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
3480 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
3481 | spin_lock_init(&lg->lg_prealloc_lock); |
3482 | } | |
3483 | ||
10f0d2a5 | 3484 | if (bdev_nonrot(sb->s_bdev)) |
196e402a HS |
3485 | sbi->s_mb_max_linear_groups = 0; |
3486 | else | |
3487 | sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; | |
79a77c5a YJ |
3488 | /* init file for buddy data */ |
3489 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
3490 | if (ret != 0) |
3491 | goto out_free_locality_groups; | |
79a77c5a | 3492 | |
7aa0baea TM |
3493 | return 0; |
3494 | ||
3495 | out_free_locality_groups: | |
3496 | free_percpu(sbi->s_locality_groups); | |
3497 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 3498 | out: |
83e80a6e JK |
3499 | kfree(sbi->s_mb_avg_fragment_size); |
3500 | kfree(sbi->s_mb_avg_fragment_size_locks); | |
196e402a HS |
3501 | kfree(sbi->s_mb_largest_free_orders); |
3502 | kfree(sbi->s_mb_largest_free_orders_locks); | |
7aa0baea TM |
3503 | kfree(sbi->s_mb_offsets); |
3504 | sbi->s_mb_offsets = NULL; | |
3505 | kfree(sbi->s_mb_maxs); | |
3506 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 3507 | return ret; |
c9de560d AT |
3508 | } |
3509 | ||
955ce5f5 | 3510 | /* need to called with the ext4 group lock held */ |
d3df1453 | 3511 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
c9de560d AT |
3512 | { |
3513 | struct ext4_prealloc_space *pa; | |
3514 | struct list_head *cur, *tmp; | |
3515 | int count = 0; | |
3516 | ||
3517 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
3518 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3519 | list_del(&pa->pa_group_list); | |
3520 | count++; | |
688f05a0 | 3521 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d | 3522 | } |
d3df1453 | 3523 | return count; |
c9de560d AT |
3524 | } |
3525 | ||
3526 | int ext4_mb_release(struct super_block *sb) | |
3527 | { | |
8df9675f | 3528 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
3529 | ext4_group_t i; |
3530 | int num_meta_group_infos; | |
df3da4ea | 3531 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 3532 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 3533 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
d3df1453 | 3534 | int count; |
c9de560d | 3535 | |
55cdd0af WJ |
3536 | if (test_opt(sb, DISCARD)) { |
3537 | /* | |
3538 | * wait the discard work to drain all of ext4_free_data | |
3539 | */ | |
3540 | flush_work(&sbi->s_discard_work); | |
3541 | WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); | |
3542 | } | |
3543 | ||
c9de560d | 3544 | if (sbi->s_group_info) { |
8df9675f | 3545 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3546 | cond_resched(); |
c9de560d | 3547 | grinfo = ext4_get_group_info(sb, i); |
5354b2af TT |
3548 | if (!grinfo) |
3549 | continue; | |
a3450215 | 3550 | mb_group_bb_bitmap_free(grinfo); |
c9de560d | 3551 | ext4_lock_group(sb, i); |
d3df1453 RH |
3552 | count = ext4_mb_cleanup_pa(grinfo); |
3553 | if (count) | |
3554 | mb_debug(sb, "mballoc: %d PAs left\n", | |
3555 | count); | |
c9de560d | 3556 | ext4_unlock_group(sb, i); |
fb1813f4 | 3557 | kmem_cache_free(cachep, grinfo); |
c9de560d | 3558 | } |
8df9675f | 3559 | num_meta_group_infos = (ngroups + |
c9de560d AT |
3560 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
3561 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
3562 | rcu_read_lock(); |
3563 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 3564 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
3565 | kfree(group_info[i]); |
3566 | kvfree(group_info); | |
3567 | rcu_read_unlock(); | |
c9de560d | 3568 | } |
83e80a6e JK |
3569 | kfree(sbi->s_mb_avg_fragment_size); |
3570 | kfree(sbi->s_mb_avg_fragment_size_locks); | |
196e402a HS |
3571 | kfree(sbi->s_mb_largest_free_orders); |
3572 | kfree(sbi->s_mb_largest_free_orders_locks); | |
c9de560d AT |
3573 | kfree(sbi->s_mb_offsets); |
3574 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 3575 | iput(sbi->s_buddy_cache); |
c9de560d | 3576 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
3577 | ext4_msg(sb, KERN_INFO, |
3578 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
3579 | atomic_read(&sbi->s_bal_allocated), |
3580 | atomic_read(&sbi->s_bal_reqs), | |
3581 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 | 3582 | ext4_msg(sb, KERN_INFO, |
a6c75eaf | 3583 | "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " |
9d8b9ec4 | 3584 | "%u 2^N hits, %u breaks, %u lost", |
c9de560d | 3585 | atomic_read(&sbi->s_bal_ex_scanned), |
a6c75eaf | 3586 | atomic_read(&sbi->s_bal_groups_scanned), |
c9de560d AT |
3587 | atomic_read(&sbi->s_bal_goals), |
3588 | atomic_read(&sbi->s_bal_2orders), | |
3589 | atomic_read(&sbi->s_bal_breaks), | |
3590 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 | 3591 | ext4_msg(sb, KERN_INFO, |
67d25186 HS |
3592 | "mballoc: %u generated and it took %llu", |
3593 | atomic_read(&sbi->s_mb_buddies_generated), | |
3594 | atomic64_read(&sbi->s_mb_generation_time)); | |
9d8b9ec4 TT |
3595 | ext4_msg(sb, KERN_INFO, |
3596 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
3597 | atomic_read(&sbi->s_mb_preallocated), |
3598 | atomic_read(&sbi->s_mb_discarded)); | |
3599 | } | |
3600 | ||
730c213c | 3601 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
3602 | |
3603 | return 0; | |
3604 | } | |
3605 | ||
77ca6cdf | 3606 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
3607 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
3608 | struct bio **biop) | |
5c521830 | 3609 | { |
5c521830 JZ |
3610 | ext4_fsblk_t discard_block; |
3611 | ||
84130193 TT |
3612 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
3613 | ext4_group_first_block_no(sb, block_group)); | |
3614 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
3615 | trace_ext4_discard_blocks(sb, |
3616 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
3617 | if (biop) { |
3618 | return __blkdev_issue_discard(sb->s_bdev, | |
3619 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
3620 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
44abff2c | 3621 | GFP_NOFS, biop); |
a0154344 DJ |
3622 | } else |
3623 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
3624 | } |
3625 | ||
a0154344 DJ |
3626 | static void ext4_free_data_in_buddy(struct super_block *sb, |
3627 | struct ext4_free_data *entry) | |
c9de560d | 3628 | { |
c9de560d | 3629 | struct ext4_buddy e4b; |
c894058d | 3630 | struct ext4_group_info *db; |
c7f2bafa | 3631 | int err, count = 0; |
c9de560d | 3632 | |
d3df1453 | 3633 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
18aadd47 | 3634 | entry->efd_count, entry->efd_group, entry); |
c9de560d | 3635 | |
18aadd47 BJ |
3636 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
3637 | /* we expect to find existing buddy because it's pinned */ | |
3638 | BUG_ON(err != 0); | |
b90f6870 | 3639 | |
d08854f5 TT |
3640 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
3641 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
3642 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 3643 | |
18aadd47 BJ |
3644 | db = e4b.bd_info; |
3645 | /* there are blocks to put in buddy to make them really free */ | |
3646 | count += entry->efd_count; | |
18aadd47 BJ |
3647 | ext4_lock_group(sb, entry->efd_group); |
3648 | /* Take it out of per group rb tree */ | |
3649 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
3650 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 3651 | |
18aadd47 BJ |
3652 | /* |
3653 | * Clear the trimmed flag for the group so that the next | |
3654 | * ext4_trim_fs can trim it. | |
3655 | * If the volume is mounted with -o discard, online discard | |
3656 | * is supported and the free blocks will be trimmed online. | |
3657 | */ | |
3658 | if (!test_opt(sb, DISCARD)) | |
3659 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 3660 | |
18aadd47 BJ |
3661 | if (!db->bb_free_root.rb_node) { |
3662 | /* No more items in the per group rb tree | |
3663 | * balance refcounts from ext4_mb_free_metadata() | |
3664 | */ | |
09cbfeaf KS |
3665 | put_page(e4b.bd_buddy_page); |
3666 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 3667 | } |
18aadd47 | 3668 | ext4_unlock_group(sb, entry->efd_group); |
18aadd47 | 3669 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 3670 | |
c7f2bafa | 3671 | mb_debug(sb, "freed %d blocks in 1 structures\n", count); |
c9de560d AT |
3672 | } |
3673 | ||
a0154344 DJ |
3674 | /* |
3675 | * This function is called by the jbd2 layer once the commit has finished, | |
3676 | * so we know we can free the blocks that were released with that commit. | |
3677 | */ | |
3678 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
3679 | { | |
3680 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3681 | struct ext4_free_data *entry, *tmp; | |
a0154344 DJ |
3682 | struct list_head freed_data_list; |
3683 | struct list_head *cut_pos = NULL; | |
55cdd0af | 3684 | bool wake; |
a0154344 DJ |
3685 | |
3686 | INIT_LIST_HEAD(&freed_data_list); | |
3687 | ||
3688 | spin_lock(&sbi->s_md_lock); | |
3689 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
3690 | if (entry->efd_tid != commit_tid) | |
3691 | break; | |
3692 | cut_pos = &entry->efd_list; | |
3693 | } | |
3694 | if (cut_pos) | |
3695 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
3696 | cut_pos); | |
3697 | spin_unlock(&sbi->s_md_lock); | |
3698 | ||
55cdd0af WJ |
3699 | list_for_each_entry(entry, &freed_data_list, efd_list) |
3700 | ext4_free_data_in_buddy(sb, entry); | |
a0154344 | 3701 | |
55cdd0af WJ |
3702 | if (test_opt(sb, DISCARD)) { |
3703 | spin_lock(&sbi->s_md_lock); | |
3704 | wake = list_empty(&sbi->s_discard_list); | |
3705 | list_splice_tail(&freed_data_list, &sbi->s_discard_list); | |
3706 | spin_unlock(&sbi->s_md_lock); | |
3707 | if (wake) | |
3708 | queue_work(system_unbound_wq, &sbi->s_discard_work); | |
3709 | } else { | |
3710 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
3711 | kmem_cache_free(ext4_free_data_cachep, entry); | |
a0154344 | 3712 | } |
a0154344 DJ |
3713 | } |
3714 | ||
5dabfc78 | 3715 | int __init ext4_init_mballoc(void) |
c9de560d | 3716 | { |
16828088 TT |
3717 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
3718 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 3719 | if (ext4_pspace_cachep == NULL) |
f283529a | 3720 | goto out; |
c9de560d | 3721 | |
16828088 TT |
3722 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
3723 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3724 | if (ext4_ac_cachep == NULL) |
3725 | goto out_pa_free; | |
c894058d | 3726 | |
18aadd47 BJ |
3727 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
3728 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3729 | if (ext4_free_data_cachep == NULL) |
3730 | goto out_ac_free; | |
3731 | ||
c9de560d | 3732 | return 0; |
f283529a RH |
3733 | |
3734 | out_ac_free: | |
3735 | kmem_cache_destroy(ext4_ac_cachep); | |
3736 | out_pa_free: | |
3737 | kmem_cache_destroy(ext4_pspace_cachep); | |
3738 | out: | |
3739 | return -ENOMEM; | |
c9de560d AT |
3740 | } |
3741 | ||
5dabfc78 | 3742 | void ext4_exit_mballoc(void) |
c9de560d | 3743 | { |
60e6679e | 3744 | /* |
3e03f9ca JDB |
3745 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
3746 | * before destroying the slab cache. | |
3747 | */ | |
3748 | rcu_barrier(); | |
c9de560d | 3749 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 3750 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 3751 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 3752 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
3753 | } |
3754 | ||
3755 | ||
3756 | /* | |
73b2c716 | 3757 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
3758 | * Returns 0 if success or error code |
3759 | */ | |
4ddfef7b ES |
3760 | static noinline_for_stack int |
3761 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 3762 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
3763 | { |
3764 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
3765 | struct ext4_group_desc *gdp; |
3766 | struct buffer_head *gdp_bh; | |
3767 | struct ext4_sb_info *sbi; | |
3768 | struct super_block *sb; | |
3769 | ext4_fsblk_t block; | |
519deca0 | 3770 | int err, len; |
c9de560d AT |
3771 | |
3772 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3773 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
3774 | ||
3775 | sb = ac->ac_sb; | |
3776 | sbi = EXT4_SB(sb); | |
c9de560d | 3777 | |
574ca174 | 3778 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e | 3779 | if (IS_ERR(bitmap_bh)) { |
fb28f9ce | 3780 | return PTR_ERR(bitmap_bh); |
9008a58e | 3781 | } |
c9de560d | 3782 | |
5d601255 | 3783 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
3784 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
3785 | EXT4_JTR_NONE); | |
c9de560d AT |
3786 | if (err) |
3787 | goto out_err; | |
3788 | ||
3789 | err = -EIO; | |
3790 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3791 | if (!gdp) | |
3792 | goto out_err; | |
3793 | ||
a9df9a49 | 3794 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3795 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3796 | |
5d601255 | 3797 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
188c299e | 3798 | err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE); |
c9de560d AT |
3799 | if (err) |
3800 | goto out_err; | |
3801 | ||
bda00de7 | 3802 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3803 | |
53accfa9 | 3804 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
ce9f24cc | 3805 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
12062ddd | 3806 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3807 | "fs metadata", block, block+len); |
519deca0 | 3808 | /* File system mounted not to panic on error |
554a5ccc | 3809 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3810 | * We leak some of the blocks here. |
3811 | */ | |
955ce5f5 | 3812 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
123e3016 | 3813 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
c3e94d1d | 3814 | ac->ac_b_ex.fe_len); |
955ce5f5 | 3815 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3816 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3817 | if (!err) |
554a5ccc | 3818 | err = -EFSCORRUPTED; |
519deca0 | 3819 | goto out_err; |
c9de560d | 3820 | } |
955ce5f5 AK |
3821 | |
3822 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3823 | #ifdef AGGRESSIVE_CHECK |
3824 | { | |
3825 | int i; | |
3826 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3827 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3828 | bitmap_bh->b_data)); | |
3829 | } | |
3830 | } | |
3831 | #endif | |
123e3016 | 3832 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
c3e94d1d | 3833 | ac->ac_b_ex.fe_len); |
8844618d TT |
3834 | if (ext4_has_group_desc_csum(sb) && |
3835 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3836 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3837 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3838 | ext4_free_clusters_after_init(sb, |
021b65bb | 3839 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3840 | } |
021b65bb TT |
3841 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3842 | ext4_free_group_clusters_set(sb, gdp, len); | |
1df9bde4 | 3843 | ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); |
feb0ab32 | 3844 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3845 | |
3846 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3847 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3848 | /* |
6bc6e63f | 3849 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3850 | */ |
6bc6e63f AK |
3851 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3852 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3853 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3854 | reserv_clstrs); | |
c9de560d | 3855 | |
772cb7c8 JS |
3856 | if (sbi->s_log_groups_per_flex) { |
3857 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3858 | ac->ac_b_ex.fe_group); | |
90ba983f | 3859 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3860 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3861 | flex_group)->free_clusters); | |
772cb7c8 JS |
3862 | } |
3863 | ||
0390131b | 3864 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3865 | if (err) |
3866 | goto out_err; | |
0390131b | 3867 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3868 | |
3869 | out_err: | |
42a10add | 3870 | brelse(bitmap_bh); |
c9de560d AT |
3871 | return err; |
3872 | } | |
3873 | ||
8016e29f HS |
3874 | /* |
3875 | * Idempotent helper for Ext4 fast commit replay path to set the state of | |
3876 | * blocks in bitmaps and update counters. | |
3877 | */ | |
3878 | void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, | |
3879 | int len, int state) | |
3880 | { | |
3881 | struct buffer_head *bitmap_bh = NULL; | |
3882 | struct ext4_group_desc *gdp; | |
