]>
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 | * | |
143 | * 2) Average fragment size rb tree (sbi->s_mb_avg_fragment_size_root) | |
144 | * | |
145 | * Locking: sbi->s_mb_rb_lock (rwlock) | |
146 | * | |
147 | * This is a red black tree consisting of group infos and the tree is sorted | |
148 | * by average fragment sizes (which is calculated as ext4_group_info->bb_free | |
149 | * / ext4_group_info->bb_fragments). | |
150 | * | |
151 | * When "mb_optimize_scan" mount option is set, mballoc consults the above data | |
152 | * structures to decide the order in which groups are to be traversed for | |
153 | * fulfilling an allocation request. | |
154 | * | |
155 | * At CR = 0, we look for groups which have the largest_free_order >= the order | |
156 | * of the request. We directly look at the largest free order list in the data | |
157 | * structure (1) above where largest_free_order = order of the request. If that | |
158 | * list is empty, we look at remaining list in the increasing order of | |
159 | * largest_free_order. This allows us to perform CR = 0 lookup in O(1) time. | |
160 | * | |
161 | * At CR = 1, we only consider groups where average fragment size > request | |
162 | * size. So, we lookup a group which has average fragment size just above or | |
163 | * equal to request size using our rb tree (data structure 2) in O(log N) time. | |
164 | * | |
165 | * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in | |
166 | * linear order which requires O(N) search time for each CR 0 and CR 1 phase. | |
167 | * | |
d7a1fee1 | 168 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 169 | * |
b713a5ec TT |
170 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
171 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
172 | * /sys/fs/ext4/<partition>/mb_order2_req | |
196e402a | 173 | * /sys/fs/ext4/<partition>/mb_linear_limit |
c9de560d | 174 | * |
b713a5ec | 175 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
176 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
177 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 178 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 179 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
180 | * stripe size. This should result in better allocation on RAID setups. If |
181 | * not, we search in the specific group using bitmap for best extents. The | |
182 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 183 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 184 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
185 | * best extent in the found extents. Searching for the blocks starts with |
186 | * the group specified as the goal value in allocation context via | |
187 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 188 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
189 | * checked. |
190 | * | |
196e402a HS |
191 | * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not |
192 | * get traversed linearly. That may result in subsequent allocations being not | |
193 | * close to each other. And so, the underlying device may get filled up in a | |
194 | * non-linear fashion. While that may not matter on non-rotational devices, for | |
195 | * rotational devices that may result in higher seek times. "mb_linear_limit" | |
196 | * tells mballoc how many groups mballoc should search linearly before | |
197 | * performing consulting above data structures for more efficient lookups. For | |
198 | * non rotational devices, this value defaults to 0 and for rotational devices | |
199 | * this is set to MB_DEFAULT_LINEAR_LIMIT. | |
200 | * | |
c9de560d AT |
201 | * Both the prealloc space are getting populated as above. So for the first |
202 | * request we will hit the buddy cache which will result in this prealloc | |
203 | * space getting filled. The prealloc space is then later used for the | |
204 | * subsequent request. | |
205 | */ | |
206 | ||
207 | /* | |
208 | * mballoc operates on the following data: | |
209 | * - on-disk bitmap | |
210 | * - in-core buddy (actually includes buddy and bitmap) | |
211 | * - preallocation descriptors (PAs) | |
212 | * | |
213 | * there are two types of preallocations: | |
214 | * - inode | |
215 | * assiged to specific inode and can be used for this inode only. | |
216 | * it describes part of inode's space preallocated to specific | |
217 | * physical blocks. any block from that preallocated can be used | |
218 | * independent. the descriptor just tracks number of blocks left | |
219 | * unused. so, before taking some block from descriptor, one must | |
220 | * make sure corresponded logical block isn't allocated yet. this | |
221 | * also means that freeing any block within descriptor's range | |
222 | * must discard all preallocated blocks. | |
223 | * - locality group | |
224 | * assigned to specific locality group which does not translate to | |
225 | * permanent set of inodes: inode can join and leave group. space | |
226 | * from this type of preallocation can be used for any inode. thus | |
227 | * it's consumed from the beginning to the end. | |
228 | * | |
229 | * relation between them can be expressed as: | |
230 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
231 | * | |
232 | * this mean blocks mballoc considers used are: | |
233 | * - allocated blocks (persistent) | |
234 | * - preallocated blocks (non-persistent) | |
235 | * | |
236 | * consistency in mballoc world means that at any time a block is either | |
237 | * free or used in ALL structures. notice: "any time" should not be read | |
238 | * literally -- time is discrete and delimited by locks. | |
239 | * | |
240 | * to keep it simple, we don't use block numbers, instead we count number of | |
241 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
242 | * | |
243 | * all operations can be expressed as: | |
244 | * - init buddy: buddy = on-disk + PAs | |
245 | * - new PA: buddy += N; PA = N | |
246 | * - use inode PA: on-disk += N; PA -= N | |
247 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
248 | * - use locality group PA on-disk += N; PA -= N | |
249 | * - discard locality group PA buddy -= PA; PA = 0 | |
250 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
251 | * is used in real operation because we can't know actual used | |
252 | * bits from PA, only from on-disk bitmap | |
253 | * | |
254 | * if we follow this strict logic, then all operations above should be atomic. | |
255 | * given some of them can block, we'd have to use something like semaphores | |
256 | * killing performance on high-end SMP hardware. let's try to relax it using | |
257 | * the following knowledge: | |
258 | * 1) if buddy is referenced, it's already initialized | |
259 | * 2) while block is used in buddy and the buddy is referenced, | |
260 | * nobody can re-allocate that block | |