3883 | struct buffer_head *gdp_bh; | |
3884 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3885 | ext4_group_t group; | |
3886 | ext4_grpblk_t blkoff; | |
a5c0e2fd | 3887 | int i, err; |
8016e29f | 3888 | int already; |
bfdc502a | 3889 | unsigned int clen, clen_changed, thisgrp_len; |
8016e29f | 3890 | |
bfdc502a RH |
3891 | while (len > 0) { |
3892 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
8016e29f | 3893 | |
bfdc502a RH |
3894 | /* |
3895 | * Check to see if we are freeing blocks across a group | |
3896 | * boundary. | |
3897 | * In case of flex_bg, this can happen that (block, len) may | |
3898 | * span across more than one group. In that case we need to | |
3899 | * get the corresponding group metadata to work with. | |
3900 | * For this we have goto again loop. | |
3901 | */ | |
3902 | thisgrp_len = min_t(unsigned int, (unsigned int)len, | |
3903 | EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); | |
3904 | clen = EXT4_NUM_B2C(sbi, thisgrp_len); | |
3905 | ||
8c91c579 RH |
3906 | if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { |
3907 | ext4_error(sb, "Marking blocks in system zone - " | |
3908 | "Block = %llu, len = %u", | |
3909 | block, thisgrp_len); | |
3910 | bitmap_bh = NULL; | |
3911 | break; | |
3912 | } | |
3913 | ||
bfdc502a RH |
3914 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
3915 | if (IS_ERR(bitmap_bh)) { | |
3916 | err = PTR_ERR(bitmap_bh); | |
3917 | bitmap_bh = NULL; | |
3918 | break; | |
3919 | } | |
8016e29f | 3920 | |
bfdc502a RH |
3921 | err = -EIO; |
3922 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
3923 | if (!gdp) | |
3924 | break; | |
8016e29f | 3925 | |
bfdc502a RH |
3926 | ext4_lock_group(sb, group); |
3927 | already = 0; | |
3928 | for (i = 0; i < clen; i++) | |
3929 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == | |
3930 | !state) | |
3931 | already++; | |
8016e29f | 3932 | |
bfdc502a RH |
3933 | clen_changed = clen - already; |
3934 | if (state) | |
123e3016 | 3935 | mb_set_bits(bitmap_bh->b_data, blkoff, clen); |
bfdc502a | 3936 | else |
bd8247ee | 3937 | mb_clear_bits(bitmap_bh->b_data, blkoff, clen); |
bfdc502a RH |
3938 | if (ext4_has_group_desc_csum(sb) && |
3939 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
3940 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
3941 | ext4_free_group_clusters_set(sb, gdp, | |
3942 | ext4_free_clusters_after_init(sb, group, gdp)); | |
3943 | } | |
3944 | if (state) | |
3945 | clen = ext4_free_group_clusters(sb, gdp) - clen_changed; | |
3946 | else | |
3947 | clen = ext4_free_group_clusters(sb, gdp) + clen_changed; | |
8016e29f | 3948 | |
bfdc502a | 3949 | ext4_free_group_clusters_set(sb, gdp, clen); |
1df9bde4 | 3950 | ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); |
bfdc502a | 3951 | ext4_group_desc_csum_set(sb, group, gdp); |
8016e29f | 3952 | |
bfdc502a | 3953 | ext4_unlock_group(sb, group); |
8016e29f | 3954 | |
bfdc502a RH |
3955 | if (sbi->s_log_groups_per_flex) { |
3956 | ext4_group_t flex_group = ext4_flex_group(sbi, group); | |
3957 | struct flex_groups *fg = sbi_array_rcu_deref(sbi, | |
3958 | s_flex_groups, flex_group); | |
8016e29f | 3959 | |
bfdc502a RH |
3960 | if (state) |
3961 | atomic64_sub(clen_changed, &fg->free_clusters); | |
3962 | else | |
3963 | atomic64_add(clen_changed, &fg->free_clusters); | |
3964 | ||
3965 | } | |
3966 | ||
3967 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
3968 | if (err) | |
3969 | break; | |
3970 | sync_dirty_buffer(bitmap_bh); | |
3971 | err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
3972 | sync_dirty_buffer(gdp_bh); | |
3973 | if (err) | |
3974 | break; | |
3975 | ||
3976 | block += thisgrp_len; | |
3977 | len -= thisgrp_len; | |
3978 | brelse(bitmap_bh); | |
3979 | BUG_ON(len < 0); | |
8016e29f HS |
3980 | } |
3981 | ||
8016e29f | 3982 | if (err) |
bfdc502a | 3983 | brelse(bitmap_bh); |
8016e29f HS |
3984 | } |
3985 | ||
c9de560d AT |
3986 | /* |
3987 | * here we normalize request for locality group | |
d7a1fee1 DE |
3988 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3989 | * s_strip if we set the same via mount option. | |
3990 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3991 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3992 | * |
3993 | * XXX: should we try to preallocate more than the group has now? | |
3994 | */ | |
3995 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3996 | { | |
3997 | struct super_block *sb = ac->ac_sb; | |
3998 | struct ext4_locality_group *lg = ac->ac_lg; | |
3999 | ||
4000 | BUG_ON(lg == NULL); | |
d7a1fee1 | 4001 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
d3df1453 | 4002 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
c9de560d AT |
4003 | } |
4004 | ||
38727786 OM |
4005 | /* |
4006 | * This function returns the next element to look at during inode | |
4007 | * PA rbtree walk. We assume that we have held the inode PA rbtree lock | |
4008 | * (ei->i_prealloc_lock) | |
4009 | * | |
4010 | * new_start The start of the range we want to compare | |
4011 | * cur_start The existing start that we are comparing against | |
4012 | * node The node of the rb_tree | |
4013 | */ | |
4014 | static inline struct rb_node* | |
4015 | ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node) | |
4016 | { | |
4017 | if (new_start < cur_start) | |
4018 | return node->rb_left; | |
4019 | else | |
4020 | return node->rb_right; | |
4021 | } | |
4022 | ||
7692094a OM |
4023 | static inline void |
4024 | ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac, | |
4025 | ext4_lblk_t start, ext4_lblk_t end) | |
4026 | { | |
4027 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4028 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
4029 | struct ext4_prealloc_space *tmp_pa; | |
4030 | ext4_lblk_t tmp_pa_start, tmp_pa_end; | |
38727786 | 4031 | struct rb_node *iter; |
7692094a | 4032 | |
38727786 OM |
4033 | read_lock(&ei->i_prealloc_lock); |
4034 | for (iter = ei->i_prealloc_node.rb_node; iter; | |
4035 | iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) { | |
4036 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4037 | pa_node.inode_node); | |
4038 | tmp_pa_start = tmp_pa->pa_lstart; | |
4039 | tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len); | |
7692094a | 4040 | |
38727786 OM |
4041 | spin_lock(&tmp_pa->pa_lock); |
4042 | if (tmp_pa->pa_deleted == 0) | |
7692094a | 4043 | BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start)); |
7692094a OM |
4044 | spin_unlock(&tmp_pa->pa_lock); |
4045 | } | |
38727786 | 4046 | read_unlock(&ei->i_prealloc_lock); |
7692094a OM |
4047 | } |
4048 | ||
0830344c OM |
4049 | /* |
4050 | * Given an allocation context "ac" and a range "start", "end", check | |
4051 | * and adjust boundaries if the range overlaps with any of the existing | |
4052 | * preallocatoins stored in the corresponding inode of the allocation context. | |
4053 | * | |
38727786 | 4054 | * Parameters: |
0830344c OM |
4055 | * ac allocation context |
4056 | * start start of the new range | |
4057 | * end end of the new range | |
4058 | */ | |
4059 | static inline void | |
4060 | ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac, | |
38727786 | 4061 | ext4_lblk_t *start, ext4_lblk_t *end) |
0830344c OM |
4062 | { |
4063 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
4064 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
38727786 OM |
4065 | struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL; |
4066 | struct rb_node *iter; | |
0830344c | 4067 | ext4_lblk_t new_start, new_end; |
38727786 | 4068 | ext4_lblk_t tmp_pa_start, tmp_pa_end, left_pa_end = -1, right_pa_start = -1; |
0830344c OM |
4069 | |
4070 | new_start = *start; | |
4071 | new_end = *end; | |
4072 | ||
38727786 OM |
4073 | /* |
4074 | * Adjust the normalized range so that it doesn't overlap with any | |
4075 | * existing preallocated blocks(PAs). Make sure to hold the rbtree lock | |
4076 | * so it doesn't change underneath us. | |
4077 | */ | |
4078 | read_lock(&ei->i_prealloc_lock); | |
4079 | ||
4080 | /* Step 1: find any one immediate neighboring PA of the normalized range */ | |
4081 | for (iter = ei->i_prealloc_node.rb_node; iter; | |
4082 | iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, | |
4083 | tmp_pa_start, iter)) { | |
4084 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4085 | pa_node.inode_node); | |
0830344c OM |
4086 | tmp_pa_start = tmp_pa->pa_lstart; |
4087 | tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len); | |
4088 | ||
4089 | /* PA must not overlap original request */ | |
38727786 OM |
4090 | spin_lock(&tmp_pa->pa_lock); |
4091 | if (tmp_pa->pa_deleted == 0) | |
4092 | BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end || | |
4093 | ac->ac_o_ex.fe_logical < tmp_pa_start)); | |
4094 | spin_unlock(&tmp_pa->pa_lock); | |
4095 | } | |
4096 | ||
4097 | /* | |
4098 | * Step 2: check if the found PA is left or right neighbor and | |
4099 | * get the other neighbor | |
4100 | */ | |
4101 | if (tmp_pa) { | |
4102 | if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) { | |
4103 | struct rb_node *tmp; | |
4104 | ||
4105 | left_pa = tmp_pa; | |
4106 | tmp = rb_next(&left_pa->pa_node.inode_node); | |
4107 | if (tmp) { | |
4108 | right_pa = rb_entry(tmp, | |
4109 | struct ext4_prealloc_space, | |
4110 | pa_node.inode_node); | |
4111 | } | |
4112 | } else { | |
4113 | struct rb_node *tmp; | |
4114 | ||
4115 | right_pa = tmp_pa; | |
4116 | tmp = rb_prev(&right_pa->pa_node.inode_node); | |
4117 | if (tmp) { | |
4118 | left_pa = rb_entry(tmp, | |
4119 | struct ext4_prealloc_space, | |
4120 | pa_node.inode_node); | |
4121 | } | |
4122 | } | |
4123 | } | |
4124 | ||
4125 | /* Step 3: get the non deleted neighbors */ | |
4126 | if (left_pa) { | |
4127 | for (iter = &left_pa->pa_node.inode_node;; | |
4128 | iter = rb_prev(iter)) { | |
4129 | if (!iter) { | |
4130 | left_pa = NULL; | |
4131 | break; | |
4132 | } | |
0830344c | 4133 | |
38727786 OM |
4134 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
4135 | pa_node.inode_node); | |
4136 | left_pa = tmp_pa; | |
4137 | spin_lock(&tmp_pa->pa_lock); | |
4138 | if (tmp_pa->pa_deleted == 0) { | |
4139 | spin_unlock(&tmp_pa->pa_lock); | |
4140 | break; | |
4141 | } | |
0830344c | 4142 | spin_unlock(&tmp_pa->pa_lock); |
0830344c | 4143 | } |
38727786 OM |
4144 | } |
4145 | ||
4146 | if (right_pa) { | |
4147 | for (iter = &right_pa->pa_node.inode_node;; | |
4148 | iter = rb_next(iter)) { | |
4149 | if (!iter) { | |
4150 | right_pa = NULL; | |
4151 | break; | |
4152 | } | |
4153 | ||
4154 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4155 | pa_node.inode_node); | |
4156 | right_pa = tmp_pa; | |
4157 | spin_lock(&tmp_pa->pa_lock); | |
4158 | if (tmp_pa->pa_deleted == 0) { | |
4159 | spin_unlock(&tmp_pa->pa_lock); | |
4160 | break; | |
4161 | } | |
4162 | spin_unlock(&tmp_pa->pa_lock); | |
0830344c | 4163 | } |
0830344c | 4164 | } |
38727786 OM |
4165 | |
4166 | if (left_pa) { | |
4167 | left_pa_end = | |
4168 | left_pa->pa_lstart + EXT4_C2B(sbi, left_pa->pa_len); | |
4169 | BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical); | |
4170 | } | |
4171 | ||
4172 | if (right_pa) { | |
4173 | right_pa_start = right_pa->pa_lstart; | |
4174 | BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical); | |
4175 | } | |
4176 | ||
4177 | /* Step 4: trim our normalized range to not overlap with the neighbors */ | |
4178 | if (left_pa) { | |
4179 | if (left_pa_end > new_start) | |
4180 | new_start = left_pa_end; | |
4181 | } | |
4182 | ||
4183 | if (right_pa) { | |
4184 | if (right_pa_start < new_end) | |
4185 | new_end = right_pa_start; | |
4186 | } | |
4187 | read_unlock(&ei->i_prealloc_lock); | |
0830344c OM |
4188 | |
4189 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
4190 | ext4_mb_pa_assert_overlap(ac, new_start, new_end); | |
4191 | ||
4192 | *start = new_start; | |
4193 | *end = new_end; | |
4194 | } | |
4195 | ||
c9de560d AT |
4196 | /* |
4197 | * Normalization means making request better in terms of | |
4198 | * size and alignment | |
4199 | */ | |
4ddfef7b ES |
4200 | static noinline_for_stack void |
4201 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
4202 | struct ext4_allocation_request *ar) |
4203 | { | |
53accfa9 | 4204 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
b07ffe69 | 4205 | struct ext4_super_block *es = sbi->s_es; |
c9de560d AT |
4206 | int bsbits, max; |
4207 | ext4_lblk_t end; | |
1592d2c5 CW |
4208 | loff_t size, start_off; |
4209 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 4210 | ext4_lblk_t start; |
c9de560d AT |
4211 | |
4212 | /* do normalize only data requests, metadata requests | |
4213 | do not need preallocation */ | |
4214 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4215 | return; | |
4216 | ||
4217 | /* sometime caller may want exact blocks */ | |
4218 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4219 | return; | |
4220 | ||
4221 | /* caller may indicate that preallocation isn't | |
4222 | * required (it's a tail, for example) */ | |
4223 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
4224 | return; | |
4225 | ||
4226 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
4227 | ext4_mb_normalize_group_request(ac); | |
4228 | return ; | |
4229 | } | |
4230 | ||
4231 | bsbits = ac->ac_sb->s_blocksize_bits; | |
4232 | ||
4233 | /* first, let's learn actual file size | |
4234 | * given current request is allocated */ | |
53accfa9 | 4235 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
4236 | size = size << bsbits; |
4237 | if (size < i_size_read(ac->ac_inode)) | |
4238 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 4239 | orig_size = size; |
c9de560d | 4240 | |
1930479c VC |
4241 | /* max size of free chunks */ |
4242 | max = 2 << bsbits; | |
c9de560d | 4243 | |
1930479c VC |
4244 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
4245 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
4246 | |
4247 | /* first, try to predict filesize */ | |
4248 | /* XXX: should this table be tunable? */ | |
4249 | start_off = 0; | |
4250 | if (size <= 16 * 1024) { | |
4251 | size = 16 * 1024; | |
4252 | } else if (size <= 32 * 1024) { | |
4253 | size = 32 * 1024; | |
4254 | } else if (size <= 64 * 1024) { | |
4255 | size = 64 * 1024; | |
4256 | } else if (size <= 128 * 1024) { | |
4257 | size = 128 * 1024; | |
4258 | } else if (size <= 256 * 1024) { | |
4259 | size = 256 * 1024; | |
4260 | } else if (size <= 512 * 1024) { | |
4261 | size = 512 * 1024; | |
4262 | } else if (size <= 1024 * 1024) { | |
4263 | size = 1024 * 1024; | |
1930479c | 4264 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 4265 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
4266 | (21 - bsbits)) << 21; |
4267 | size = 2 * 1024 * 1024; | |
4268 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
4269 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4270 | (22 - bsbits)) << 22; | |
4271 | size = 4 * 1024 * 1024; | |
4272 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 4273 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
4274 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4275 | (23 - bsbits)) << 23; | |
4276 | size = 8 * 1024 * 1024; | |
4277 | } else { | |
b27b1535 | 4278 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
91a48aaf | 4279 | size = (loff_t) EXT4_C2B(sbi, |
b27b1535 | 4280 | ac->ac_o_ex.fe_len) << bsbits; |
c9de560d | 4281 | } |
5a0790c2 AK |
4282 | size = size >> bsbits; |
4283 | start = start_off >> bsbits; | |
c9de560d | 4284 | |
a08f789d BL |
4285 | /* |
4286 | * For tiny groups (smaller than 8MB) the chosen allocation | |
4287 | * alignment may be larger than group size. Make sure the | |
4288 | * alignment does not move allocation to a different group which | |
4289 | * makes mballoc fail assertions later. | |
4290 | */ | |
4291 | start = max(start, rounddown(ac->ac_o_ex.fe_logical, | |
4292 | (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); | |
4293 | ||
c9de560d AT |
4294 | /* don't cover already allocated blocks in selected range */ |
4295 | if (ar->pleft && start <= ar->lleft) { | |
4296 | size -= ar->lleft + 1 - start; | |
4297 | start = ar->lleft + 1; | |
4298 | } | |
4299 | if (ar->pright && start + size - 1 >= ar->lright) | |
4300 | size -= start + size - ar->lright; | |
4301 | ||
cd648b8a JK |
4302 | /* |
4303 | * Trim allocation request for filesystems with artificially small | |
4304 | * groups. | |
4305 | */ | |
4306 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
4307 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
4308 | ||
c9de560d AT |
4309 | end = start + size; |
4310 | ||
0830344c | 4311 | ext4_mb_pa_adjust_overlap(ac, &start, &end); |
c9de560d | 4312 | |
c9de560d AT |
4313 | size = end - start; |
4314 | ||
cf4ff938 BL |
4315 | /* |
4316 | * In this function "start" and "size" are normalized for better | |
4317 | * alignment and length such that we could preallocate more blocks. | |
4318 | * This normalization is done such that original request of | |
4319 | * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and | |
4320 | * "size" boundaries. | |
4321 | * (Note fe_len can be relaxed since FS block allocation API does not | |
4322 | * provide gurantee on number of contiguous blocks allocation since that | |
4323 | * depends upon free space left, etc). | |
4324 | * In case of inode pa, later we use the allocated blocks | |
1221b235 | 4325 | * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated |
cf4ff938 BL |
4326 | * range of goal/best blocks [start, size] to put it at the |
4327 | * ac_o_ex.fe_logical extent of this inode. | |
4328 | * (See ext4_mb_use_inode_pa() for more details) | |
4329 | */ | |
4330 | if (start + size <= ac->ac_o_ex.fe_logical || | |
c9de560d | 4331 | start > ac->ac_o_ex.fe_logical) { |
9d8b9ec4 TT |
4332 | ext4_msg(ac->ac_sb, KERN_ERR, |
4333 | "start %lu, size %lu, fe_logical %lu", | |
4334 | (unsigned long) start, (unsigned long) size, | |
4335 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 4336 | BUG(); |
c9de560d | 4337 | } |
b5b60778 | 4338 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
4339 | |
4340 | /* now prepare goal request */ | |
4341 | ||
4342 | /* XXX: is it better to align blocks WRT to logical | |
4343 | * placement or satisfy big request as is */ | |
4344 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 4345 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
4346 | |
4347 | /* define goal start in order to merge */ | |
b07ffe69 KS |
4348 | if (ar->pright && (ar->lright == (start + size)) && |
4349 | ar->pright >= size && | |
4350 | ar->pright - size >= le32_to_cpu(es->s_first_data_block)) { | |
c9de560d AT |
4351 | /* merge to the right */ |
4352 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
b07ffe69 KS |
4353 | &ac->ac_g_ex.fe_group, |
4354 | &ac->ac_g_ex.fe_start); | |
c9de560d AT |
4355 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; |
4356 | } | |
b07ffe69 KS |
4357 | if (ar->pleft && (ar->lleft + 1 == start) && |
4358 | ar->pleft + 1 < ext4_blocks_count(es)) { | |
c9de560d AT |
4359 | /* merge to the left */ |
4360 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
b07ffe69 KS |
4361 | &ac->ac_g_ex.fe_group, |
4362 | &ac->ac_g_ex.fe_start); | |
c9de560d AT |
4363 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; |
4364 | } | |
4365 | ||
d3df1453 RH |
4366 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
4367 | orig_size, start); | |
c9de560d AT |
4368 | } |
4369 | ||
4370 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
4371 | { | |
4372 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4373 | ||
a6c75eaf | 4374 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { |
c9de560d AT |
4375 | atomic_inc(&sbi->s_bal_reqs); |
4376 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 4377 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
4378 | atomic_inc(&sbi->s_bal_success); |
4379 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
a6c75eaf | 4380 | atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); |
c9de560d AT |
4381 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && |
4382 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
4383 | atomic_inc(&sbi->s_bal_goals); | |
4384 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
4385 | atomic_inc(&sbi->s_bal_breaks); | |
4386 | } | |
4387 | ||
296c355c TT |
4388 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
4389 | trace_ext4_mballoc_alloc(ac); | |
4390 | else | |
4391 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
4392 | } |
4393 | ||
b844167e CW |
4394 | /* |
4395 | * Called on failure; free up any blocks from the inode PA for this | |
4396 | * context. We don't need this for MB_GROUP_PA because we only change | |
4397 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
4398 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
4399 | */ | |
4400 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
4401 | { | |
4402 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
4403 | struct ext4_buddy e4b; |
4404 | int err; | |
b844167e | 4405 | |
86f0afd4 | 4406 | if (pa == NULL) { |
c99d1e6e TT |
4407 | if (ac->ac_f_ex.fe_len == 0) |
4408 | return; | |
86f0afd4 | 4409 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
19b8b035 TT |
4410 | if (WARN_RATELIMIT(err, |
4411 | "ext4: mb_load_buddy failed (%d)", err)) | |
86f0afd4 TT |
4412 | /* |
4413 | * This should never happen since we pin the | |
4414 | * pages in the ext4_allocation_context so | |
4415 | * ext4_mb_load_buddy() should never fail. | |
4416 | */ | |
86f0afd4 | 4417 | return; |
86f0afd4 TT |
4418 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); |
4419 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
4420 | ac->ac_f_ex.fe_len); | |
4421 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 4422 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
4423 | return; |
4424 | } | |
36cb0f52 KS |
4425 | if (pa->pa_type == MB_INODE_PA) { |
4426 | spin_lock(&pa->pa_lock); | |
400db9d3 | 4427 | pa->pa_free += ac->ac_b_ex.fe_len; |
36cb0f52 KS |
4428 | spin_unlock(&pa->pa_lock); |
4429 | } | |
b844167e CW |
4430 | } |
4431 | ||
c9de560d AT |
4432 | /* |
4433 | * use blocks preallocated to inode | |
4434 | */ | |
4435 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
4436 | struct ext4_prealloc_space *pa) | |
4437 | { | |
53accfa9 | 4438 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
4439 | ext4_fsblk_t start; |
4440 | ext4_fsblk_t end; | |
4441 | int len; | |
4442 | ||
4443 | /* found preallocated blocks, use them */ | |
4444 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
4445 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
4446 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
4447 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
4448 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
4449 | &ac->ac_b_ex.fe_start); | |
4450 | ac->ac_b_ex.fe_len = len; | |
4451 | ac->ac_status = AC_STATUS_FOUND; | |
4452 | ac->ac_pa = pa; | |
4453 | ||
4454 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 4455 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d | 4456 | BUG_ON(pa->pa_free < len); |
93cdf49f | 4457 | BUG_ON(ac->ac_b_ex.fe_len <= 0); |
c9de560d AT |
4458 | pa->pa_free -= len; |
4459 | ||
d3df1453 | 4460 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
4461 | } |
4462 | ||
4463 | /* | |
4464 | * use blocks preallocated to locality group | |
4465 | */ | |
4466 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
4467 | struct ext4_prealloc_space *pa) | |
4468 | { | |
03cddb80 | 4469 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 4470 | |
c9de560d AT |
4471 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
4472 | &ac->ac_b_ex.fe_group, | |
4473 | &ac->ac_b_ex.fe_start); | |
4474 | ac->ac_b_ex.fe_len = len; | |
4475 | ac->ac_status = AC_STATUS_FOUND; | |
4476 | ac->ac_pa = pa; | |
4477 | ||
1221b235 | 4478 | /* we don't correct pa_pstart or pa_len here to avoid |
26346ff6 | 4479 | * possible race when the group is being loaded concurrently |
c9de560d | 4480 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
4481 | * in on-disk bitmap -- see ext4_mb_release_context() |
4482 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 4483 | */ |
d3df1453 | 4484 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
1afdc588 | 4485 | pa->pa_lstart, len, pa); |
c9de560d AT |
4486 | } |
4487 | ||
5e745b04 AK |
4488 | /* |
4489 | * Return the prealloc space that have minimal distance | |
4490 | * from the goal block. @cpa is the prealloc | |
4491 | * space that is having currently known minimal distance | |
4492 | * from the goal block. | |
4493 | */ | |
4494 | static struct ext4_prealloc_space * | |
4495 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
4496 | struct ext4_prealloc_space *pa, | |
4497 | struct ext4_prealloc_space *cpa) | |
4498 | { | |
4499 | ext4_fsblk_t cur_distance, new_distance; | |
4500 | ||
4501 | if (cpa == NULL) { | |
4502 | atomic_inc(&pa->pa_count); | |
4503 | return pa; | |
4504 | } | |
79211c8e AM |
4505 | cur_distance = abs(goal_block - cpa->pa_pstart); |
4506 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 4507 | |
5a54b2f1 | 4508 | if (cur_distance <= new_distance) |
5e745b04 AK |
4509 | return cpa; |
4510 | ||
4511 | /* drop the previous reference */ | |
4512 | atomic_dec(&cpa->pa_count); | |
4513 | atomic_inc(&pa->pa_count); | |
4514 | return pa; | |
4515 | } | |
4516 | ||
c9de560d AT |
4517 | /* |
4518 | * search goal blocks in preallocated space | |
4519 | */ | |
4fca8f07 | 4520 | static noinline_for_stack bool |
4ddfef7b | 4521 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 4522 | { |
53accfa9 | 4523 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 4524 | int order, i; |
c9de560d AT |
4525 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
4526 | struct ext4_locality_group *lg; | |
bcf43499 OM |
4527 | struct ext4_prealloc_space *tmp_pa, *cpa = NULL; |
4528 | ext4_lblk_t tmp_pa_start, tmp_pa_end; | |
38727786 | 4529 | struct rb_node *iter; |
5e745b04 | 4530 | ext4_fsblk_t goal_block; |
c9de560d AT |
4531 | |
4532 | /* only data can be preallocated */ | |
4533 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 4534 | return false; |
c9de560d AT |
4535 | |
4536 | /* first, try per-file preallocation */ | |
38727786 OM |
4537 | read_lock(&ei->i_prealloc_lock); |
4538 | for (iter = ei->i_prealloc_node.rb_node; iter; | |
4539 | iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, | |
4540 | tmp_pa_start, iter)) { | |
4541 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4542 | pa_node.inode_node); | |
c9de560d AT |
4543 | |
4544 | /* all fields in this condition don't change, | |
4545 | * so we can skip locking for them */ | |
bcf43499 OM |
4546 | tmp_pa_start = tmp_pa->pa_lstart; |
4547 | tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len); | |
4548 | ||
38727786 | 4549 | /* original request start doesn't lie in this PA */ |
bcf43499 OM |
4550 | if (ac->ac_o_ex.fe_logical < tmp_pa_start || |
4551 | ac->ac_o_ex.fe_logical >= tmp_pa_end) | |
c9de560d AT |
4552 | continue; |
4553 | ||
fb0a387d | 4554 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 4555 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
bcf43499 | 4556 | (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) > |
e86a7182 OM |
4557 | EXT4_MAX_BLOCK_FILE_PHYS)) { |
4558 | /* | |
4559 | * Since PAs don't overlap, we won't find any | |
4560 | * other PA to satisfy this. | |
4561 | */ | |
4562 | break; | |
4563 | } | |
fb0a387d | 4564 | |
c9de560d | 4565 | /* found preallocated blocks, use them */ |
bcf43499 OM |
4566 | spin_lock(&tmp_pa->pa_lock); |
4567 | if (tmp_pa->pa_deleted == 0 && tmp_pa->pa_free) { | |
4568 | atomic_inc(&tmp_pa->pa_count); | |
4569 | ext4_mb_use_inode_pa(ac, tmp_pa); | |
4570 | spin_unlock(&tmp_pa->pa_lock); | |
c9de560d | 4571 | ac->ac_criteria = 10; |
38727786 | 4572 | read_unlock(&ei->i_prealloc_lock); |
4fca8f07 | 4573 | return true; |
c9de560d | 4574 | } |
bcf43499 | 4575 | spin_unlock(&tmp_pa->pa_lock); |
c9de560d | 4576 | } |
38727786 | 4577 | read_unlock(&ei->i_prealloc_lock); |
c9de560d AT |
4578 | |
4579 | /* can we use group allocation? */ | |
4580 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 4581 | return false; |
c9de560d AT |
4582 | |
4583 | /* inode may have no locality group for some reason */ | |
4584 | lg = ac->ac_lg; | |
4585 | if (lg == NULL) | |
4fca8f07 | 4586 | return false; |
6be2ded1 AK |
4587 | order = fls(ac->ac_o_ex.fe_len) - 1; |
4588 | if (order > PREALLOC_TB_SIZE - 1) | |
4589 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4590 | order = PREALLOC_TB_SIZE - 1; | |
4591 | ||
bda00de7 | 4592 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
4593 | /* |
4594 | * search for the prealloc space that is having | |
4595 | * minimal distance from the goal block. | |
4596 | */ | |
6be2ded1 AK |
4597 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
4598 | rcu_read_lock(); | |
bcf43499 | 4599 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i], |
a8e38fd3 | 4600 | pa_node.lg_list) { |
bcf43499 OM |
4601 | spin_lock(&tmp_pa->pa_lock); |
4602 | if (tmp_pa->pa_deleted == 0 && | |
4603 | tmp_pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
4604 | |
4605 | cpa = ext4_mb_check_group_pa(goal_block, | |
bcf43499 | 4606 | tmp_pa, cpa); |
6be2ded1 | 4607 | } |
bcf43499 | 4608 | spin_unlock(&tmp_pa->pa_lock); |
c9de560d | 4609 | } |
6be2ded1 | 4610 | rcu_read_unlock(); |
c9de560d | 4611 | } |
5e745b04 AK |
4612 | if (cpa) { |
4613 | ext4_mb_use_group_pa(ac, cpa); | |
4614 | ac->ac_criteria = 20; | |
4fca8f07 | 4615 | return true; |
5e745b04 | 4616 | } |
4fca8f07 | 4617 | return false; |
c9de560d AT |
4618 | } |
4619 | ||
7a2fcbf7 AK |
4620 | /* |
4621 | * the function goes through all block freed in the group | |
4622 | * but not yet committed and marks them used in in-core bitmap. | |
4623 | * buddy must be generated from this bitmap | |
955ce5f5 | 4624 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
4625 | */ |
4626 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
4627 | ext4_group_t group) | |
4628 | { | |
4629 | struct rb_node *n; | |
4630 | struct ext4_group_info *grp; | |
4631 | struct ext4_free_data *entry; | |
4632 | ||
4633 | grp = ext4_get_group_info(sb, group); | |
5354b2af TT |
4634 | if (!grp) |
4635 | return; | |
7a2fcbf7 AK |
4636 | n = rb_first(&(grp->bb_free_root)); |
4637 | ||
4638 | while (n) { | |
18aadd47 | 4639 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
123e3016 | 4640 | mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); |
7a2fcbf7 AK |
4641 | n = rb_next(n); |
4642 | } | |
4643 | return; | |
4644 | } | |
4645 | ||
c9de560d AT |
4646 | /* |
4647 | * the function goes through all preallocation in this group and marks them | |
4648 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 4649 | * Need to be called with ext4 group lock held |
c9de560d | 4650 | */ |
089ceecc ES |
4651 | static noinline_for_stack |
4652 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
4653 | ext4_group_t group) |
4654 | { | |
4655 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4656 | struct ext4_prealloc_space *pa; | |
4657 | struct list_head *cur; | |
4658 | ext4_group_t groupnr; | |
4659 | ext4_grpblk_t start; | |
4660 | int preallocated = 0; | |
c9de560d AT |
4661 | int len; |
4662 | ||
5354b2af TT |
4663 | if (!grp) |
4664 | return; | |
4665 | ||
c9de560d AT |
4666 | /* all form of preallocation discards first load group, |
4667 | * so the only competing code is preallocation use. | |
4668 | * we don't need any locking here | |
4669 | * notice we do NOT ignore preallocations with pa_deleted | |
4670 | * otherwise we could leave used blocks available for | |
4671 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
4672 | * is dropping preallocation | |
4673 | */ | |
4674 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4675 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
4676 | spin_lock(&pa->pa_lock); | |
4677 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4678 | &groupnr, &start); | |
4679 | len = pa->pa_len; | |
4680 | spin_unlock(&pa->pa_lock); | |
4681 | if (unlikely(len == 0)) | |
4682 | continue; | |
4683 | BUG_ON(groupnr != group); | |
123e3016 | 4684 | mb_set_bits(bitmap, start, len); |
c9de560d | 4685 | preallocated += len; |
c9de560d | 4686 | } |
d3df1453 | 4687 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
4688 | } |
4689 | ||
27bc446e | 4690 | static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
4691 | struct ext4_prealloc_space *pa) | |
4692 | { | |
4693 | struct ext4_inode_info *ei; | |
4694 | ||
4695 | if (pa->pa_deleted) { | |
4696 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", | |
4697 | pa->pa_type, pa->pa_pstart, pa->pa_lstart, | |
4698 | pa->pa_len); | |
4699 | return; | |
4700 | } | |
4701 | ||
4702 | pa->pa_deleted = 1; | |
4703 | ||
4704 | if (pa->pa_type == MB_INODE_PA) { | |
4705 | ei = EXT4_I(pa->pa_inode); | |
4706 | atomic_dec(&ei->i_prealloc_active); | |
4707 | } | |
4708 | } | |
4709 | ||
82089725 | 4710 | static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa) |
c9de560d | 4711 | { |
82089725 | 4712 | BUG_ON(!pa); |
4e8d2139 JR |
4713 | BUG_ON(atomic_read(&pa->pa_count)); |
4714 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
4715 | kmem_cache_free(ext4_pspace_cachep, pa); |
4716 | } | |
4717 | ||
82089725 OM |
4718 | static void ext4_mb_pa_callback(struct rcu_head *head) |
4719 | { | |
4720 | struct ext4_prealloc_space *pa; | |
4721 | ||
4722 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4723 | ext4_mb_pa_free(pa); | |
4724 | } | |
4725 | ||