261 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
262 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
263 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
264 | * block | |
265 | * | |
266 | * so, now we're building a concurrency table: | |
267 | * - init buddy vs. | |
268 | * - new PA | |
269 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
270 | * until PA is linked to allocation group to avoid concurrent buddy init | |
271 | * - use inode PA | |
272 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
273 | * given (3) we care that PA-=N operation doesn't interfere with init | |
274 | * - discard inode PA | |
275 | * the simplest way would be to have buddy initialized by the discard | |
276 | * - use locality group PA | |
277 | * again PA-=N must be serialized with init | |
278 | * - discard locality group PA | |
279 | * the simplest way would be to have buddy initialized by the discard | |
280 | * - new PA vs. | |
281 | * - use inode PA | |
282 | * i_data_sem serializes them | |
283 | * - discard inode PA | |
284 | * discard process must wait until PA isn't used by another process | |
285 | * - use locality group PA | |
286 | * some mutex should serialize them | |
287 | * - discard locality group PA | |
288 | * discard process must wait until PA isn't used by another process | |
289 | * - use inode PA | |
290 | * - use inode PA | |
291 | * i_data_sem or another mutex should serializes them | |
292 | * - discard inode PA | |
293 | * discard process must wait until PA isn't used by another process | |
294 | * - use locality group PA | |
295 | * nothing wrong here -- they're different PAs covering different blocks | |
296 | * - discard locality group PA | |
297 | * discard process must wait until PA isn't used by another process | |
298 | * | |
299 | * now we're ready to make few consequences: | |
300 | * - PA is referenced and while it is no discard is possible | |
301 | * - PA is referenced until block isn't marked in on-disk bitmap | |
302 | * - PA changes only after on-disk bitmap | |
303 | * - discard must not compete with init. either init is done before | |
304 | * any discard or they're serialized somehow | |
305 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
306 | * | |
307 | * a special case when we've used PA to emptiness. no need to modify buddy | |
308 | * in this case, but we should care about concurrent init | |
309 | * | |
310 | */ | |
311 | ||
312 | /* | |
313 | * Logic in few words: | |
314 | * | |
315 | * - allocation: | |
316 | * load group | |
317 | * find blocks | |
318 | * mark bits in on-disk bitmap | |
319 | * release group | |
320 | * | |
321 | * - use preallocation: | |
322 | * find proper PA (per-inode or group) | |
323 | * load group | |
324 | * mark bits in on-disk bitmap | |
325 | * release group | |
326 | * release PA | |
327 | * | |
328 | * - free: | |
329 | * load group | |
330 | * mark bits in on-disk bitmap | |
331 | * release group | |
332 | * | |
333 | * - discard preallocations in group: | |
334 | * mark PAs deleted | |
335 | * move them onto local list | |
336 | * load on-disk bitmap | |
337 | * load group | |
338 | * remove PA from object (inode or locality group) | |
339 | * mark free blocks in-core | |
340 | * | |
341 | * - discard inode's preallocations: | |
342 | */ | |
343 | ||
344 | /* | |
345 | * Locking rules | |
346 | * | |
347 | * Locks: | |
348 | * - bitlock on a group (group) | |
349 | * - object (inode/locality) (object) | |
350 | * - per-pa lock (pa) | |
196e402a HS |
351 | * - cr0 lists lock (cr0) |
352 | * - cr1 tree lock (cr1) | |
c9de560d AT |
353 | * |
354 | * Paths: | |
355 | * - new pa | |
356 | * object | |
357 | * group | |
358 | * | |
359 | * - find and use pa: | |
360 | * pa | |
361 | * | |
362 | * - release consumed pa: | |
363 | * pa | |
364 | * group | |
365 | * object | |
366 | * | |
367 | * - generate in-core bitmap: | |
368 | * group | |
369 | * pa | |
370 | * | |
371 | * - discard all for given object (inode, locality group): | |
372 | * object | |
373 | * pa | |
374 | * group | |
375 | * | |
376 | * - discard all for given group: | |
377 | * group | |
378 | * pa | |
379 | * group | |
380 | * object | |
381 | * | |
196e402a HS |
382 | * - allocation path (ext4_mb_regular_allocator) |
383 | * group | |
384 | * cr0/cr1 | |
c9de560d | 385 | */ |
c3a326a6 AK |
386 | static struct kmem_cache *ext4_pspace_cachep; |
387 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 388 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
389 | |
390 | /* We create slab caches for groupinfo data structures based on the | |
391 | * superblock block size. There will be one per mounted filesystem for | |
392 | * each unique s_blocksize_bits */ | |
2892c15d | 393 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
394 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
395 | ||
d6006186 | 396 | static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
2892c15d ES |
397 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", |
398 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
399 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
400 | }; | |
401 | ||
c3a326a6 AK |
402 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
403 | ext4_group_t group); | |
7a2fcbf7 AK |
404 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
405 | ext4_group_t group); | |
53f86b17 | 406 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); |
c3a326a6 | 407 | |
196e402a HS |
408 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
409 | ext4_group_t group, int cr); | |
410 | ||
55cdd0af WJ |
411 | static int ext4_try_to_trim_range(struct super_block *sb, |
412 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
413 | ext4_grpblk_t max, ext4_grpblk_t minblocks); | |
414 | ||
07b5b8e1 RH |
415 | /* |
416 | * The algorithm using this percpu seq counter goes below: | |
417 | * 1. We sample the percpu discard_pa_seq counter before trying for block | |
418 | * allocation in ext4_mb_new_blocks(). | |
419 | * 2. We increment this percpu discard_pa_seq counter when we either allocate | |
420 | * or free these blocks i.e. while marking those blocks as used/free in | |
421 | * mb_mark_used()/mb_free_blocks(). | |
422 | * 3. We also increment this percpu seq counter when we successfully identify | |
423 | * that the bb_prealloc_list is not empty and hence proceed for discarding | |
424 | * of those PAs inside ext4_mb_discard_group_preallocations(). | |
425 | * | |
426 | * Now to make sure that the regular fast path of block allocation is not | |
427 | * affected, as a small optimization we only sample the percpu seq counter | |
428 | * on that cpu. Only when the block allocation fails and when freed blocks | |
429 | * found were 0, that is when we sample percpu seq counter for all cpus using | |