c9de560d AT |
4726 | /* |
4727 | * drops a reference to preallocated space descriptor | |
4728 | * if this was the last reference and the space is consumed | |
4729 | */ | |
4730 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
4731 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
4732 | { | |
a9df9a49 | 4733 | ext4_group_t grp; |
d33a1976 | 4734 | ext4_fsblk_t grp_blk; |
38727786 | 4735 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
c9de560d | 4736 | |
c9de560d AT |
4737 | /* in this short window concurrent discard can set pa_deleted */ |
4738 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
4739 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
4740 | spin_unlock(&pa->pa_lock); | |
4741 | return; | |
4742 | } | |
4743 | ||
c9de560d AT |
4744 | if (pa->pa_deleted == 1) { |
4745 | spin_unlock(&pa->pa_lock); | |
4746 | return; | |
4747 | } | |
4748 | ||
27bc446e | 4749 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4750 | spin_unlock(&pa->pa_lock); |
4751 | ||
d33a1976 | 4752 | grp_blk = pa->pa_pstart; |
60e6679e | 4753 | /* |
cc0fb9ad AK |
4754 | * If doing group-based preallocation, pa_pstart may be in the |
4755 | * next group when pa is used up | |
4756 | */ | |
4757 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
4758 | grp_blk--; |
4759 | ||
bd86298e | 4760 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
4761 | |
4762 | /* | |
4763 | * possible race: | |
4764 | * | |
4765 | * P1 (buddy init) P2 (regular allocation) | |
4766 | * find block B in PA | |
4767 | * copy on-disk bitmap to buddy | |
4768 | * mark B in on-disk bitmap | |
4769 | * drop PA from group | |
4770 | * mark all PAs in buddy | |
4771 | * | |
4772 | * thus, P1 initializes buddy with B available. to prevent this | |
4773 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
4774 | * against that pair | |
4775 | */ | |
4776 | ext4_lock_group(sb, grp); | |
4777 | list_del(&pa->pa_group_list); | |
4778 | ext4_unlock_group(sb, grp); | |
4779 | ||
a8e38fd3 | 4780 | if (pa->pa_type == MB_INODE_PA) { |
38727786 OM |
4781 | write_lock(pa->pa_node_lock.inode_lock); |
4782 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); | |
4783 | write_unlock(pa->pa_node_lock.inode_lock); | |
4784 | ext4_mb_pa_free(pa); | |
a8e38fd3 OM |
4785 | } else { |
4786 | spin_lock(pa->pa_node_lock.lg_lock); | |
4787 | list_del_rcu(&pa->pa_node.lg_list); | |
4788 | spin_unlock(pa->pa_node_lock.lg_lock); | |
38727786 OM |
4789 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
4790 | } | |
4791 | } | |
4792 | ||
4793 | static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new) | |
4794 | { | |
4795 | struct rb_node **iter = &root->rb_node, *parent = NULL; | |
4796 | struct ext4_prealloc_space *iter_pa, *new_pa; | |
4797 | ext4_lblk_t iter_start, new_start; | |
4798 | ||
4799 | while (*iter) { | |
4800 | iter_pa = rb_entry(*iter, struct ext4_prealloc_space, | |
4801 | pa_node.inode_node); | |
4802 | new_pa = rb_entry(new, struct ext4_prealloc_space, | |
4803 | pa_node.inode_node); | |
4804 | iter_start = iter_pa->pa_lstart; | |
4805 | new_start = new_pa->pa_lstart; | |
4806 | ||
4807 | parent = *iter; | |
4808 | if (new_start < iter_start) | |
4809 | iter = &((*iter)->rb_left); | |
4810 | else | |
4811 | iter = &((*iter)->rb_right); | |
a8e38fd3 | 4812 | } |
c9de560d | 4813 | |
38727786 OM |
4814 | rb_link_node(new, parent, iter); |
4815 | rb_insert_color(new, root); | |
c9de560d AT |
4816 | } |
4817 | ||
4818 | /* | |
4819 | * creates new preallocated space for given inode | |
4820 | */ | |
53f86b17 | 4821 | static noinline_for_stack void |
4ddfef7b | 4822 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4823 | { |
4824 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 4825 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
4826 | struct ext4_prealloc_space *pa; |
4827 | struct ext4_group_info *grp; | |
4828 | struct ext4_inode_info *ei; | |
4829 | ||
4830 | /* preallocate only when found space is larger then requested */ | |
4831 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4832 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4833 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4834 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4835 | |
53f86b17 | 4836 | pa = ac->ac_pa; |
c9de560d AT |
4837 | |
4838 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
93cdf49f OM |
4839 | int new_bex_start; |
4840 | int new_bex_end; | |
c9de560d AT |
4841 | |
4842 | /* we can't allocate as much as normalizer wants. | |
4843 | * so, found space must get proper lstart | |
4844 | * to cover original request */ | |
4845 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
4846 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
4847 | ||
93cdf49f OM |
4848 | /* |
4849 | * Use the below logic for adjusting best extent as it keeps | |
4850 | * fragmentation in check while ensuring logical range of best | |
4851 | * extent doesn't overflow out of goal extent: | |
4852 | * | |
4853 | * 1. Check if best ex can be kept at end of goal and still | |
4854 | * cover original start | |
4855 | * 2. Else, check if best ex can be kept at start of goal and | |
4856 | * still cover original start | |
4857 | * 3. Else, keep the best ex at start of original request. | |
4858 | */ | |
4859 | new_bex_end = ac->ac_g_ex.fe_logical + | |
4860 | EXT4_C2B(sbi, ac->ac_g_ex.fe_len); | |
4861 | new_bex_start = new_bex_end - EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
4862 | if (ac->ac_o_ex.fe_logical >= new_bex_start) | |
4863 | goto adjust_bex; | |
c9de560d | 4864 | |
93cdf49f OM |
4865 | new_bex_start = ac->ac_g_ex.fe_logical; |
4866 | new_bex_end = | |
4867 | new_bex_start + EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
4868 | if (ac->ac_o_ex.fe_logical < new_bex_end) | |
4869 | goto adjust_bex; | |
c9de560d | 4870 | |
93cdf49f OM |
4871 | new_bex_start = ac->ac_o_ex.fe_logical; |
4872 | new_bex_end = | |
4873 | new_bex_start + EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d | 4874 | |
93cdf49f OM |
4875 | adjust_bex: |
4876 | ac->ac_b_ex.fe_logical = new_bex_start; | |
c9de560d | 4877 | |
c9de560d AT |
4878 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
4879 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
93cdf49f OM |
4880 | BUG_ON(new_bex_end > (ac->ac_g_ex.fe_logical + |
4881 | EXT4_C2B(sbi, ac->ac_g_ex.fe_len))); | |
c9de560d AT |
4882 | } |
4883 | ||
c9de560d AT |
4884 | pa->pa_lstart = ac->ac_b_ex.fe_logical; |
4885 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4886 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4887 | pa->pa_free = pa->pa_len; | |
c9de560d | 4888 | spin_lock_init(&pa->pa_lock); |
d794bf8e | 4889 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 4890 | pa->pa_deleted = 0; |
cc0fb9ad | 4891 | pa->pa_type = MB_INODE_PA; |
c9de560d | 4892 | |
d3df1453 RH |
4893 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4894 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4895 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d | 4896 | |
53accfa9 | 4897 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
abc075d4 | 4898 | ext4_mb_use_inode_pa(ac, pa); |
c9de560d AT |
4899 | |
4900 | ei = EXT4_I(ac->ac_inode); | |
4901 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
5354b2af TT |
4902 | if (!grp) |
4903 | return; | |
c9de560d | 4904 | |
a8e38fd3 | 4905 | pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock; |
c9de560d AT |
4906 | pa->pa_inode = ac->ac_inode; |
4907 | ||
c9de560d | 4908 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 4909 | |
38727786 OM |
4910 | write_lock(pa->pa_node_lock.inode_lock); |
4911 | ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node); | |
4912 | write_unlock(pa->pa_node_lock.inode_lock); | |
27bc446e | 4913 | atomic_inc(&ei->i_prealloc_active); |
c9de560d AT |
4914 | } |
4915 | ||
4916 | /* | |
4917 | * creates new preallocated space for locality group inodes belongs to | |
4918 | */ | |
53f86b17 | 4919 | static noinline_for_stack void |
4ddfef7b | 4920 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4921 | { |
4922 | struct super_block *sb = ac->ac_sb; | |
4923 | struct ext4_locality_group *lg; | |
4924 | struct ext4_prealloc_space *pa; | |
4925 | struct ext4_group_info *grp; | |
4926 | ||
4927 | /* preallocate only when found space is larger then requested */ | |
4928 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4929 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4930 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4931 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4932 | |
53f86b17 | 4933 | pa = ac->ac_pa; |
c9de560d | 4934 | |
c9de560d AT |
4935 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4936 | pa->pa_lstart = pa->pa_pstart; | |
4937 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4938 | pa->pa_free = pa->pa_len; | |
c9de560d | 4939 | spin_lock_init(&pa->pa_lock); |
a8e38fd3 | 4940 | INIT_LIST_HEAD(&pa->pa_node.lg_list); |
d794bf8e | 4941 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 4942 | pa->pa_deleted = 0; |
cc0fb9ad | 4943 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 4944 | |
d3df1453 RH |
4945 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4946 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4947 | trace_ext4_mb_new_group_pa(ac, pa); |
c9de560d AT |
4948 | |
4949 | ext4_mb_use_group_pa(ac, pa); | |
4950 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
4951 | ||
4952 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
5354b2af TT |
4953 | if (!grp) |
4954 | return; | |
c9de560d AT |
4955 | lg = ac->ac_lg; |
4956 | BUG_ON(lg == NULL); | |
4957 | ||
a8e38fd3 | 4958 | pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock; |
c9de560d AT |
4959 | pa->pa_inode = NULL; |
4960 | ||
c9de560d | 4961 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 4962 | |
6be2ded1 AK |
4963 | /* |
4964 | * We will later add the new pa to the right bucket | |
4965 | * after updating the pa_free in ext4_mb_release_context | |
4966 | */ | |
c9de560d AT |
4967 | } |
4968 | ||
53f86b17 | 4969 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
c9de560d | 4970 | { |
c9de560d | 4971 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
53f86b17 | 4972 | ext4_mb_new_group_pa(ac); |
c9de560d | 4973 | else |
53f86b17 | 4974 | ext4_mb_new_inode_pa(ac); |
c9de560d AT |
4975 | } |
4976 | ||
4977 | /* | |
4978 | * finds all unused blocks in on-disk bitmap, frees them in | |
4979 | * in-core bitmap and buddy. | |
4980 | * @pa must be unlinked from inode and group lists, so that | |
4981 | * nobody else can find/use it. | |
4982 | * the caller MUST hold group/inode locks. | |
4983 | * TODO: optimize the case when there are no in-core structures yet | |
4984 | */ | |
4ddfef7b ES |
4985 | static noinline_for_stack int |
4986 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 4987 | struct ext4_prealloc_space *pa) |
c9de560d | 4988 | { |
c9de560d AT |
4989 | struct super_block *sb = e4b->bd_sb; |
4990 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
4991 | unsigned int end; |
4992 | unsigned int next; | |
c9de560d AT |
4993 | ext4_group_t group; |
4994 | ext4_grpblk_t bit; | |
ba80b101 | 4995 | unsigned long long grp_blk_start; |
c9de560d AT |
4996 | int free = 0; |
4997 | ||
4998 | BUG_ON(pa->pa_deleted == 0); | |
4999 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 5000 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
5001 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
5002 | end = bit + pa->pa_len; | |
5003 | ||
c9de560d | 5004 | while (bit < end) { |
ffad0a44 | 5005 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
5006 | if (bit >= end) |
5007 | break; | |
ffad0a44 | 5008 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
d3df1453 | 5009 | mb_debug(sb, "free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
5010 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
5011 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
5012 | free += next - bit; |
5013 | ||
3e1e5f50 | 5014 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
5015 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
5016 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 5017 | next - bit); |
c9de560d AT |
5018 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
5019 | bit = next + 1; | |
5020 | } | |
5021 | if (free != pa->pa_free) { | |
9d8b9ec4 | 5022 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 5023 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
5024 | pa, (unsigned long) pa->pa_lstart, |
5025 | (unsigned long) pa->pa_pstart, | |
36bad423 | 5026 | pa->pa_len); |
e29136f8 | 5027 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 5028 | free, pa->pa_free); |
e56eb659 AK |
5029 | /* |
5030 | * pa is already deleted so we use the value obtained | |
5031 | * from the bitmap and continue. | |
5032 | */ | |
c9de560d | 5033 | } |
c9de560d AT |
5034 | atomic_add(free, &sbi->s_mb_discarded); |
5035 | ||
863c37fc | 5036 | return 0; |
c9de560d AT |
5037 | } |
5038 | ||
4ddfef7b ES |
5039 | static noinline_for_stack int |
5040 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 5041 | struct ext4_prealloc_space *pa) |
c9de560d | 5042 | { |
c9de560d AT |
5043 | struct super_block *sb = e4b->bd_sb; |
5044 | ext4_group_t group; | |
5045 | ext4_grpblk_t bit; | |
5046 | ||
60e07cf5 | 5047 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
5048 | BUG_ON(pa->pa_deleted == 0); |
5049 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
5050 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
5051 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
5052 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 5053 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
5054 | |
5055 | return 0; | |
5056 | } | |
5057 | ||
5058 | /* | |
5059 | * releases all preallocations in given group | |
5060 | * | |
5061 | * first, we need to decide discard policy: | |
5062 | * - when do we discard | |
5063 | * 1) ENOSPC | |
5064 | * - how many do we discard | |
5065 | * 1) how many requested | |
5066 | */ | |
4ddfef7b ES |
5067 | static noinline_for_stack int |
5068 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
8c80fb31 | 5069 | ext4_group_t group, int *busy) |
c9de560d AT |
5070 | { |
5071 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
5072 | struct buffer_head *bitmap_bh = NULL; | |
5073 | struct ext4_prealloc_space *pa, *tmp; | |
5074 | struct list_head list; | |
5075 | struct ext4_buddy e4b; | |
38727786 | 5076 | struct ext4_inode_info *ei; |
c9de560d | 5077 | int err; |
8c80fb31 | 5078 | int free = 0; |
c9de560d | 5079 | |
5354b2af TT |
5080 | if (!grp) |
5081 | return 0; | |
d3df1453 | 5082 | mb_debug(sb, "discard preallocation for group %u\n", group); |
c9de560d | 5083 | if (list_empty(&grp->bb_prealloc_list)) |
bbc4ec77 | 5084 | goto out_dbg; |
c9de560d | 5085 | |
574ca174 | 5086 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5087 | if (IS_ERR(bitmap_bh)) { |
5088 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5089 | ext4_error_err(sb, -err, |
5090 | "Error %d reading block bitmap for %u", | |
5091 | err, group); | |
bbc4ec77 | 5092 | goto out_dbg; |
c9de560d AT |
5093 | } |
5094 | ||
5095 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 5096 | if (err) { |
9651e6b2 KK |
5097 | ext4_warning(sb, "Error %d loading buddy information for %u", |
5098 | err, group); | |
ce89f46c | 5099 | put_bh(bitmap_bh); |
bbc4ec77 | 5100 | goto out_dbg; |
ce89f46c | 5101 | } |
c9de560d | 5102 | |
c9de560d | 5103 | INIT_LIST_HEAD(&list); |
c9de560d AT |
5104 | ext4_lock_group(sb, group); |
5105 | list_for_each_entry_safe(pa, tmp, | |
5106 | &grp->bb_prealloc_list, pa_group_list) { | |
5107 | spin_lock(&pa->pa_lock); | |
5108 | if (atomic_read(&pa->pa_count)) { | |
5109 | spin_unlock(&pa->pa_lock); | |
8c80fb31 | 5110 | *busy = 1; |
c9de560d AT |
5111 | continue; |
5112 | } | |
5113 | if (pa->pa_deleted) { | |
5114 | spin_unlock(&pa->pa_lock); | |
5115 | continue; | |
5116 | } | |
5117 | ||
5118 | /* seems this one can be freed ... */ | |
27bc446e | 5119 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 5120 | |
70022da8 YB |
5121 | if (!free) |
5122 | this_cpu_inc(discard_pa_seq); | |
5123 | ||
c9de560d AT |
5124 | /* we can trust pa_free ... */ |
5125 | free += pa->pa_free; | |
5126 | ||
5127 | spin_unlock(&pa->pa_lock); | |
5128 | ||
5129 | list_del(&pa->pa_group_list); | |
5130 | list_add(&pa->u.pa_tmp_list, &list); | |
5131 | } | |
5132 | ||
c9de560d AT |
5133 | /* now free all selected PAs */ |
5134 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
5135 | ||
5136 | /* remove from object (inode or locality group) */ | |
a8e38fd3 OM |
5137 | if (pa->pa_type == MB_GROUP_PA) { |
5138 | spin_lock(pa->pa_node_lock.lg_lock); | |
5139 | list_del_rcu(&pa->pa_node.lg_list); | |
5140 | spin_unlock(pa->pa_node_lock.lg_lock); | |
5141 | } else { | |
38727786 OM |
5142 | write_lock(pa->pa_node_lock.inode_lock); |
5143 | ei = EXT4_I(pa->pa_inode); | |
5144 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); | |
5145 | write_unlock(pa->pa_node_lock.inode_lock); | |
a8e38fd3 | 5146 | } |
c9de560d | 5147 | |
38727786 OM |
5148 | list_del(&pa->u.pa_tmp_list); |
5149 | ||
5150 | if (pa->pa_type == MB_GROUP_PA) { | |
3e1e5f50 | 5151 | ext4_mb_release_group_pa(&e4b, pa); |
38727786 OM |
5152 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
5153 | } else { | |
3e1e5f50 | 5154 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
38727786 OM |
5155 | ext4_mb_pa_free(pa); |
5156 | } | |
c9de560d AT |
5157 | } |
5158 | ||
c9de560d | 5159 | ext4_unlock_group(sb, group); |
e39e07fd | 5160 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 5161 | put_bh(bitmap_bh); |
bbc4ec77 | 5162 | out_dbg: |
d3df1453 | 5163 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
8c80fb31 CX |
5164 | free, group, grp->bb_free); |
5165 | return free; | |
c9de560d AT |
5166 | } |
5167 | ||
5168 | /* | |
5169 | * releases all non-used preallocated blocks for given inode | |
5170 | * | |
5171 | * It's important to discard preallocations under i_data_sem | |
5172 | * We don't want another block to be served from the prealloc | |
5173 | * space when we are discarding the inode prealloc space. | |
5174 | * | |
5175 | * FIXME!! Make sure it is valid at all the call sites | |
5176 | */ | |
27bc446e | 5177 | void ext4_discard_preallocations(struct inode *inode, unsigned int needed) |
c9de560d AT |
5178 | { |
5179 | struct ext4_inode_info *ei = EXT4_I(inode); | |
5180 | struct super_block *sb = inode->i_sb; | |
5181 | struct buffer_head *bitmap_bh = NULL; | |
5182 | struct ext4_prealloc_space *pa, *tmp; | |
5183 | ext4_group_t group = 0; | |
5184 | struct list_head list; | |
5185 | struct ext4_buddy e4b; | |
38727786 | 5186 | struct rb_node *iter; |
c9de560d AT |
5187 | int err; |
5188 | ||
c2ea3fde | 5189 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
5190 | return; |
5191 | } | |
5192 | ||
8016e29f HS |
5193 | if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) |
5194 | return; | |
5195 | ||
d3df1453 RH |
5196 | mb_debug(sb, "discard preallocation for inode %lu\n", |
5197 | inode->i_ino); | |
27bc446e | 5198 | trace_ext4_discard_preallocations(inode, |
5199 | atomic_read(&ei->i_prealloc_active), needed); | |
c9de560d AT |
5200 | |
5201 | INIT_LIST_HEAD(&list); | |
5202 | ||
27bc446e | 5203 | if (needed == 0) |
5204 | needed = UINT_MAX; | |
5205 | ||
c9de560d AT |
5206 | repeat: |
5207 | /* first, collect all pa's in the inode */ | |
38727786 OM |
5208 | write_lock(&ei->i_prealloc_lock); |
5209 | for (iter = rb_first(&ei->i_prealloc_node); iter && needed; | |
5210 | iter = rb_next(iter)) { | |
5211 | pa = rb_entry(iter, struct ext4_prealloc_space, | |
5212 | pa_node.inode_node); | |
a8e38fd3 | 5213 | BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock); |
38727786 | 5214 | |
c9de560d AT |
5215 | spin_lock(&pa->pa_lock); |
5216 | if (atomic_read(&pa->pa_count)) { | |
5217 | /* this shouldn't happen often - nobody should | |
5218 | * use preallocation while we're discarding it */ | |
5219 | spin_unlock(&pa->pa_lock); | |
38727786 | 5220 | write_unlock(&ei->i_prealloc_lock); |
9d8b9ec4 TT |
5221 | ext4_msg(sb, KERN_ERR, |
5222 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
5223 | WARN_ON(1); |
5224 | schedule_timeout_uninterruptible(HZ); | |
5225 | goto repeat; | |
5226 | ||
5227 | } | |
5228 | if (pa->pa_deleted == 0) { | |
27bc446e | 5229 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 5230 | spin_unlock(&pa->pa_lock); |
38727786 | 5231 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); |
c9de560d | 5232 | list_add(&pa->u.pa_tmp_list, &list); |
27bc446e | 5233 | needed--; |
c9de560d AT |
5234 | continue; |
5235 | } | |
5236 | ||
5237 | /* someone is deleting pa right now */ | |
5238 | spin_unlock(&pa->pa_lock); | |
38727786 | 5239 | write_unlock(&ei->i_prealloc_lock); |
c9de560d AT |
5240 | |
5241 | /* we have to wait here because pa_deleted | |
5242 | * doesn't mean pa is already unlinked from | |
5243 | * the list. as we might be called from | |
5244 | * ->clear_inode() the inode will get freed | |
5245 | * and concurrent thread which is unlinking | |
5246 | * pa from inode's list may access already | |
5247 | * freed memory, bad-bad-bad */ | |
5248 | ||
5249 | /* XXX: if this happens too often, we can | |
5250 | * add a flag to force wait only in case | |
5251 | * of ->clear_inode(), but not in case of | |
5252 | * regular truncate */ | |
5253 | schedule_timeout_uninterruptible(HZ); | |
5254 | goto repeat; | |
5255 | } | |
38727786 | 5256 | write_unlock(&ei->i_prealloc_lock); |
c9de560d AT |
5257 | |
5258 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 5259 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 5260 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 5261 | |
9651e6b2 KK |
5262 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5263 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 5264 | if (err) { |
54d3adbc TT |
5265 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5266 | err, group); | |
ce89f46c AK |
5267 | continue; |
5268 | } | |
c9de560d | 5269 | |
574ca174 | 5270 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5271 | if (IS_ERR(bitmap_bh)) { |
5272 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5273 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
5274 | err, group); | |
e39e07fd | 5275 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 5276 | continue; |
c9de560d AT |
5277 | } |
5278 | ||
5279 | ext4_lock_group(sb, group); | |
5280 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5281 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
5282 | ext4_unlock_group(sb, group); |
5283 | ||
e39e07fd | 5284 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
5285 | put_bh(bitmap_bh); |
5286 | ||
5287 | list_del(&pa->u.pa_tmp_list); | |
38727786 | 5288 | ext4_mb_pa_free(pa); |
c9de560d AT |
5289 | } |
5290 | } | |
5291 | ||
53f86b17 RH |
5292 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
5293 | { | |
5294 | struct ext4_prealloc_space *pa; | |
5295 | ||
5296 | BUG_ON(ext4_pspace_cachep == NULL); | |
5297 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); | |
5298 | if (!pa) | |
5299 | return -ENOMEM; | |
5300 | atomic_set(&pa->pa_count, 1); | |
5301 | ac->ac_pa = pa; | |
5302 | return 0; | |
5303 | } | |
5304 | ||
82089725 | 5305 | static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac) |
53f86b17 RH |
5306 | { |
5307 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
5308 | ||
5309 | BUG_ON(!pa); | |
5310 | ac->ac_pa = NULL; | |
5311 | WARN_ON(!atomic_dec_and_test(&pa->pa_count)); | |
82089725 OM |
5312 | /* |
5313 | * current function is only called due to an error or due to | |
5314 | * len of found blocks < len of requested blocks hence the PA has not | |
5315 | * been added to grp->bb_prealloc_list. So we don't need to lock it | |
5316 | */ | |
5317 | pa->pa_deleted = 1; | |
5318 | ext4_mb_pa_free(pa); | |
53f86b17 RH |
5319 | } |
5320 | ||
6ba495e9 | 5321 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 5322 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 5323 | { |
e68cf40c | 5324 | ext4_group_t i, ngroups; |
c9de560d | 5325 | |
9b5f6c9b | 5326 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e3570639 ES |
5327 | return; |
5328 | ||
8df9675f | 5329 | ngroups = ext4_get_groups_count(sb); |
d3df1453 | 5330 | mb_debug(sb, "groups: "); |
8df9675f | 5331 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
5332 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
5333 | struct ext4_prealloc_space *pa; | |
5334 | ext4_grpblk_t start; | |
5335 | struct list_head *cur; | |
5354b2af TT |
5336 | |
5337 | if (!grp) | |
5338 | continue; | |
c9de560d AT |
5339 | ext4_lock_group(sb, i); |
5340 | list_for_each(cur, &grp->bb_prealloc_list) { | |
5341 | pa = list_entry(cur, struct ext4_prealloc_space, | |
5342 | pa_group_list); | |
5343 | spin_lock(&pa->pa_lock); | |
5344 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
5345 | NULL, &start); | |
5346 | spin_unlock(&pa->pa_lock); | |
d3df1453 RH |
5347 | mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
5348 | pa->pa_len); | |
c9de560d | 5349 | } |
60bd63d1 | 5350 | ext4_unlock_group(sb, i); |
d3df1453 RH |
5351 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
5352 | grp->bb_fragments); | |
c9de560d | 5353 | } |
c9de560d | 5354 | } |
e68cf40c RH |
5355 | |
5356 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
5357 | { | |
5358 | struct super_block *sb = ac->ac_sb; | |
5359 | ||
9b5f6c9b | 5360 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e68cf40c RH |
5361 | return; |
5362 | ||
d3df1453 | 5363 | mb_debug(sb, "Can't allocate:" |
e68cf40c | 5364 | " Allocation context details:"); |
d3df1453 | 5365 | mb_debug(sb, "status %u flags 0x%x", |
e68cf40c | 5366 | ac->ac_status, ac->ac_flags); |
d3df1453 | 5367 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
e68cf40c RH |
5368 | "goal %lu/%lu/%lu@%lu, " |
5369 | "best %lu/%lu/%lu@%lu cr %d", | |
5370 | (unsigned long)ac->ac_o_ex.fe_group, | |
5371 | (unsigned long)ac->ac_o_ex.fe_start, | |
5372 | (unsigned long)ac->ac_o_ex.fe_len, | |
5373 | (unsigned long)ac->ac_o_ex.fe_logical, | |
5374 | (unsigned long)ac->ac_g_ex.fe_group, | |
5375 | (unsigned long)ac->ac_g_ex.fe_start, | |
5376 | (unsigned long)ac->ac_g_ex.fe_len, | |
5377 | (unsigned long)ac->ac_g_ex.fe_logical, | |
5378 | (unsigned long)ac->ac_b_ex.fe_group, | |
5379 | (unsigned long)ac->ac_b_ex.fe_start, | |
5380 | (unsigned long)ac->ac_b_ex.fe_len, | |
5381 | (unsigned long)ac->ac_b_ex.fe_logical, | |
5382 | (int)ac->ac_criteria); | |
d3df1453 | 5383 | mb_debug(sb, "%u found", ac->ac_found); |
e68cf40c RH |
5384 | ext4_mb_show_pa(sb); |
5385 | } | |
c9de560d | 5386 | #else |
e68cf40c RH |
5387 | static inline void ext4_mb_show_pa(struct super_block *sb) |
5388 | { | |
5389 | return; | |
5390 | } | |
c9de560d AT |
5391 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
5392 | { | |
e68cf40c | 5393 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
5394 | return; |
5395 | } | |
5396 | #endif | |
5397 | ||
5398 | /* | |
5399 | * We use locality group preallocation for small size file. The size of the | |
5400 | * file is determined by the current size or the resulting size after | |
5401 | * allocation which ever is larger | |
5402 | * | |
b713a5ec | 5403 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
5404 | */ |
5405 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
5406 | { | |
5407 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
5408 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
5409 | loff_t size, isize; | |
a9f2a293 | 5410 | bool inode_pa_eligible, group_pa_eligible; |
c9de560d AT |
5411 | |
5412 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
5413 | return; | |
5414 | ||
4ba74d00 TT |
5415 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
5416 | return; | |
5417 | ||
a9f2a293 JK |
5418 | group_pa_eligible = sbi->s_mb_group_prealloc > 0; |
5419 | inode_pa_eligible = true; | |
53accfa9 | 5420 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
5421 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
5422 | >> bsbits; | |
c9de560d | 5423 | |
a9f2a293 | 5424 | /* No point in using inode preallocation for closed files */ |
82dd124c | 5425 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
a9f2a293 JK |
5426 | !inode_is_open_for_write(ac->ac_inode)) |
5427 | inode_pa_eligible = false; | |
50797481 | 5428 | |
71780577 | 5429 | size = max(size, isize); |
a9f2a293 JK |
5430 | /* Don't use group allocation for large files */ |
5431 | if (size > sbi->s_mb_stream_request) | |
5432 | group_pa_eligible = false; | |
5433 | ||
5434 | if (!group_pa_eligible) { | |
5435 | if (inode_pa_eligible) | |
5436 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
5437 | else | |
5438 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
c9de560d | 5439 | return; |
4ba74d00 | 5440 | } |
c9de560d AT |
5441 | |
5442 | BUG_ON(ac->ac_lg != NULL); | |
5443 | /* | |
5444 | * locality group prealloc space are per cpu. The reason for having | |
5445 | * per cpu locality group is to reduce the contention between block | |
5446 | * request from multiple CPUs. | |
5447 | */ | |
a0b6bc63 | 5448 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
5449 | |
5450 | /* we're going to use group allocation */ | |
5451 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
5452 | ||
5453 | /* serialize all allocations in the group */ | |
5454 | mutex_lock(&ac->ac_lg->lg_mutex); | |
5455 | } | |
5456 | ||
d73eff68 | 5457 | static noinline_for_stack void |
4ddfef7b | 5458 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, |
c9de560d AT |
5459 | struct ext4_allocation_request *ar) |
5460 | { | |
5461 | struct super_block *sb = ar->inode->i_sb; | |
5462 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
5463 | struct ext4_super_block *es = sbi->s_es; | |
5464 | ext4_group_t group; | |
498e5f24 TT |
5465 | unsigned int len; |
5466 | ext4_fsblk_t goal; | |
c9de560d AT |
5467 | ext4_grpblk_t block; |
5468 | ||
5469 | /* we can't allocate > group size */ | |
5470 | len = ar->len; | |
5471 | ||
5472 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
5473 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
5474 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
5475 | |
5476 | /* start searching from the goal */ | |
5477 | goal = ar->goal; | |
5478 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5479 | goal >= ext4_blocks_count(es)) | |
5480 | goal = le32_to_cpu(es->s_first_data_block); | |
5481 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
5482 | ||
5483 | /* set up allocation goals */ | |
f5a44db5 | 5484 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 5485 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
5486 | ac->ac_sb = sb; |
5487 | ac->ac_inode = ar->inode; | |
53accfa9 | 5488 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
5489 | ac->ac_o_ex.fe_group = group; |
5490 | ac->ac_o_ex.fe_start = block; | |
5491 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 5492 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 5493 | ac->ac_flags = ar->flags; |
c9de560d | 5494 | |
3cb77bd2 | 5495 | /* we have to define context: we'll work with a file or |
c9de560d AT |
5496 | * locality group. this is a policy, actually */ |
5497 | ext4_mb_group_or_file(ac); | |
5498 | ||
d3df1453 | 5499 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
c9de560d AT |
5500 | "left: %u/%u, right %u/%u to %swritable\n", |
5501 | (unsigned) ar->len, (unsigned) ar->logical, | |
5502 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
5503 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
5504 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 5505 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
5506 | } |
5507 | ||
6be2ded1 AK |
5508 | static noinline_for_stack void |
5509 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
5510 | struct ext4_locality_group *lg, | |
5511 | int order, int total_entries) | |
5512 | { | |
5513 | ext4_group_t group = 0; | |
5514 | struct ext4_buddy e4b; | |
5515 | struct list_head discard_list; | |
5516 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 5517 | |
d3df1453 | 5518 | mb_debug(sb, "discard locality group preallocation\n"); |
6be2ded1 AK |
5519 | |
5520 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
5521 | |
5522 | spin_lock(&lg->lg_prealloc_lock); | |
5523 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
a8e38fd3 | 5524 | pa_node.lg_list, |
92e9c58c | 5525 | lockdep_is_held(&lg->lg_prealloc_lock)) { |
6be2ded1 AK |
5526 | spin_lock(&pa->pa_lock); |
5527 | if (atomic_read(&pa->pa_count)) { | |
5528 | /* | |
5529 | * This is the pa that we just used | |
5530 | * for block allocation. So don't | |
5531 | * free that | |
5532 | */ | |
5533 | spin_unlock(&pa->pa_lock); | |
5534 | continue; | |
5535 | } | |
5536 | if (pa->pa_deleted) { | |
5537 | spin_unlock(&pa->pa_lock); | |
5538 | continue; | |
5539 | } | |
5540 | /* only lg prealloc space */ | |
cc0fb9ad | 5541 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
5542 | |
5543 | /* seems this one can be freed ... */ | |
27bc446e | 5544 | ext4_mb_mark_pa_deleted(sb, pa); |
6be2ded1 AK |
5545 | spin_unlock(&pa->pa_lock); |
5546 | ||
a8e38fd3 | 5547 | list_del_rcu(&pa->pa_node.lg_list); |
6be2ded1 AK |
5548 | list_add(&pa->u.pa_tmp_list, &discard_list); |
5549 | ||
5550 | total_entries--; | |
5551 | if (total_entries <= 5) { | |
5552 | /* | |
5553 | * we want to keep only 5 entries | |
5554 | * allowing it to grow to 8. This | |
5555 | * mak sure we don't call discard | |
5556 | * soon for this list. | |