430 | * below function ext4_get_discard_pa_seq_sum(). This happens after making | |
431 | * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. | |
432 | */ | |
433 | static DEFINE_PER_CPU(u64, discard_pa_seq); | |
434 | static inline u64 ext4_get_discard_pa_seq_sum(void) | |
435 | { | |
436 | int __cpu; | |
437 | u64 __seq = 0; | |
438 | ||
439 | for_each_possible_cpu(__cpu) | |
440 | __seq += per_cpu(discard_pa_seq, __cpu); | |
441 | return __seq; | |
442 | } | |
443 | ||
ffad0a44 AK |
444 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
445 | { | |
c9de560d | 446 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
447 | *bit += ((unsigned long) addr & 7UL) << 3; |
448 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 449 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
450 | *bit += ((unsigned long) addr & 3UL) << 3; |
451 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
452 | #else |
453 | #error "how many bits you are?!" | |
454 | #endif | |
ffad0a44 AK |
455 | return addr; |
456 | } | |
c9de560d AT |
457 | |
458 | static inline int mb_test_bit(int bit, void *addr) | |
459 | { | |
460 | /* | |
461 | * ext4_test_bit on architecture like powerpc | |
462 | * needs unsigned long aligned address | |
463 | */ | |
ffad0a44 | 464 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
465 | return ext4_test_bit(bit, addr); |
466 | } | |
467 | ||
468 | static inline void mb_set_bit(int bit, void *addr) | |
469 | { | |
ffad0a44 | 470 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
471 | ext4_set_bit(bit, addr); |
472 | } | |
473 | ||
c9de560d AT |
474 | static inline void mb_clear_bit(int bit, void *addr) |
475 | { | |
ffad0a44 | 476 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
477 | ext4_clear_bit(bit, addr); |
478 | } | |
479 | ||
eabe0444 AS |
480 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
481 | { | |
482 | addr = mb_correct_addr_and_bit(&bit, addr); | |
483 | return ext4_test_and_clear_bit(bit, addr); | |
484 | } | |
485 | ||
ffad0a44 AK |
486 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
487 | { | |
e7dfb246 | 488 | int fix = 0, ret, tmpmax; |
ffad0a44 | 489 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 490 | tmpmax = max + fix; |
ffad0a44 AK |
491 | start += fix; |
492 | ||
e7dfb246 AK |
493 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
494 | if (ret > max) | |
495 | return max; | |
496 | return ret; | |
ffad0a44 AK |
497 | } |
498 | ||
499 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
500 | { | |
e7dfb246 | 501 | int fix = 0, ret, tmpmax; |
ffad0a44 | 502 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 503 | tmpmax = max + fix; |
ffad0a44 AK |
504 | start += fix; |
505 | ||
e7dfb246 AK |
506 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
507 | if (ret > max) | |
508 | return max; | |
509 | return ret; | |
ffad0a44 AK |
510 | } |
511 | ||
c9de560d AT |
512 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
513 | { | |
514 | char *bb; | |
515 | ||
c5e8f3f3 | 516 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
517 | BUG_ON(max == NULL); |
518 | ||
519 | if (order > e4b->bd_blkbits + 1) { | |
520 | *max = 0; | |
521 | return NULL; | |
522 | } | |
523 | ||
524 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
525 | if (order == 0) { |
526 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 527 | return e4b->bd_bitmap; |
84b775a3 | 528 | } |
c9de560d | 529 | |
c5e8f3f3 | 530 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
531 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
532 | ||
533 | return bb; | |
534 | } | |
535 | ||
536 | #ifdef DOUBLE_CHECK | |
537 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
538 | int first, int count) | |
539 | { | |
540 | int i; | |
541 | struct super_block *sb = e4b->bd_sb; | |
542 | ||
543 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
544 | return; | |
bc8e6740 | 545 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
546 | for (i = 0; i < count; i++) { |
547 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
548 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
549 | |
550 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 551 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 552 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
553 | inode ? inode->i_ino : 0, |
554 | blocknr, | |
555 | "freeing block already freed " | |
556 | "(bit %u)", | |
557 | first + i); | |
736dedbb WS |
558 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
559 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
560 | } |
561 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
562 | } | |
563 | } | |
564 | ||
565 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
566 | { | |
567 | int i; | |
568 | ||
569 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
570 | return; | |
bc8e6740 | 571 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
572 | for (i = 0; i < count; i++) { |
573 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
574 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
575 | } | |
576 | } | |
577 | ||
578 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
579 | { | |
eb2b8ebb RH |
580 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
581 | return; | |
c9de560d AT |
582 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { |
583 | unsigned char *b1, *b2; | |
584 | int i; | |
585 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
586 | b2 = (unsigned char *) bitmap; | |
587 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
588 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
589 | ext4_msg(e4b->bd_sb, KERN_ERR, |
590 | "corruption in group %u " | |
591 | "at byte %u(%u): %x in copy != %x " | |
592 | "on disk/prealloc", | |
593 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
594 | BUG(); |
595 | } | |
596 | } | |
597 | } | |
598 | } | |
599 | ||
a3450215 RH |
600 | static void mb_group_bb_bitmap_alloc(struct super_block *sb, |
601 | struct ext4_group_info *grp, ext4_group_t group) | |
602 | { | |
603 | struct buffer_head *bh; | |
604 | ||
605 | grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); | |
eb2b8ebb RH |
606 | if (!grp->bb_bitmap) |
607 | return; | |
a3450215 RH |
608 | |
609 | bh = ext4_read_block_bitmap(sb, group); | |
eb2b8ebb RH |
610 | if (IS_ERR_OR_NULL(bh)) { |
611 | kfree(grp->bb_bitmap); | |
612 | grp->bb_bitmap = NULL; | |
613 | return; | |
614 | } | |
a3450215 RH |
615 | |
616 | memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); | |
617 | put_bh(bh); | |
618 | } | |
619 | ||
620 | static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
621 | { | |
622 | kfree(grp->bb_bitmap); | |
623 | } | |
624 | ||
c9de560d AT |
625 | #else |
626 | static inline void mb_free_blocks_double(struct inode *inode, | |
627 | struct ext4_buddy *e4b, int first, int count) | |