5557 | */ | |
5558 | break; | |
5559 | } | |
5560 | } | |
5561 | spin_unlock(&lg->lg_prealloc_lock); | |
5562 | ||
5563 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 5564 | int err; |
6be2ded1 | 5565 | |
bd86298e | 5566 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
5567 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5568 | GFP_NOFS|__GFP_NOFAIL); | |
5569 | if (err) { | |
54d3adbc TT |
5570 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5571 | err, group); | |
6be2ded1 AK |
5572 | continue; |
5573 | } | |
5574 | ext4_lock_group(sb, group); | |
5575 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5576 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
5577 | ext4_unlock_group(sb, group); |
5578 | ||
e39e07fd | 5579 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
5580 | list_del(&pa->u.pa_tmp_list); |
5581 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5582 | } | |
6be2ded1 AK |
5583 | } |
5584 | ||
5585 | /* | |
5586 | * We have incremented pa_count. So it cannot be freed at this | |
5587 | * point. Also we hold lg_mutex. So no parallel allocation is | |
5588 | * possible from this lg. That means pa_free cannot be updated. | |
5589 | * | |
5590 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
5591 | * which can cause the lg_prealloc_list to be updated. | |
5592 | */ | |
5593 | ||
5594 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
5595 | { | |
5596 | int order, added = 0, lg_prealloc_count = 1; | |
5597 | struct super_block *sb = ac->ac_sb; | |
5598 | struct ext4_locality_group *lg = ac->ac_lg; | |
5599 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
5600 | ||
5601 | order = fls(pa->pa_free) - 1; | |
5602 | if (order > PREALLOC_TB_SIZE - 1) | |
5603 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
5604 | order = PREALLOC_TB_SIZE - 1; | |
5605 | /* Add the prealloc space to lg */ | |
f1167009 | 5606 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 5607 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
a8e38fd3 | 5608 | pa_node.lg_list, |
92e9c58c | 5609 | lockdep_is_held(&lg->lg_prealloc_lock)) { |
6be2ded1 AK |
5610 | spin_lock(&tmp_pa->pa_lock); |
5611 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 5612 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
5613 | continue; |
5614 | } | |
5615 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
5616 | /* Add to the tail of the previous entry */ | |
a8e38fd3 OM |
5617 | list_add_tail_rcu(&pa->pa_node.lg_list, |
5618 | &tmp_pa->pa_node.lg_list); | |
6be2ded1 AK |
5619 | added = 1; |
5620 | /* | |
5621 | * we want to count the total | |
5622 | * number of entries in the list | |
5623 | */ | |
5624 | } | |
5625 | spin_unlock(&tmp_pa->pa_lock); | |
5626 | lg_prealloc_count++; | |
5627 | } | |
5628 | if (!added) | |
a8e38fd3 | 5629 | list_add_tail_rcu(&pa->pa_node.lg_list, |
6be2ded1 | 5630 | &lg->lg_prealloc_list[order]); |
f1167009 | 5631 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
5632 | |
5633 | /* Now trim the list to be not more than 8 elements */ | |
5634 | if (lg_prealloc_count > 8) { | |
5635 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 5636 | order, lg_prealloc_count); |
6be2ded1 AK |
5637 | return; |
5638 | } | |
5639 | return ; | |
5640 | } | |
5641 | ||
c9de560d AT |
5642 | /* |
5643 | * release all resource we used in allocation | |
5644 | */ | |
5645 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
5646 | { | |
53accfa9 | 5647 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
5648 | struct ext4_prealloc_space *pa = ac->ac_pa; |
5649 | if (pa) { | |
cc0fb9ad | 5650 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 5651 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 5652 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
5653 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
5654 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
5655 | pa->pa_free -= ac->ac_b_ex.fe_len; |
5656 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
5657 | spin_unlock(&pa->pa_lock); | |
66d5e027 | 5658 | |
5659 | /* | |
5660 | * We want to add the pa to the right bucket. | |
5661 | * Remove it from the list and while adding | |
5662 | * make sure the list to which we are adding | |
5663 | * doesn't grow big. | |
5664 | */ | |
5665 | if (likely(pa->pa_free)) { | |
a8e38fd3 OM |
5666 | spin_lock(pa->pa_node_lock.lg_lock); |
5667 | list_del_rcu(&pa->pa_node.lg_list); | |
5668 | spin_unlock(pa->pa_node_lock.lg_lock); | |
66d5e027 | 5669 | ext4_mb_add_n_trim(ac); |
5670 | } | |
ba443916 | 5671 | } |
27bc446e | 5672 | |
ba443916 AK |
5673 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
5674 | } | |
c9de560d | 5675 | if (ac->ac_bitmap_page) |
09cbfeaf | 5676 | put_page(ac->ac_bitmap_page); |
c9de560d | 5677 | if (ac->ac_buddy_page) |
09cbfeaf | 5678 | put_page(ac->ac_buddy_page); |
c9de560d AT |
5679 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
5680 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
5681 | ext4_mb_collect_stats(ac); | |
5682 | return 0; | |
5683 | } | |
5684 | ||
5685 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
5686 | { | |
8df9675f | 5687 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d | 5688 | int ret; |
8c80fb31 CX |
5689 | int freed = 0, busy = 0; |
5690 | int retry = 0; | |
c9de560d | 5691 | |
9bffad1e | 5692 | trace_ext4_mb_discard_preallocations(sb, needed); |
8c80fb31 CX |
5693 | |
5694 | if (needed == 0) | |
5695 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; | |
5696 | repeat: | |
8df9675f | 5697 | for (i = 0; i < ngroups && needed > 0; i++) { |
8c80fb31 | 5698 | ret = ext4_mb_discard_group_preallocations(sb, i, &busy); |
c9de560d AT |
5699 | freed += ret; |
5700 | needed -= ret; | |
8c80fb31 CX |
5701 | cond_resched(); |
5702 | } | |
5703 | ||
5704 | if (needed > 0 && busy && ++retry < 3) { | |
5705 | busy = 0; | |
5706 | goto repeat; | |
c9de560d AT |
5707 | } |
5708 | ||
5709 | return freed; | |
5710 | } | |
5711 | ||
cf5e2ca6 | 5712 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
07b5b8e1 | 5713 | struct ext4_allocation_context *ac, u64 *seq) |
cf5e2ca6 RH |
5714 | { |
5715 | int freed; | |
07b5b8e1 RH |
5716 | u64 seq_retry = 0; |
5717 | bool ret = false; | |
cf5e2ca6 RH |
5718 | |
5719 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
07b5b8e1 RH |
5720 | if (freed) { |
5721 | ret = true; | |
5722 | goto out_dbg; | |
5723 | } | |
5724 | seq_retry = ext4_get_discard_pa_seq_sum(); | |
99377830 RH |
5725 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
5726 | ac->ac_flags |= EXT4_MB_STRICT_CHECK; | |
07b5b8e1 RH |
5727 | *seq = seq_retry; |
5728 | ret = true; | |
5729 | } | |
5730 | ||
5731 | out_dbg: | |
5732 | mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); | |
5733 | return ret; | |
cf5e2ca6 RH |
5734 | } |
5735 | ||
8016e29f HS |
5736 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, |
5737 | struct ext4_allocation_request *ar, int *errp); | |
5738 | ||
c9de560d AT |
5739 | /* |
5740 | * Main entry point into mballoc to allocate blocks | |
5741 | * it tries to use preallocation first, then falls back | |
5742 | * to usual allocation | |
5743 | */ | |
5744 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 5745 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 5746 | { |
256bdb49 | 5747 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
5748 | struct ext4_sb_info *sbi; |
5749 | struct super_block *sb; | |
5750 | ext4_fsblk_t block = 0; | |
60e58e0f | 5751 | unsigned int inquota = 0; |
53accfa9 | 5752 | unsigned int reserv_clstrs = 0; |
80fa46d6 | 5753 | int retries = 0; |
07b5b8e1 | 5754 | u64 seq; |
c9de560d | 5755 | |
b10a44c3 | 5756 | might_sleep(); |
c9de560d AT |
5757 | sb = ar->inode->i_sb; |
5758 | sbi = EXT4_SB(sb); | |
5759 | ||
9bffad1e | 5760 | trace_ext4_request_blocks(ar); |
8016e29f HS |
5761 | if (sbi->s_mount_state & EXT4_FC_REPLAY) |
5762 | return ext4_mb_new_blocks_simple(handle, ar, errp); | |
ba80b101 | 5763 | |
45dc63e7 | 5764 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 5765 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
5766 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
5767 | ||
e3cf5d5d | 5768 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
5769 | /* Without delayed allocation we need to verify |
5770 | * there is enough free blocks to do block allocation | |
5771 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 5772 | */ |
55f020db | 5773 | while (ar->len && |
e7d5f315 | 5774 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 5775 | |
030ba6bc | 5776 | /* let others to free the space */ |
bb8b20ed | 5777 | cond_resched(); |
030ba6bc AK |
5778 | ar->len = ar->len >> 1; |
5779 | } | |
5780 | if (!ar->len) { | |
bbc4ec77 | 5781 | ext4_mb_show_pa(sb); |
a30d542a AK |
5782 | *errp = -ENOSPC; |
5783 | return 0; | |
5784 | } | |
53accfa9 | 5785 | reserv_clstrs = ar->len; |
55f020db | 5786 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
5787 | dquot_alloc_block_nofail(ar->inode, |
5788 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
5789 | } else { |
5790 | while (ar->len && | |
53accfa9 TT |
5791 | dquot_alloc_block(ar->inode, |
5792 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
5793 | |
5794 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
5795 | ar->len--; | |
5796 | } | |
60e58e0f MC |
5797 | } |
5798 | inquota = ar->len; | |
5799 | if (ar->len == 0) { | |
5800 | *errp = -EDQUOT; | |
6c7a120a | 5801 | goto out; |
60e58e0f | 5802 | } |
07031431 | 5803 | } |
d2a17637 | 5804 | |
85556c9a | 5805 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 5806 | if (!ac) { |
363d4251 | 5807 | ar->len = 0; |
256bdb49 | 5808 | *errp = -ENOMEM; |
6c7a120a | 5809 | goto out; |
256bdb49 ES |
5810 | } |
5811 | ||
d73eff68 | 5812 | ext4_mb_initialize_context(ac, ar); |
c9de560d | 5813 | |
256bdb49 | 5814 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
81198536 | 5815 | seq = this_cpu_read(discard_pa_seq); |
256bdb49 | 5816 | if (!ext4_mb_use_preallocated(ac)) { |
256bdb49 ES |
5817 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
5818 | ext4_mb_normalize_request(ac, ar); | |
53f86b17 RH |
5819 | |
5820 | *errp = ext4_mb_pa_alloc(ac); | |
5821 | if (*errp) | |
5822 | goto errout; | |
c9de560d AT |
5823 | repeat: |
5824 | /* allocate space in core */ | |
6c7a120a | 5825 | *errp = ext4_mb_regular_allocator(ac); |
53f86b17 RH |
5826 | /* |
5827 | * pa allocated above is added to grp->bb_prealloc_list only | |
5828 | * when we were able to allocate some block i.e. when | |
5829 | * ac->ac_status == AC_STATUS_FOUND. | |
5830 | * And error from above mean ac->ac_status != AC_STATUS_FOUND | |
5831 | * So we have to free this pa here itself. | |
5832 | */ | |
2c00ef3e | 5833 | if (*errp) { |
82089725 | 5834 | ext4_mb_pa_put_free(ac); |
2c00ef3e AK |
5835 | ext4_discard_allocated_blocks(ac); |
5836 | goto errout; | |
5837 | } | |
53f86b17 RH |
5838 | if (ac->ac_status == AC_STATUS_FOUND && |
5839 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) | |
82089725 | 5840 | ext4_mb_pa_put_free(ac); |
c9de560d | 5841 | } |
256bdb49 | 5842 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 5843 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 5844 | if (*errp) { |
b844167e | 5845 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
5846 | goto errout; |
5847 | } else { | |
519deca0 AK |
5848 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
5849 | ar->len = ac->ac_b_ex.fe_len; | |
5850 | } | |
c9de560d | 5851 | } else { |
80fa46d6 TT |
5852 | if (++retries < 3 && |
5853 | ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) | |
c9de560d | 5854 | goto repeat; |
53f86b17 RH |
5855 | /* |
5856 | * If block allocation fails then the pa allocated above | |
5857 | * needs to be freed here itself. | |
5858 | */ | |
82089725 | 5859 | ext4_mb_pa_put_free(ac); |
c9de560d | 5860 | *errp = -ENOSPC; |
6c7a120a AK |
5861 | } |
5862 | ||
5863 | if (*errp) { | |
aaae558d | 5864 | errout: |
256bdb49 | 5865 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 5866 | ar->len = 0; |
256bdb49 | 5867 | ext4_mb_show_ac(ac); |
c9de560d | 5868 | } |
256bdb49 | 5869 | ext4_mb_release_context(ac); |
aaae558d | 5870 | kmem_cache_free(ext4_ac_cachep, ac); |
6c7a120a | 5871 | out: |
60e58e0f | 5872 | if (inquota && ar->len < inquota) |
53accfa9 | 5873 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 5874 | if (!ar->len) { |
e3cf5d5d | 5875 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 5876 | /* release all the reserved blocks if non delalloc */ |
57042651 | 5877 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 5878 | reserv_clstrs); |
0087d9fb | 5879 | } |
c9de560d | 5880 | |
9bffad1e | 5881 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 5882 | |
c9de560d AT |
5883 | return block; |
5884 | } | |
c9de560d | 5885 | |
c894058d AK |
5886 | /* |
5887 | * We can merge two free data extents only if the physical blocks | |
5888 | * are contiguous, AND the extents were freed by the same transaction, | |
5889 | * AND the blocks are associated with the same group. | |
5890 | */ | |
a0154344 DJ |
5891 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
5892 | struct ext4_free_data *entry, | |
5893 | struct ext4_free_data *new_entry, | |
5894 | struct rb_root *entry_rb_root) | |
c894058d | 5895 | { |
a0154344 DJ |
5896 | if ((entry->efd_tid != new_entry->efd_tid) || |
5897 | (entry->efd_group != new_entry->efd_group)) | |
5898 | return; | |
5899 | if (entry->efd_start_cluster + entry->efd_count == | |
5900 | new_entry->efd_start_cluster) { | |
5901 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
5902 | new_entry->efd_count += entry->efd_count; | |
5903 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
5904 | entry->efd_start_cluster) { | |
5905 | new_entry->efd_count += entry->efd_count; | |
5906 | } else | |
5907 | return; | |
5908 | spin_lock(&sbi->s_md_lock); | |
5909 | list_del(&entry->efd_list); | |
5910 | spin_unlock(&sbi->s_md_lock); | |
5911 | rb_erase(&entry->efd_node, entry_rb_root); | |
5912 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
5913 | } |
5914 | ||
85b67ffb | 5915 | static noinline_for_stack void |
4ddfef7b | 5916 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, |
7a2fcbf7 | 5917 | struct ext4_free_data *new_entry) |
c9de560d | 5918 | { |
e29136f8 | 5919 | ext4_group_t group = e4b->bd_group; |
84130193 | 5920 | ext4_grpblk_t cluster; |
d08854f5 | 5921 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 5922 | struct ext4_free_data *entry; |
c9de560d AT |
5923 | struct ext4_group_info *db = e4b->bd_info; |
5924 | struct super_block *sb = e4b->bd_sb; | |
5925 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
5926 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
5927 | struct rb_node *parent = NULL, *new_node; | |
5928 | ||
0390131b | 5929 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
5930 | BUG_ON(e4b->bd_bitmap_page == NULL); |
5931 | BUG_ON(e4b->bd_buddy_page == NULL); | |
5932 | ||
18aadd47 BJ |
5933 | new_node = &new_entry->efd_node; |
5934 | cluster = new_entry->efd_start_cluster; | |
c894058d | 5935 | |
c894058d AK |
5936 | if (!*n) { |
5937 | /* first free block exent. We need to | |
5938 | protect buddy cache from being freed, | |
5939 | * otherwise we'll refresh it from | |
5940 | * on-disk bitmap and lose not-yet-available | |
5941 | * blocks */ | |
09cbfeaf KS |
5942 | get_page(e4b->bd_buddy_page); |
5943 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
5944 | } |
5945 | while (*n) { | |
5946 | parent = *n; | |
18aadd47 BJ |
5947 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
5948 | if (cluster < entry->efd_start_cluster) | |
c894058d | 5949 | n = &(*n)->rb_left; |
18aadd47 | 5950 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
5951 | n = &(*n)->rb_right; |
5952 | else { | |
e29136f8 | 5953 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
5954 | ext4_group_first_block_no(sb, group) + |
5955 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 5956 | "Block already on to-be-freed list"); |
cca41553 | 5957 | kmem_cache_free(ext4_free_data_cachep, new_entry); |
85b67ffb | 5958 | return; |
c9de560d | 5959 | } |
c894058d | 5960 | } |
c9de560d | 5961 | |
c894058d AK |
5962 | rb_link_node(new_node, parent, n); |
5963 | rb_insert_color(new_node, &db->bb_free_root); | |
5964 | ||
5965 | /* Now try to see the extent can be merged to left and right */ | |
5966 | node = rb_prev(new_node); | |
5967 | if (node) { | |
18aadd47 | 5968 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5969 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5970 | &(db->bb_free_root)); | |
c894058d | 5971 | } |
c9de560d | 5972 | |
c894058d AK |
5973 | node = rb_next(new_node); |
5974 | if (node) { | |
18aadd47 | 5975 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5976 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5977 | &(db->bb_free_root)); | |