628 | { | |
629 | return; | |
630 | } | |
631 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
632 | int first, int count) | |
633 | { | |
634 | return; | |
635 | } | |
636 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
637 | { | |
638 | return; | |
639 | } | |
a3450215 RH |
640 | |
641 | static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, | |
642 | struct ext4_group_info *grp, ext4_group_t group) | |
643 | { | |
644 | return; | |
645 | } | |
646 | ||
647 | static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
648 | { | |
649 | return; | |
650 | } | |
c9de560d AT |
651 | #endif |
652 | ||
653 | #ifdef AGGRESSIVE_CHECK | |
654 | ||
655 | #define MB_CHECK_ASSERT(assert) \ | |
656 | do { \ | |
657 | if (!(assert)) { \ | |
658 | printk(KERN_EMERG \ | |
659 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
660 | function, file, line, # assert); \ | |
661 | BUG(); \ | |
662 | } \ | |
663 | } while (0) | |
664 | ||
665 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
666 | const char *function, int line) | |
667 | { | |
668 | struct super_block *sb = e4b->bd_sb; | |
669 | int order = e4b->bd_blkbits + 1; | |
670 | int max; | |
671 | int max2; | |
672 | int i; | |
673 | int j; | |
674 | int k; | |
675 | int count; | |
676 | struct ext4_group_info *grp; | |
677 | int fragments = 0; | |
678 | int fstart; | |
679 | struct list_head *cur; | |
680 | void *buddy; | |
681 | void *buddy2; | |
682 | ||
addd752c CX |
683 | if (e4b->bd_info->bb_check_counter++ % 10) |
684 | return 0; | |
c9de560d AT |
685 | |
686 | while (order > 1) { | |
687 | buddy = mb_find_buddy(e4b, order, &max); | |
688 | MB_CHECK_ASSERT(buddy); | |
689 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
690 | MB_CHECK_ASSERT(buddy2); | |
691 | MB_CHECK_ASSERT(buddy != buddy2); | |
692 | MB_CHECK_ASSERT(max * 2 == max2); | |
693 | ||
694 | count = 0; | |
695 | for (i = 0; i < max; i++) { | |
696 | ||
697 | if (mb_test_bit(i, buddy)) { | |
af2b3275 | 698 | /* only single bit in buddy2 may be 0 */ |
c9de560d AT |
699 | if (!mb_test_bit(i << 1, buddy2)) { |
700 | MB_CHECK_ASSERT( | |
701 | mb_test_bit((i<<1)+1, buddy2)); | |
c9de560d AT |
702 | } |
703 | continue; | |
704 | } | |
705 | ||
0a10da73 | 706 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
707 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
708 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
709 | ||
710 | for (j = 0; j < (1 << order); j++) { | |
711 | k = (i * (1 << order)) + j; | |
712 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 713 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
714 | } |
715 | count++; | |
716 | } | |
717 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
718 | order--; | |
719 | } | |
720 | ||
721 | fstart = -1; | |
722 | buddy = mb_find_buddy(e4b, 0, &max); | |
723 | for (i = 0; i < max; i++) { | |
724 | if (!mb_test_bit(i, buddy)) { | |
725 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
726 | if (fstart == -1) { | |
727 | fragments++; | |
728 | fstart = i; | |
729 | } | |
730 | continue; | |
731 | } | |
732 | fstart = -1; | |
733 | /* check used bits only */ | |
734 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
735 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
736 | k = i >> j; | |
737 | MB_CHECK_ASSERT(k < max2); | |
738 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
739 | } | |
740 | } | |
741 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
742 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
743 | ||
744 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
745 | list_for_each(cur, &grp->bb_prealloc_list) { |
746 | ext4_group_t groupnr; | |
747 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
748 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
749 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 750 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 751 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
752 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
753 | } | |
754 | return 0; | |
755 | } | |
756 | #undef MB_CHECK_ASSERT | |
757 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 758 | __FILE__, __func__, __LINE__) |
c9de560d AT |
759 | #else |
760 | #define mb_check_buddy(e4b) | |
761 | #endif | |
762 | ||
7c786059 CL |
763 | /* |
764 | * Divide blocks started from @first with length @len into | |
765 | * smaller chunks with power of 2 blocks. | |
766 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
767 | * then increase bb_counters[] for corresponded chunk size. | |
768 | */ | |
c9de560d | 769 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 770 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
771 | struct ext4_group_info *grp) |
772 | { | |
773 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
774 | ext4_grpblk_t min; |
775 | ext4_grpblk_t max; | |
776 | ext4_grpblk_t chunk; | |
69e43e8c | 777 | unsigned int border; |
c9de560d | 778 | |
7137d7a4 | 779 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
780 | |
781 | border = 2 << sb->s_blocksize_bits; | |
782 | ||
783 | while (len > 0) { | |
784 | /* find how many blocks can be covered since this position */ | |
785 | max = ffs(first | border) - 1; | |
786 | ||
787 | /* find how many blocks of power 2 we need to mark */ | |
788 | min = fls(len) - 1; | |
789 | ||
790 | if (max < min) | |
791 | min = max; | |
792 | chunk = 1 << min; | |
793 | ||
794 | /* mark multiblock chunks only */ | |
795 | grp->bb_counters[min]++; | |
796 | if (min > 0) | |
797 | mb_clear_bit(first >> min, | |
798 | buddy + sbi->s_mb_offsets[min]); | |
799 | ||
800 | len -= chunk; | |
801 | first += chunk; | |
802 | } | |
803 | } | |
804 | ||
196e402a HS |
805 | static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *new, |
806 | int (*cmp)(struct rb_node *, struct rb_node *)) | |
807 | { | |
808 | struct rb_node **iter = &root->rb_node, *parent = NULL; | |
809 | ||
810 | while (*iter) { | |
811 | parent = *iter; | |
812 | if (cmp(new, *iter) > 0) | |
813 | iter = &((*iter)->rb_left); | |
814 | else | |
815 | iter = &((*iter)->rb_right); | |
816 | } | |
817 | ||
818 | rb_link_node(new, parent, iter); | |
819 | rb_insert_color(new, root); | |
820 | } | |
821 | ||
822 | static int | |
823 | ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2) | |
824 | { | |
825 | struct ext4_group_info *grp1 = rb_entry(rb1, | |
826 | struct ext4_group_info, | |
827 | bb_avg_fragment_size_rb); | |
828 | struct ext4_group_info *grp2 = rb_entry(rb2, | |
829 | struct ext4_group_info, | |
830 | bb_avg_fragment_size_rb); | |
831 | int num_frags_1, num_frags_2; | |
832 | ||
833 | num_frags_1 = grp1->bb_fragments ? | |
834 | grp1->bb_free / grp1->bb_fragments : 0; | |
835 | num_frags_2 = grp2->bb_fragments ? | |
836 | grp2->bb_free / grp2->bb_fragments : 0; | |
837 | ||
838 | return (num_frags_2 - num_frags_1); | |
839 | } | |
840 | ||
841 | /* | |
842 | * Reinsert grpinfo into the avg_fragment_size tree with new average | |
843 | * fragment size. | |
844 | */ | |
845 | static void | |
846 | mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) | |
847 | { | |
848 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
849 | ||
850 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) | |
851 | return; | |
852 | ||
853 | write_lock(&sbi->s_mb_rb_lock); | |
854 | if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) { | |
855 | rb_erase(&grp->bb_avg_fragment_size_rb, | |
856 | &sbi->s_mb_avg_fragment_size_root); | |
857 | RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb); | |
858 | } | |
859 | ||
860 | ext4_mb_rb_insert(&sbi->s_mb_avg_fragment_size_root, | |
861 | &grp->bb_avg_fragment_size_rb, | |
862 | ext4_mb_avg_fragment_size_cmp); | |
863 | write_unlock(&sbi->s_mb_rb_lock); | |
864 | } | |
865 | ||
866 | /* | |
867 | * Choose next group by traversing largest_free_order lists. Updates *new_cr if | |
868 | * cr level needs an update. | |
869 | */ | |
870 | static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, | |
871 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
872 | { | |
873 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
874 | struct ext4_group_info *iter, *grp; | |
875 | int i; | |
876 | ||
877 | if (ac->ac_status == AC_STATUS_FOUND) | |
878 | return; | |
879 | ||
880 | if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED)) | |
881 | atomic_inc(&sbi->s_bal_cr0_bad_suggestions); | |
882 | ||
883 | grp = NULL; | |
884 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
885 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) | |
886 | continue; | |
887 | read_lock(&sbi->s_mb_largest_free_orders_locks[i]); | |
888 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) { | |
889 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
890 | continue; | |
891 | } | |
892 | grp = NULL; | |
893 | list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], | |
894 | bb_largest_free_order_node) { | |
895 | if (sbi->s_mb_stats) | |
896 | atomic64_inc(&sbi->s_bal_cX_groups_considered[0]); | |
897 | if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) { | |
898 | grp = iter; | |
899 | break; | |
900 | } | |
901 | } | |
902 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
903 | if (grp) | |
904 | break; | |
905 | } | |
906 | ||
907 | if (!grp) { | |
908 | /* Increment cr and search again */ | |
909 | *new_cr = 1; | |
910 | } else { | |
911 | *group = grp->bb_group; | |
912 | ac->ac_last_optimal_group = *group; | |
913 | ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED; | |
914 | } | |
915 | } | |
916 | ||
917 | /* | |
918 | * Choose next group by traversing average fragment size tree. Updates *new_cr | |
919 | * if cr lvel needs an update. Sets EXT4_MB_SEARCH_NEXT_LINEAR to indicate that | |
920 | * the linear search should continue for one iteration since there's lock | |
921 | * contention on the rb tree lock. | |
922 | */ | |
923 | static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, | |
924 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
925 | { | |
926 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
927 | int avg_fragment_size, best_so_far; | |
928 | struct rb_node *node, *found; | |
929 | struct ext4_group_info *grp; | |
930 | ||
931 | /* | |
932 | * If there is contention on the lock, instead of waiting for the lock | |
933 | * to become available, just continue searching lineraly. We'll resume | |
934 | * our rb tree search later starting at ac->ac_last_optimal_group. | |
935 | */ | |
936 | if (!read_trylock(&sbi->s_mb_rb_lock)) { | |
937 | ac->ac_flags |= EXT4_MB_SEARCH_NEXT_LINEAR; | |
938 | return; | |
939 | } | |
940 | ||
941 | if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) { | |
942 | if (sbi->s_mb_stats) | |
943 | atomic_inc(&sbi->s_bal_cr1_bad_suggestions); | |
944 | /* We have found something at CR 1 in the past */ | |
945 | grp = ext4_get_group_info(ac->ac_sb, ac->ac_last_optimal_group); | |
946 | for (found = rb_next(&grp->bb_avg_fragment_size_rb); found != NULL; | |
947 | found = rb_next(found)) { | |
948 | grp = rb_entry(found, struct ext4_group_info, | |
949 | bb_avg_fragment_size_rb); | |
950 | if (sbi->s_mb_stats) | |
951 | atomic64_inc(&sbi->s_bal_cX_groups_considered[1]); | |
952 | if (likely(ext4_mb_good_group(ac, grp->bb_group, 1))) | |
953 | break; | |
954 | } | |
955 | goto done; | |
956 | } | |
957 | ||
958 | node = sbi->s_mb_avg_fragment_size_root.rb_node; | |
959 | best_so_far = 0; | |
960 | found = NULL; | |
961 | ||
962 | while (node) { | |
963 | grp = rb_entry(node, struct ext4_group_info, | |
964 | bb_avg_fragment_size_rb); | |
965 | avg_fragment_size = 0; | |
966 | if (ext4_mb_good_group(ac, grp->bb_group, 1)) { | |
967 | avg_fragment_size = grp->bb_fragments ? | |
968 | grp->bb_free / grp->bb_fragments : 0; | |
969 | if (!best_so_far || avg_fragment_size < best_so_far) { | |
970 | best_so_far = avg_fragment_size; | |
971 | found = node; | |
972 | } | |
973 | } | |
974 | if (avg_fragment_size > ac->ac_g_ex.fe_len) | |
975 | node = node->rb_right; | |
976 | else | |
977 | node = node->rb_left; | |
978 | } | |
979 | ||
980 | done: | |
981 | if (found) { | |
982 | grp = rb_entry(found, struct ext4_group_info, | |
983 | bb_avg_fragment_size_rb); | |
984 | *group = grp->bb_group; | |
985 | ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED; | |
986 | } else { | |
987 | *new_cr = 2; | |
988 | } | |
989 | ||
990 | read_unlock(&sbi->s_mb_rb_lock); | |
991 | ac->ac_last_optimal_group = *group; | |
992 | } | |
993 | ||
994 | static inline int should_optimize_scan(struct ext4_allocation_context *ac) | |
995 | { | |
996 | if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) | |
997 | return 0; | |
998 | if (ac->ac_criteria >= 2) | |
999 | return 0; | |
077d0c2c | 1000 | if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) |
196e402a HS |
1001 | return 0; |
1002 | return 1; | |
1003 | } | |
1004 | ||
1005 | /* | |
1006 | * Return next linear group for allocation. If linear traversal should not be | |
1007 | * performed, this function just returns the same group | |
1008 | */ | |
1009 | static int | |
1010 | next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups) | |
1011 | { | |
1012 | if (!should_optimize_scan(ac)) | |
1013 | goto inc_and_return; | |
1014 | ||
1015 | if (ac->ac_groups_linear_remaining) { | |
1016 | ac->ac_groups_linear_remaining--; | |
1017 | goto inc_and_return; | |
1018 | } | |
1019 | ||
1020 | if (ac->ac_flags & EXT4_MB_SEARCH_NEXT_LINEAR) { | |
1021 | ac->ac_flags &= ~EXT4_MB_SEARCH_NEXT_LINEAR; | |
1022 | goto inc_and_return; | |
1023 | } | |
1024 | ||
1025 | return group; | |