c9de560d | 5978 | } |
a0154344 | 5979 | |
d08854f5 | 5980 | spin_lock(&sbi->s_md_lock); |
a0154344 | 5981 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
5982 | sbi->s_mb_free_pending += clusters; |
5983 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
5984 | } |
5985 | ||
8016e29f HS |
5986 | /* |
5987 | * Simple allocator for Ext4 fast commit replay path. It searches for blocks | |
5988 | * linearly starting at the goal block and also excludes the blocks which | |
5989 | * are going to be in use after fast commit replay. | |
5990 | */ | |
5991 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, | |
5992 | struct ext4_allocation_request *ar, int *errp) | |
5993 | { | |
5994 | struct buffer_head *bitmap_bh; | |
5995 | struct super_block *sb = ar->inode->i_sb; | |
5996 | ext4_group_t group; | |
5997 | ext4_grpblk_t blkoff; | |
31a074a0 XY |
5998 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
5999 | ext4_grpblk_t i = 0; | |
8016e29f HS |
6000 | ext4_fsblk_t goal, block; |
6001 | struct ext4_super_block *es = EXT4_SB(sb)->s_es; | |
6002 | ||
6003 | goal = ar->goal; | |
6004 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
6005 | goal >= ext4_blocks_count(es)) | |
6006 | goal = le32_to_cpu(es->s_first_data_block); | |
6007 | ||
6008 | ar->len = 0; | |
6009 | ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); | |
6010 | for (; group < ext4_get_groups_count(sb); group++) { | |
6011 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
6012 | if (IS_ERR(bitmap_bh)) { | |
6013 | *errp = PTR_ERR(bitmap_bh); | |
6014 | pr_warn("Failed to read block bitmap\n"); | |
6015 | return 0; | |
6016 | } | |
6017 | ||
31a074a0 XY |
6018 | while (1) { |
6019 | i = mb_find_next_zero_bit(bitmap_bh->b_data, max, | |
8016e29f | 6020 | blkoff); |
31a074a0 XY |
6021 | if (i >= max) |
6022 | break; | |
6023 | if (ext4_fc_replay_check_excluded(sb, | |
6024 | ext4_group_first_block_no(sb, group) + i)) { | |
6025 | blkoff = i + 1; | |
6026 | } else | |
6027 | break; | |
6028 | } | |
8016e29f | 6029 | brelse(bitmap_bh); |
31a074a0 XY |
6030 | if (i < max) |
6031 | break; | |
253cacb0 KS |
6032 | |
6033 | blkoff = 0; | |
8016e29f HS |
6034 | } |
6035 | ||
31a074a0 XY |
6036 | if (group >= ext4_get_groups_count(sb) || i >= max) { |
6037 | *errp = -ENOSPC; | |
8016e29f | 6038 | return 0; |
31a074a0 | 6039 | } |
8016e29f HS |
6040 | |
6041 | block = ext4_group_first_block_no(sb, group) + i; | |
6042 | ext4_mb_mark_bb(sb, block, 1, 1); | |
6043 | ar->len = 1; | |
6044 | ||
6045 | return block; | |
6046 | } | |
6047 | ||
6048 | static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, | |
6049 | unsigned long count) | |
6050 | { | |
6051 | struct buffer_head *bitmap_bh; | |
6052 | struct super_block *sb = inode->i_sb; | |
6053 | struct ext4_group_desc *gdp; | |
6054 | struct buffer_head *gdp_bh; | |
6055 | ext4_group_t group; | |
6056 | ext4_grpblk_t blkoff; | |
6057 | int already_freed = 0, err, i; | |
6058 | ||
6059 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
6060 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
6061 | if (IS_ERR(bitmap_bh)) { | |
8016e29f HS |
6062 | pr_warn("Failed to read block bitmap\n"); |
6063 | return; | |
6064 | } | |
6065 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
6066 | if (!gdp) | |
1b5c9d34 | 6067 | goto err_out; |
8016e29f HS |
6068 | |
6069 | for (i = 0; i < count; i++) { | |
6070 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data)) | |
6071 | already_freed++; | |
6072 | } | |
6073 | mb_clear_bits(bitmap_bh->b_data, blkoff, count); | |
6074 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
6075 | if (err) | |
1b5c9d34 | 6076 | goto err_out; |
8016e29f HS |
6077 | ext4_free_group_clusters_set( |
6078 | sb, gdp, ext4_free_group_clusters(sb, gdp) + | |
6079 | count - already_freed); | |
1df9bde4 | 6080 | ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); |
8016e29f HS |
6081 | ext4_group_desc_csum_set(sb, group, gdp); |
6082 | ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
6083 | sync_dirty_buffer(bitmap_bh); | |
6084 | sync_dirty_buffer(gdp_bh); | |
1b5c9d34 KS |
6085 | |
6086 | err_out: | |
8016e29f HS |
6087 | brelse(bitmap_bh); |
6088 | } | |
6089 | ||
44338711 | 6090 | /** |
8ac3939d RH |
6091 | * ext4_mb_clear_bb() -- helper function for freeing blocks. |
6092 | * Used by ext4_free_blocks() | |
44338711 TT |
6093 | * @handle: handle for this transaction |
6094 | * @inode: inode | |
c60990b3 TT |
6095 | * @block: starting physical block to be freed |
6096 | * @count: number of blocks to be freed | |
5def1360 | 6097 | * @flags: flags used by ext4_free_blocks |
c9de560d | 6098 | */ |
8ac3939d RH |
6099 | static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, |
6100 | ext4_fsblk_t block, unsigned long count, | |
6101 | int flags) | |
c9de560d | 6102 | { |
26346ff6 | 6103 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 6104 | struct super_block *sb = inode->i_sb; |
c9de560d | 6105 | struct ext4_group_desc *gdp; |
5354b2af | 6106 | struct ext4_group_info *grp; |
498e5f24 | 6107 | unsigned int overflow; |
c9de560d AT |
6108 | ext4_grpblk_t bit; |
6109 | struct buffer_head *gd_bh; | |
6110 | ext4_group_t block_group; | |
6111 | struct ext4_sb_info *sbi; | |
6112 | struct ext4_buddy e4b; | |
84130193 | 6113 | unsigned int count_clusters; |
c9de560d AT |
6114 | int err = 0; |
6115 | int ret; | |
6116 | ||
8016e29f HS |
6117 | sbi = EXT4_SB(sb); |
6118 | ||
1e1c2b86 LC |
6119 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
6120 | !ext4_inode_block_valid(inode, block, count)) { | |
6121 | ext4_error(sb, "Freeing blocks in system zone - " | |
6122 | "Block = %llu, count = %lu", block, count); | |
6123 | /* err = 0. ext4_std_error should be a no op */ | |
6124 | goto error_return; | |
6125 | } | |
6126 | flags |= EXT4_FREE_BLOCKS_VALIDATED; | |
6127 | ||
c9de560d AT |
6128 | do_more: |
6129 | overflow = 0; | |
6130 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
6131 | ||
5354b2af TT |
6132 | grp = ext4_get_group_info(sb, block_group); |
6133 | if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
163a203d DW |
6134 | return; |
6135 | ||
c9de560d AT |
6136 | /* |
6137 | * Check to see if we are freeing blocks across a group | |
6138 | * boundary. | |
6139 | */ | |
84130193 TT |
6140 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
6141 | overflow = EXT4_C2B(sbi, bit) + count - | |
6142 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d | 6143 | count -= overflow; |
1e1c2b86 LC |
6144 | /* The range changed so it's no longer validated */ |
6145 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d | 6146 | } |
810da240 | 6147 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 6148 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
6149 | if (IS_ERR(bitmap_bh)) { |
6150 | err = PTR_ERR(bitmap_bh); | |
6151 | bitmap_bh = NULL; | |
c9de560d | 6152 | goto error_return; |
ce89f46c | 6153 | } |
c9de560d | 6154 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
6155 | if (!gdp) { |
6156 | err = -EIO; | |
c9de560d | 6157 | goto error_return; |
ce89f46c | 6158 | } |
c9de560d | 6159 | |
1e1c2b86 LC |
6160 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
6161 | !ext4_inode_block_valid(inode, block, count)) { | |
12062ddd | 6162 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 6163 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
6164 | /* err = 0. ext4_std_error should be a no op */ |
6165 | goto error_return; | |
c9de560d AT |
6166 | } |
6167 | ||
6168 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
188c299e JK |
6169 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
6170 | EXT4_JTR_NONE); | |
c9de560d AT |
6171 | if (err) |
6172 | goto error_return; | |
6173 | ||
6174 | /* | |
6175 | * We are about to modify some metadata. Call the journal APIs | |
6176 | * to unshare ->b_data if a currently-committing transaction is | |
6177 | * using it | |
6178 | */ | |
6179 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 6180 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
c9de560d AT |
6181 | if (err) |
6182 | goto error_return; | |
c9de560d AT |
6183 | #ifdef AGGRESSIVE_CHECK |
6184 | { | |
6185 | int i; | |
84130193 | 6186 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
6187 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
6188 | } | |
6189 | #endif | |
84130193 | 6190 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 6191 | |
adb7ef60 KK |
6192 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
6193 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
6194 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
6195 | if (err) |
6196 | goto error_return; | |
e6362609 | 6197 | |
f96c450d DJ |
6198 | /* |
6199 | * We need to make sure we don't reuse the freed block until after the | |
6200 | * transaction is committed. We make an exception if the inode is to be | |
6201 | * written in writeback mode since writeback mode has weak data | |
6202 | * consistency guarantees. | |
6203 | */ | |
6204 | if (ext4_handle_valid(handle) && | |
6205 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
6206 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
6207 | struct ext4_free_data *new_entry; |
6208 | /* | |
7444a072 MH |
6209 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
6210 | * to fail. | |
7a2fcbf7 | 6211 | */ |
7444a072 MH |
6212 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
6213 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
6214 | new_entry->efd_start_cluster = bit; |
6215 | new_entry->efd_group = block_group; | |
6216 | new_entry->efd_count = count_clusters; | |
6217 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 6218 | |
7a2fcbf7 | 6219 | ext4_lock_group(sb, block_group); |
84130193 | 6220 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 6221 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 6222 | } else { |
7a2fcbf7 AK |
6223 | /* need to update group_info->bb_free and bitmap |
6224 | * with group lock held. generate_buddy look at | |
6225 | * them with group lock_held | |
6226 | */ | |
d71c1ae2 | 6227 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
6228 | err = ext4_issue_discard(sb, block_group, bit, count, |
6229 | NULL); | |
d71c1ae2 LC |
6230 | if (err && err != -EOPNOTSUPP) |
6231 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
a00b482b | 6232 | " group:%u block:%d count:%lu failed" |
d71c1ae2 LC |
6233 | " with %d", block_group, bit, count, |
6234 | err); | |
8f9ff189 LC |
6235 | } else |
6236 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 6237 | |
955ce5f5 | 6238 | ext4_lock_group(sb, block_group); |
84130193 TT |
6239 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
6240 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
6241 | } |
6242 | ||
021b65bb TT |
6243 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
6244 | ext4_free_group_clusters_set(sb, gdp, ret); | |
1df9bde4 | 6245 | ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); |
feb0ab32 | 6246 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 6247 | ext4_unlock_group(sb, block_group); |
c9de560d | 6248 | |
772cb7c8 JS |
6249 | if (sbi->s_log_groups_per_flex) { |
6250 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 6251 | atomic64_add(count_clusters, |
7c990728 SJS |
6252 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6253 | flex_group)->free_clusters); | |
772cb7c8 JS |
6254 | } |
6255 | ||
9fe67149 EW |
6256 | /* |
6257 | * on a bigalloc file system, defer the s_freeclusters_counter | |
6258 | * update to the caller (ext4_remove_space and friends) so they | |
6259 | * can determine if a cluster freed here should be rereserved | |
6260 | */ | |
6261 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
6262 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
6263 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
6264 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
6265 | count_clusters); | |
6266 | } | |
7d734532 JK |
6267 | |
6268 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 6269 | |
7a2fcbf7 AK |
6270 | /* We dirtied the bitmap block */ |
6271 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6272 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6273 | ||
c9de560d AT |
6274 | /* And the group descriptor block */ |
6275 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 6276 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
6277 | if (!err) |
6278 | err = ret; | |
6279 | ||
6280 | if (overflow && !err) { | |
6281 | block += count; | |
6282 | count = overflow; | |
6283 | put_bh(bitmap_bh); | |
1e1c2b86 LC |
6284 | /* The range changed so it's no longer validated */ |
6285 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d AT |
6286 | goto do_more; |
6287 | } | |
c9de560d AT |
6288 | error_return: |
6289 | brelse(bitmap_bh); | |
6290 | ext4_std_error(sb, err); | |
6291 | return; | |
6292 | } | |
7360d173 | 6293 | |
8ac3939d RH |
6294 | /** |
6295 | * ext4_free_blocks() -- Free given blocks and update quota | |
6296 | * @handle: handle for this transaction | |
6297 | * @inode: inode | |
6298 | * @bh: optional buffer of the block to be freed | |
6299 | * @block: starting physical block to be freed | |
6300 | * @count: number of blocks to be freed | |
6301 | * @flags: flags used by ext4_free_blocks | |
6302 | */ | |
6303 | void ext4_free_blocks(handle_t *handle, struct inode *inode, | |
6304 | struct buffer_head *bh, ext4_fsblk_t block, | |
6305 | unsigned long count, int flags) | |
6306 | { | |
6307 | struct super_block *sb = inode->i_sb; | |
6308 | unsigned int overflow; | |
6309 | struct ext4_sb_info *sbi; | |
6310 | ||
6311 | sbi = EXT4_SB(sb); | |
6312 | ||
6313 | if (sbi->s_mount_state & EXT4_FC_REPLAY) { | |
6314 | ext4_free_blocks_simple(inode, block, count); | |
6315 | return; | |
6316 | } | |
6317 | ||
6318 | might_sleep(); | |
6319 | if (bh) { | |
6320 | if (block) | |
6321 | BUG_ON(block != bh->b_blocknr); | |
6322 | else | |
6323 | block = bh->b_blocknr; | |
6324 | } | |
6325 | ||
6326 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && | |
6327 | !ext4_inode_block_valid(inode, block, count)) { | |
6328 | ext4_error(sb, "Freeing blocks not in datazone - " | |
6329 | "block = %llu, count = %lu", block, count); | |
6330 | return; | |
6331 | } | |
1e1c2b86 | 6332 | flags |= EXT4_FREE_BLOCKS_VALIDATED; |
8ac3939d RH |
6333 | |
6334 | ext4_debug("freeing block %llu\n", block); | |
6335 | trace_ext4_free_blocks(inode, block, count, flags); | |
6336 | ||
6337 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6338 | BUG_ON(count > 1); | |
6339 | ||
6340 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
6341 | inode, bh, block); | |
6342 | } | |
6343 | ||
6344 | /* | |
6345 | * If the extent to be freed does not begin on a cluster | |
6346 | * boundary, we need to deal with partial clusters at the | |
6347 | * beginning and end of the extent. Normally we will free | |
6348 | * blocks at the beginning or the end unless we are explicitly | |
6349 | * requested to avoid doing so. | |
6350 | */ | |
6351 | overflow = EXT4_PBLK_COFF(sbi, block); | |
6352 | if (overflow) { | |
6353 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
6354 | overflow = sbi->s_cluster_ratio - overflow; | |
6355 | block += overflow; | |
6356 | if (count > overflow) | |
6357 | count -= overflow; | |
6358 | else | |
6359 | return; | |
6360 | } else { | |
6361 | block -= overflow; | |
6362 | count += overflow; | |
6363 | } | |
1e1c2b86 LC |
6364 | /* The range changed so it's no longer validated */ |
6365 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6366 | } |
6367 | overflow = EXT4_LBLK_COFF(sbi, count); | |
6368 | if (overflow) { | |
6369 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
6370 | if (count > overflow) | |
6371 | count -= overflow; | |
6372 | else | |
6373 | return; | |
6374 | } else | |
6375 | count += sbi->s_cluster_ratio - overflow; | |
1e1c2b86 LC |
6376 | /* The range changed so it's no longer validated */ |
6377 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6378 | } |
6379 | ||
6380 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6381 | int i; | |
6382 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; | |
6383 | ||
6384 | for (i = 0; i < count; i++) { | |
6385 | cond_resched(); | |
6386 | if (is_metadata) | |
6387 | bh = sb_find_get_block(inode->i_sb, block + i); | |
6388 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
6389 | } | |
6390 | } | |
6391 | ||
6392 | ext4_mb_clear_bb(handle, inode, block, count, flags); | |
6393 | return; | |
6394 | } | |
6395 | ||
2846e820 | 6396 | /** |
0529155e | 6397 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
6398 | * @handle: handle to this transaction |
6399 | * @sb: super block | |