1026 | inc_and_return: | |
1027 | /* | |
1028 | * Artificially restricted ngroups for non-extent | |
1029 | * files makes group > ngroups possible on first loop. | |
1030 | */ | |
1031 | return group + 1 >= ngroups ? 0 : group + 1; | |
1032 | } | |
1033 | ||
1034 | /* | |
1035 | * ext4_mb_choose_next_group: choose next group for allocation. | |
1036 | * | |
1037 | * @ac Allocation Context | |
1038 | * @new_cr This is an output parameter. If the there is no good group | |
1039 | * available at current CR level, this field is updated to indicate | |
1040 | * the new cr level that should be used. | |
1041 | * @group This is an input / output parameter. As an input it indicates the | |
1042 | * next group that the allocator intends to use for allocation. As | |
1043 | * output, this field indicates the next group that should be used as | |
1044 | * determined by the optimization functions. | |
1045 | * @ngroups Total number of groups | |
1046 | */ | |
1047 | static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, | |
1048 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
1049 | { | |
1050 | *new_cr = ac->ac_criteria; | |
1051 | ||
4fca50d4 JK |
1052 | if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) { |
1053 | *group = next_linear_group(ac, *group, ngroups); | |
196e402a | 1054 | return; |
4fca50d4 | 1055 | } |
196e402a HS |
1056 | |
1057 | if (*new_cr == 0) { | |
1058 | ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups); | |
1059 | } else if (*new_cr == 1) { | |
1060 | ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups); | |
1061 | } else { | |
1062 | /* | |
1063 | * TODO: For CR=2, we can arrange groups in an rb tree sorted by | |
1064 | * bb_free. But until that happens, we should never come here. | |
1065 | */ | |
1066 | WARN_ON(1); | |
1067 | } | |
1068 | } | |
1069 | ||
8a57d9d6 CW |
1070 | /* |
1071 | * Cache the order of the largest free extent we have available in this block | |
1072 | * group. | |
1073 | */ | |
1074 | static void | |
1075 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
1076 | { | |
196e402a | 1077 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
8a57d9d6 | 1078 | int i; |
8a57d9d6 | 1079 | |
1940265e JK |
1080 | for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) |
1081 | if (grp->bb_counters[i] > 0) | |
1082 | break; | |
1083 | /* No need to move between order lists? */ | |
1084 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || | |
1085 | i == grp->bb_largest_free_order) { | |
1086 | grp->bb_largest_free_order = i; | |
1087 | return; | |
1088 | } | |
1089 | ||
1090 | if (grp->bb_largest_free_order >= 0) { | |
196e402a HS |
1091 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1092 | grp->bb_largest_free_order]); | |
1093 | list_del_init(&grp->bb_largest_free_order_node); | |
1094 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1095 | grp->bb_largest_free_order]); | |
1096 | } | |
1940265e JK |
1097 | grp->bb_largest_free_order = i; |
1098 | if (grp->bb_largest_free_order >= 0 && grp->bb_free) { | |
196e402a HS |
1099 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1100 | grp->bb_largest_free_order]); | |
1101 | list_add_tail(&grp->bb_largest_free_order_node, | |
1102 | &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); | |
1103 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1104 | grp->bb_largest_free_order]); | |
1105 | } | |
8a57d9d6 CW |
1106 | } |
1107 | ||
089ceecc ES |
1108 | static noinline_for_stack |
1109 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
1110 | void *buddy, void *bitmap, ext4_group_t group) |
1111 | { | |
1112 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 1113 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 1114 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
1115 | ext4_grpblk_t i = 0; |
1116 | ext4_grpblk_t first; | |
1117 | ext4_grpblk_t len; | |
c9de560d AT |
1118 | unsigned free = 0; |
1119 | unsigned fragments = 0; | |
1120 | unsigned long long period = get_cycles(); | |
1121 | ||
1122 | /* initialize buddy from bitmap which is aggregation | |
1123 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 1124 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
1125 | grp->bb_first_free = i; |
1126 | while (i < max) { | |
1127 | fragments++; | |
1128 | first = i; | |
ffad0a44 | 1129 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
1130 | len = i - first; |
1131 | free += len; | |
1132 | if (len > 1) | |
1133 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
1134 | else | |
1135 | grp->bb_counters[0]++; | |
1136 | if (i < max) | |
ffad0a44 | 1137 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
1138 | } |
1139 | grp->bb_fragments = fragments; | |
1140 | ||
1141 | if (free != grp->bb_free) { | |
e29136f8 | 1142 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
1143 | "block bitmap and bg descriptor " |
1144 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 1145 | free, grp->bb_free); |
e56eb659 | 1146 | /* |
163a203d | 1147 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
1148 | * corrupt and update bb_free using bitmap value |
1149 | */ | |
c9de560d | 1150 | grp->bb_free = free; |
db79e6d1 WS |
1151 | ext4_mark_group_bitmap_corrupted(sb, group, |
1152 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 1153 | } |
8a57d9d6 | 1154 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
1155 | |
1156 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
1157 | ||
1158 | period = get_cycles() - period; | |
67d25186 HS |
1159 | atomic_inc(&sbi->s_mb_buddies_generated); |
1160 | atomic64_add(period, &sbi->s_mb_generation_time); | |
196e402a | 1161 | mb_update_avg_fragment_size(sb, grp); |
c9de560d AT |
1162 | } |
1163 | ||
1164 | /* The buddy information is attached the buddy cache inode | |
1165 | * for convenience. The information regarding each group | |
1166 | * is loaded via ext4_mb_load_buddy. The information involve | |
1167 | * block bitmap and buddy information. The information are | |
1168 | * stored in the inode as | |
1169 | * | |
1170 | * { page } | |
c3a326a6 | 1171 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
1172 | * |
1173 | * | |
1174 | * one block each for bitmap and buddy information. | |
1175 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 1176 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
1177 | * So it can have information regarding groups_per_page which |
1178 | * is blocks_per_page/2 | |
8a57d9d6 CW |
1179 | * |
1180 | * Locking note: This routine takes the block group lock of all groups | |
1181 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
1182 | */ |
1183 | ||
adb7ef60 | 1184 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 1185 | { |
8df9675f | 1186 | ext4_group_t ngroups; |
c9de560d AT |
1187 | int blocksize; |
1188 | int blocks_per_page; | |