4907cb7b | 6400 | * @block: start physical block to add to the block group |
2846e820 AG |
6401 | * @count: number of blocks to free |
6402 | * | |
e73a347b | 6403 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 6404 | */ |
cc7365df | 6405 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
6406 | ext4_fsblk_t block, unsigned long count) |
6407 | { | |
6408 | struct buffer_head *bitmap_bh = NULL; | |
6409 | struct buffer_head *gd_bh; | |
6410 | ext4_group_t block_group; | |
6411 | ext4_grpblk_t bit; | |
6412 | unsigned int i; | |
6413 | struct ext4_group_desc *desc; | |
6414 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 6415 | struct ext4_buddy e4b; |
d77147ff | 6416 | int err = 0, ret, free_clusters_count; |
6417 | ext4_grpblk_t clusters_freed; | |
6418 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
6419 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
6420 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
6421 | |
6422 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
6423 | ||
4740b830 YY |
6424 | if (count == 0) |
6425 | return 0; | |
6426 | ||
2846e820 | 6427 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
6428 | /* |
6429 | * Check to see if we are freeing blocks across a group | |
6430 | * boundary. | |
6431 | */ | |
d77147ff | 6432 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
6433 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
6434 | block_group); |
6435 | err = -EINVAL; | |
2846e820 | 6436 | goto error_return; |
cc7365df | 6437 | } |
2cd05cc3 | 6438 | |
2846e820 | 6439 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
6440 | if (IS_ERR(bitmap_bh)) { |
6441 | err = PTR_ERR(bitmap_bh); | |
6442 | bitmap_bh = NULL; | |
2846e820 | 6443 | goto error_return; |
cc7365df YY |
6444 | } |
6445 | ||
2846e820 | 6446 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
6447 | if (!desc) { |
6448 | err = -EIO; | |
2846e820 | 6449 | goto error_return; |
cc7365df | 6450 | } |
2846e820 | 6451 | |
a00b482b | 6452 | if (!ext4_sb_block_valid(sb, NULL, block, count)) { |
2846e820 AG |
6453 | ext4_error(sb, "Adding blocks in system zones - " |
6454 | "Block = %llu, count = %lu", | |
6455 | block, count); | |
cc7365df | 6456 | err = -EINVAL; |
2846e820 AG |
6457 | goto error_return; |
6458 | } | |
6459 | ||
2cd05cc3 | 6460 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
6461 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
6462 | EXT4_JTR_NONE); | |
2846e820 AG |
6463 | if (err) |
6464 | goto error_return; | |
6465 | ||
6466 | /* | |
6467 | * We are about to modify some metadata. Call the journal APIs | |
6468 | * to unshare ->b_data if a currently-committing transaction is | |
6469 | * using it | |
6470 | */ | |
6471 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 6472 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
2846e820 AG |
6473 | if (err) |
6474 | goto error_return; | |
e73a347b | 6475 | |
d77147ff | 6476 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 6477 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 6478 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
6479 | ext4_error(sb, "bit already cleared for block %llu", |
6480 | (ext4_fsblk_t)(block + i)); | |
6481 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
6482 | } else { | |
d77147ff | 6483 | clusters_freed++; |
2846e820 AG |
6484 | } |
6485 | } | |
e73a347b AG |
6486 | |
6487 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
6488 | if (err) | |
6489 | goto error_return; | |
6490 | ||
6491 | /* | |
6492 | * need to update group_info->bb_free and bitmap | |
6493 | * with group lock held. generate_buddy look at | |
6494 | * them with group lock_held | |
6495 | */ | |
2846e820 | 6496 | ext4_lock_group(sb, block_group); |
d77147ff | 6497 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
6498 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
6499 | free_clusters_count = clusters_freed + | |
6500 | ext4_free_group_clusters(sb, desc); | |
6501 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
1df9bde4 | 6502 | ext4_block_bitmap_csum_set(sb, desc, bitmap_bh); |
feb0ab32 | 6503 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 6504 | ext4_unlock_group(sb, block_group); |
57042651 | 6505 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 6506 | clusters_freed); |
2846e820 AG |
6507 | |
6508 | if (sbi->s_log_groups_per_flex) { | |
6509 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 6510 | atomic64_add(clusters_freed, |
7c990728 SJS |
6511 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6512 | flex_group)->free_clusters); | |
2846e820 | 6513 | } |
e73a347b AG |
6514 | |
6515 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
6516 | |
6517 | /* We dirtied the bitmap block */ | |
6518 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6519 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6520 | ||
6521 | /* And the group descriptor block */ | |
6522 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
6523 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
6524 | if (!err) | |
6525 | err = ret; | |
6526 | ||
6527 | error_return: | |
6528 | brelse(bitmap_bh); | |
6529 | ext4_std_error(sb, err); | |
cc7365df | 6530 | return err; |
2846e820 AG |
6531 | } |
6532 | ||
7360d173 LC |
6533 | /** |
6534 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
6535 | * @sb: super block for the file system | |
6536 | * @start: starting block of the free extent in the alloc. group | |
6537 | * @count: number of blocks to TRIM | |
7360d173 LC |
6538 | * @e4b: ext4 buddy for the group |
6539 | * | |
6540 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
6541 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
6542 | * be called with under the group lock. | |
6543 | */ | |
bd2eea8d WJ |
6544 | static int ext4_trim_extent(struct super_block *sb, |
6545 | int start, int count, struct ext4_buddy *e4b) | |
e2cbd587 | 6546 | __releases(bitlock) |
6547 | __acquires(bitlock) | |
7360d173 LC |
6548 | { |
6549 | struct ext4_free_extent ex; | |
bd2eea8d | 6550 | ext4_group_t group = e4b->bd_group; |
d71c1ae2 | 6551 | int ret = 0; |
7360d173 | 6552 | |
b3d4c2b1 TM |
6553 | trace_ext4_trim_extent(sb, group, start, count); |
6554 | ||
7360d173 LC |
6555 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
6556 | ||
6557 | ex.fe_start = start; | |
6558 | ex.fe_group = group; | |
6559 | ex.fe_len = count; | |
6560 | ||
6561 | /* | |
6562 | * Mark blocks used, so no one can reuse them while | |
6563 | * being trimmed. | |
6564 | */ | |
6565 | mb_mark_used(e4b, &ex); | |
6566 | ext4_unlock_group(sb, group); | |
a0154344 | 6567 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
6568 | ext4_lock_group(sb, group); |
6569 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 6570 | return ret; |
7360d173 LC |
6571 | } |
6572 | ||
6920b391 WJ |
6573 | static int ext4_try_to_trim_range(struct super_block *sb, |
6574 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
6575 | ext4_grpblk_t max, ext4_grpblk_t minblocks) | |
a5fda113 TT |
6576 | __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) |
6577 | __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) | |
6920b391 WJ |
6578 | { |
6579 | ext4_grpblk_t next, count, free_count; | |
6580 | void *bitmap; | |
6920b391 WJ |
6581 | |
6582 | bitmap = e4b->bd_bitmap; | |
6583 | start = (e4b->bd_info->bb_first_free > start) ? | |
6584 | e4b->bd_info->bb_first_free : start; | |
6585 | count = 0; | |
6586 | free_count = 0; | |
6587 | ||
6588 | while (start <= max) { | |
6589 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
6590 | if (start > max) | |
6591 | break; | |
6592 | next = mb_find_next_bit(bitmap, max + 1, start); | |
6593 | ||
6594 | if ((next - start) >= minblocks) { | |
afcc4e32 LB |
6595 | int ret = ext4_trim_extent(sb, start, next - start, e4b); |
6596 | ||
6920b391 WJ |
6597 | if (ret && ret != -EOPNOTSUPP) |
6598 | break; | |
6920b391 WJ |
6599 | count += next - start; |
6600 | } | |
6601 | free_count += next - start; | |
6602 | start = next + 1; | |
6603 | ||
6604 | if (fatal_signal_pending(current)) { | |
6605 | count = -ERESTARTSYS; | |
6606 | break; | |
6607 | } | |
6608 | ||
6609 | if (need_resched()) { | |
6610 | ext4_unlock_group(sb, e4b->bd_group); | |
6611 | cond_resched(); | |
6612 | ext4_lock_group(sb, e4b->bd_group); | |
6613 | } | |
6614 | ||
6615 | if ((e4b->bd_info->bb_free - free_count) < minblocks) | |
6616 | break; | |
6617 | } | |
6618 | ||
6619 | return count; | |
6620 | } | |
6621 | ||
7360d173 LC |
6622 | /** |
6623 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
6624 | * @sb: super block for file system | |
22612283 | 6625 | * @group: group to be trimmed |
7360d173 LC |
6626 | * @start: first group block to examine |
6627 | * @max: last group block to examine | |
6628 | * @minblocks: minimum extent block count | |
d63c00ea | 6629 | * @set_trimmed: set the trimmed flag if at least one block is trimmed |
7360d173 | 6630 | * |
7360d173 LC |
6631 | * ext4_trim_all_free walks through group's block bitmap searching for free |
6632 | * extents. When the free extent is found, mark it as used in group buddy | |
6633 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
b6f5558c | 6634 | * the group buddy bitmap. |
7360d173 | 6635 | */ |
0b75a840 | 6636 | static ext4_grpblk_t |
78944086 LC |
6637 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
6638 | ext4_grpblk_t start, ext4_grpblk_t max, | |
d63c00ea | 6639 | ext4_grpblk_t minblocks, bool set_trimmed) |
7360d173 | 6640 | { |
78944086 | 6641 | struct ext4_buddy e4b; |
6920b391 | 6642 | int ret; |
7360d173 | 6643 | |
b3d4c2b1 TM |
6644 | trace_ext4_trim_all_free(sb, group, start, max); |
6645 | ||
78944086 LC |
6646 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
6647 | if (ret) { | |
9651e6b2 KK |
6648 | ext4_warning(sb, "Error %d loading buddy information for %u", |
6649 | ret, group); | |
78944086 LC |
6650 | return ret; |
6651 | } | |
28739eea LC |
6652 | |
6653 | ext4_lock_group(sb, group); | |
7360d173 | 6654 | |
6920b391 | 6655 | if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || |
2327fb2e | 6656 | minblocks < EXT4_SB(sb)->s_last_trim_minblks) { |
6920b391 | 6657 | ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); |
d63c00ea | 6658 | if (ret >= 0 && set_trimmed) |
6920b391 WJ |
6659 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
6660 | } else { | |
6661 | ret = 0; | |
7360d173 | 6662 | } |
3d56b8d2 | 6663 | |
7360d173 | 6664 | ext4_unlock_group(sb, group); |
78944086 | 6665 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
6666 | |
6667 | ext4_debug("trimmed %d blocks in the group %d\n", | |
6920b391 | 6668 | ret, group); |
7360d173 | 6669 | |
d71c1ae2 | 6670 | return ret; |
7360d173 LC |
6671 | } |
6672 | ||
6673 | /** | |
6674 | * ext4_trim_fs() -- trim ioctl handle function | |
6675 | * @sb: superblock for filesystem | |
6676 | * @range: fstrim_range structure | |
6677 | * | |
6678 | * start: First Byte to trim | |
6679 | * len: number of Bytes to trim from start | |
6680 | * minlen: minimum extent length in Bytes | |
6681 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
6682 | * start to start+len. For each such a group ext4_trim_all_free function | |
6683 | * is invoked to trim all free space. | |
6684 | */ | |
6685 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
6686 | { | |
7b47ef52 | 6687 | unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); |
78944086 | 6688 | struct ext4_group_info *grp; |
913eed83 | 6689 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 6690 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 6691 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
6692 | ext4_fsblk_t first_data_blk = |
6693 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 6694 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
d63c00ea | 6695 | bool whole_group, eof = false; |
7360d173 LC |
6696 | int ret = 0; |
6697 | ||
6698 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 6699 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
6700 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
6701 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 6702 | |
5de35e8d LC |
6703 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
6704 | start >= max_blks || | |
6705 | range->len < sb->s_blocksize) | |
7360d173 | 6706 | return -EINVAL; |
173b6e38 | 6707 | /* No point to try to trim less than discard granularity */ |
7b47ef52 | 6708 | if (range->minlen < discard_granularity) { |
173b6e38 | 6709 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
7b47ef52 | 6710 | discard_granularity >> sb->s_blocksize_bits); |
173b6e38 JK |
6711 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) |
6712 | goto out; | |
6713 | } | |
d63c00ea | 6714 | if (end >= max_blks - 1) { |
913eed83 | 6715 | end = max_blks - 1; |
d63c00ea DM |
6716 | eof = true; |
6717 | } | |
913eed83 | 6718 | if (end <= first_data_blk) |
22f10457 | 6719 | goto out; |
913eed83 | 6720 | if (start < first_data_blk) |
0f0a25bf | 6721 | start = first_data_blk; |
7360d173 | 6722 | |
913eed83 | 6723 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 6724 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 6725 | &first_group, &first_cluster); |
913eed83 | 6726 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 6727 | &last_group, &last_cluster); |
7360d173 | 6728 | |
913eed83 LC |
6729 | /* end now represents the last cluster to discard in this group */ |
6730 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
d63c00ea | 6731 | whole_group = true; |
7360d173 LC |
6732 | |
6733 | for (group = first_group; group <= last_group; group++) { | |
78944086 | 6734 | grp = ext4_get_group_info(sb, group); |
5354b2af TT |
6735 | if (!grp) |
6736 | continue; | |
78944086 LC |
6737 | /* We only do this if the grp has never been initialized */ |
6738 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 6739 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
6740 | if (ret) |
6741 | break; | |
7360d173 LC |
6742 | } |
6743 | ||
0ba08517 | 6744 | /* |
913eed83 LC |
6745 | * For all the groups except the last one, last cluster will |
6746 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
6747 | * change it for the last group, note that last_cluster is | |
6748 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 6749 | */ |
d63c00ea | 6750 | if (group == last_group) { |
913eed83 | 6751 | end = last_cluster; |
d63c00ea DM |
6752 | whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1; |
6753 | } | |
78944086 | 6754 | if (grp->bb_free >= minlen) { |
7137d7a4 | 6755 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
d63c00ea | 6756 | end, minlen, whole_group); |
7360d173 LC |
6757 | if (cnt < 0) { |
6758 | ret = cnt; | |
7360d173 LC |
6759 | break; |
6760 | } | |
21e7fd22 | 6761 | trimmed += cnt; |
7360d173 | 6762 | } |
913eed83 LC |
6763 | |
6764 | /* | |
6765 | * For every group except the first one, we are sure | |
6766 | * that the first cluster to discard will be cluster #0. | |
6767 | */ | |
7137d7a4 | 6768 | first_cluster = 0; |
7360d173 | 6769 | } |
7360d173 | 6770 | |
3d56b8d2 | 6771 | if (!ret) |
2327fb2e | 6772 | EXT4_SB(sb)->s_last_trim_minblks = minlen; |
3d56b8d2 | 6773 | |
22f10457 | 6774 | out: |
aaf7d73e | 6775 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
6776 | return ret; |
6777 | } | |
0c9ec4be DW |
6778 | |
6779 | /* Iterate all the free extents in the group. */ | |
6780 | int | |
6781 | ext4_mballoc_query_range( | |
6782 | struct super_block *sb, | |
6783 | ext4_group_t group, | |
6784 | ext4_grpblk_t start, | |
6785 | ext4_grpblk_t end, | |
6786 | ext4_mballoc_query_range_fn formatter, | |
6787 | void *priv) | |
6788 | { | |
6789 | void *bitmap; | |
6790 | ext4_grpblk_t next; | |
6791 | struct ext4_buddy e4b; | |
6792 | int error; | |
6793 | ||
6794 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
6795 | if (error) | |
6796 | return error; | |
6797 | bitmap = e4b.bd_bitmap; | |
6798 | ||
6799 | ext4_lock_group(sb, group); | |
6800 | ||
6801 | start = (e4b.bd_info->bb_first_free > start) ? | |
6802 | e4b.bd_info->bb_first_free : start; | |
6803 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
6804 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
6805 | ||
6806 | while (start <= end) { | |
6807 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
6808 | if (start > end) | |
6809 | break; | |
6810 | next = mb_find_next_bit(bitmap, end + 1, start); | |
6811 | ||
6812 | ext4_unlock_group(sb, group); | |
6813 | error = formatter(sb, group, start, next - start, priv); | |
6814 | if (error) | |
6815 | goto out_unload; | |
6816 | ext4_lock_group(sb, group); | |
6817 | ||
6818 | start = next + 1; | |
6819 | } | |
6820 | ||
6821 | ext4_unlock_group(sb, group); | |
6822 | out_unload: | |
6823 | ext4_mb_unload_buddy(&e4b); | |
6824 | ||
6825 | return error; | |
6826 | } |