1189 | int groups_per_page; | |
1190 | int err = 0; | |
1191 | int i; | |
813e5727 | 1192 | ext4_group_t first_group, group; |
c9de560d AT |
1193 | int first_block; |
1194 | struct super_block *sb; | |
1195 | struct buffer_head *bhs; | |
fa77dcfa | 1196 | struct buffer_head **bh = NULL; |
c9de560d AT |
1197 | struct inode *inode; |
1198 | char *data; | |
1199 | char *bitmap; | |
9b8b7d35 | 1200 | struct ext4_group_info *grinfo; |
c9de560d | 1201 | |
c9de560d AT |
1202 | inode = page->mapping->host; |
1203 | sb = inode->i_sb; | |
8df9675f | 1204 | ngroups = ext4_get_groups_count(sb); |
93407472 | 1205 | blocksize = i_blocksize(inode); |
09cbfeaf | 1206 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d | 1207 | |
d3df1453 RH |
1208 | mb_debug(sb, "init page %lu\n", page->index); |
1209 | ||
c9de560d AT |
1210 | groups_per_page = blocks_per_page >> 1; |
1211 | if (groups_per_page == 0) | |
1212 | groups_per_page = 1; | |
1213 | ||
1214 | /* allocate buffer_heads to read bitmaps */ | |
1215 | if (groups_per_page > 1) { | |
c9de560d | 1216 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 1217 | bh = kzalloc(i, gfp); |
813e5727 TT |
1218 | if (bh == NULL) { |
1219 | err = -ENOMEM; | |
c9de560d | 1220 | goto out; |
813e5727 | 1221 | } |
c9de560d AT |
1222 | } else |
1223 | bh = &bhs; | |
1224 | ||
1225 | first_group = page->index * blocks_per_page / 2; | |
1226 | ||
1227 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
1228 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
1229 | if (group >= ngroups) | |
c9de560d AT |
1230 | break; |
1231 | ||
813e5727 | 1232 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
1233 | /* |
1234 | * If page is uptodate then we came here after online resize | |
1235 | * which added some new uninitialized group info structs, so | |
1236 | * we must skip all initialized uptodate buddies on the page, | |
1237 | * which may be currently in use by an allocating task. | |
1238 | */ | |
1239 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
1240 | bh[i] = NULL; | |
1241 | continue; | |
1242 | } | |
cfd73237 | 1243 | bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
9008a58e DW |
1244 | if (IS_ERR(bh[i])) { |
1245 | err = PTR_ERR(bh[i]); | |
1246 | bh[i] = NULL; | |
c9de560d | 1247 | goto out; |
2ccb5fb9 | 1248 | } |
d3df1453 | 1249 | mb_debug(sb, "read bitmap for group %u\n", group); |
c9de560d AT |
1250 | } |
1251 | ||
1252 | /* wait for I/O completion */ | |
813e5727 | 1253 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
1254 | int err2; |
1255 | ||
1256 | if (!bh[i]) | |
1257 | continue; | |
1258 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
1259 | if (!err) | |
1260 | err = err2; | |
813e5727 | 1261 | } |
c9de560d AT |
1262 | |
1263 | first_block = page->index * blocks_per_page; | |
1264 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 1265 | group = (first_block + i) >> 1; |
8df9675f | 1266 | if (group >= ngroups) |
c9de560d AT |
1267 | break; |
1268 | ||
9b8b7d35 AG |
1269 | if (!bh[group - first_group]) |
1270 | /* skip initialized uptodate buddy */ | |
1271 | continue; | |
1272 | ||
bbdc322f LC |
1273 | if (!buffer_verified(bh[group - first_group])) |
1274 | /* Skip faulty bitmaps */ | |
1275 | continue; | |
1276 | err = 0; | |
1277 | ||
c9de560d AT |
1278 | /* |
1279 | * data carry information regarding this | |
1280 | * particular group in the format specified | |
1281 | * above | |
1282 | * | |
1283 | */ | |
1284 | data = page_address(page) + (i * blocksize); | |
1285 | bitmap = bh[group - first_group]->b_data; | |
1286 | ||
1287 | /* | |
1288 | * We place the buddy block and bitmap block | |
1289 | * close together | |
1290 | */ | |
1291 | if ((first_block + i) & 1) { | |
1292 | /* this is block of buddy */ | |
1293 | BUG_ON(incore == NULL); | |
d3df1453 | 1294 | mb_debug(sb, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 1295 | group, page->index, i * blocksize); |
f307333e | 1296 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
1297 | grinfo = ext4_get_group_info(sb, group); |
1298 | grinfo->bb_fragments = 0; | |
1299 | memset(grinfo->bb_counters, 0, | |
1927805e | 1300 | sizeof(*grinfo->bb_counters) * |
4b68f6df | 1301 | (MB_NUM_ORDERS(sb))); |
c9de560d AT |
1302 | /* |
1303 | * incore got set to the group block bitmap below | |
1304 | */ | |
7a2fcbf7 | 1305 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
1306 | /* init the buddy */ |
1307 | memset(data, 0xff, blocksize); | |
c9de560d | 1308 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 1309 | ext4_unlock_group(sb, group); |
c9de560d AT |
1310 | incore = NULL; |
1311 | } else { | |
1312 | /* this is block of bitmap */ | |
1313 | BUG_ON(incore != NULL); | |
d3df1453 | 1314 | mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 1315 | group, page->index, i * blocksize); |
f307333e | 1316 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
1317 | |
1318 | /* see comments in ext4_mb_put_pa() */ | |
1319 | ext4_lock_group(sb, group); | |
1320 | memcpy(data, bitmap, blocksize); | |
1321 | ||
1322 | /* mark all preallocated blks used in in-core bitmap */ | |
1323 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 1324 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
1325 | ext4_unlock_group(sb, group); |
1326 | ||
1327 | /* set incore so that the buddy information can be | |
1328 | * generated using this | |
1329 | */ | |
1330 | incore = data; | |
1331 | } | |
1332 | } | |
1333 | SetPageUptodate(page); | |
1334 | ||
1335 | out: | |
1336 | if (bh) { | |
9b8b7d35 | 1337 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
1338 | brelse(bh[i]); |
1339 | if (bh != &bhs) | |
1340 | kfree(bh); | |
1341 | } | |
1342 | return err; | |
1343 | } | |
1344 | ||
eee4adc7 | 1345 | /* |
2de8807b AG |
1346 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
1347 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1348 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
1349 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 1350 | */ |
2de8807b | 1351 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1352 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1353 | { |
2de8807b AG |
1354 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1355 | int block, pnum, poff; | |
eee4adc7 | 1356 | int blocks_per_page; |
2de8807b AG |
1357 | struct page *page; |
1358 | ||
1359 | e4b->bd_buddy_page = NULL; | |
1360 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 1361 | |
09cbfeaf | 1362 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1363 | /* |
1364 | * the buddy cache inode stores the block bitmap | |
1365 | * and buddy information in consecutive blocks. | |
1366 | * So for each group we need two blocks. | |
1367 | */ | |
1368 | block = group * 2; | |
1369 | pnum = block / blocks_per_page; | |
2de8807b | 1370 | poff = block % blocks_per_page; |
adb7ef60 | 1371 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1372 | if (!page) |
c57ab39b | 1373 | return -ENOMEM; |
2de8807b AG |
1374 | BUG_ON(page->mapping != inode->i_mapping); |
1375 | e4b->bd_bitmap_page = page; | |
1376 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1377 | ||
1378 | if (blocks_per_page >= 2) { | |
1379 | /* buddy and bitmap are on the same page */ | |
1380 | return 0; | |
eee4adc7 | 1381 | } |
2de8807b AG |
1382 | |
1383 | block++; | |
1384 | pnum = block / blocks_per_page; | |
adb7ef60 | 1385 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1386 | if (!page) |
c57ab39b | 1387 | return -ENOMEM; |
2de8807b AG |
1388 | BUG_ON(page->mapping != inode->i_mapping); |
1389 | e4b->bd_buddy_page = page; | |
1390 | return 0; | |
eee4adc7 ES |
1391 | } |
1392 | ||
2de8807b | 1393 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1394 | { |
2de8807b AG |
1395 | if (e4b->bd_bitmap_page) { |
1396 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1397 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1398 | } |
1399 | if (e4b->bd_buddy_page) { | |
1400 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1401 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1402 | } |
eee4adc7 ES |
1403 | } |
1404 | ||
8a57d9d6 CW |
1405 | /* |
1406 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1407 | * block group lock of all groups for this page; do not hold the BG lock when | |
1408 | * calling this routine! | |
1409 | */ | |
b6a758ec | 1410 | static noinline_for_stack |
adb7ef60 | 1411 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1412 | { |
1413 | ||
b6a758ec | 1414 | struct ext4_group_info *this_grp; |
2de8807b AG |
1415 | struct ext4_buddy e4b; |
1416 | struct page *page; | |
1417 | int ret = 0; | |
b6a758ec | 1418 | |
b10a44c3 | 1419 | might_sleep(); |
d3df1453 | 1420 | mb_debug(sb, "init group %u\n", group); |
b6a758ec AK |
1421 | this_grp = ext4_get_group_info(sb, group); |
1422 | /* | |
08c3a813 AK |
1423 | * This ensures that we don't reinit the buddy cache |
1424 | * page which map to the group from which we are already | |
1425 | * allocating. If we are looking at the buddy cache we would | |
1426 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1427 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1428 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1429 | * page accessed. | |
b6a758ec | 1430 | */ |
adb7ef60 | 1431 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1432 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1433 | /* |
1434 | * somebody initialized the group | |
1435 | * return without doing anything | |
1436 | */ | |
b6a758ec AK |
1437 | goto err; |
1438 | } | |
2de8807b AG |
1439 | |
1440 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1441 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1442 | if (ret) |
1443 | goto err; | |
1444 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1445 | ret = -EIO; |
1446 | goto err; | |
1447 | } | |
b6a758ec | 1448 | |
2de8807b | 1449 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1450 | /* |
1451 | * If both the bitmap and buddy are in | |
1452 | * the same page we don't need to force | |
1453 | * init the buddy | |
1454 | */ | |
2de8807b AG |
1455 | ret = 0; |
1456 | goto err; | |
b6a758ec | 1457 | } |
2de8807b AG |
1458 | /* init buddy cache */ |
1459 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1460 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1461 | if (ret) |
1462 | goto err; | |
1463 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1464 | ret = -EIO; |
1465 | goto err; | |
1466 | } | |
b6a758ec | 1467 | err: |
2de8807b | 1468 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1469 | return ret; |
1470 | } | |
1471 | ||
8a57d9d6 CW |
1472 | /* |
1473 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1474 | * block group lock of all groups for this page; do not hold the BG lock when | |
1475 | * calling this routine! | |
1476 | */ | |
4ddfef7b | 1477 | static noinline_for_stack int |
adb7ef60 KK |
1478 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1479 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1480 | { |
c9de560d AT |
1481 | int blocks_per_page; |
1482 | int block; | |
1483 | int pnum; | |
1484 | int poff; | |
1485 | struct page *page; | |
fdf6c7a7 | 1486 | int ret; |
920313a7 AK |
1487 | struct ext4_group_info *grp; |
1488 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1489 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1490 | |
b10a44c3 | 1491 | might_sleep(); |
d3df1453 | 1492 | mb_debug(sb, "load group %u\n", group); |
c9de560d | 1493 | |
09cbfeaf | 1494 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1495 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1496 | |
1497 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1498 | e4b->bd_info = grp; |
c9de560d AT |
1499 | e4b->bd_sb = sb; |
1500 | e4b->bd_group = group; | |
1501 | e4b->bd_buddy_page = NULL; | |
1502 | e4b->bd_bitmap_page = NULL; | |
1503 | ||
f41c0750 | 1504 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1505 | /* |
1506 | * we need full data about the group | |
1507 | * to make a good selection | |
1508 | */ | |
adb7ef60 | 1509 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1510 | if (ret) |
1511 | return ret; | |
f41c0750 AK |
1512 | } |
1513 | ||
c9de560d AT |
1514 | /* |
1515 | * the buddy cache inode stores the block bitmap | |
1516 | * and buddy information in consecutive blocks. | |
1517 | * So for each group we need two blocks. | |
1518 | */ | |
1519 | block = group * 2; | |
1520 | pnum = block / blocks_per_page; | |
1521 | poff = block % blocks_per_page; | |
1522 | ||
1523 | /* we could use find_or_create_page(), but it locks page | |
1524 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1525 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1526 | if (page == NULL || !PageUptodate(page)) { |
1527 | if (page) | |
920313a7 AK |
1528 | /* |
1529 | * drop the page reference and try | |
1530 | * to get the page with lock. If we | |
1531 | * are not uptodate that implies | |
1532 | * somebody just created the page but | |
1533 | * is yet to initialize the same. So | |
1534 | * wait for it to initialize. | |
1535 | */ | |
09cbfeaf | 1536 | put_page(page); |
adb7ef60 | 1537 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1538 | if (page) { |
1539 | BUG_ON(page->mapping != inode->i_mapping); | |
1540 | if (!PageUptodate(page)) { | |
adb7ef60 | 1541 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1542 | if (ret) { |
1543 | unlock_page(page); | |
1544 | goto err; | |
1545 | } | |
c9de560d AT |
1546 | mb_cmp_bitmaps(e4b, page_address(page) + |
1547 | (poff * sb->s_blocksize)); | |
1548 | } | |
1549 | unlock_page(page); | |
1550 | } | |
1551 | } | |
c57ab39b YL |
1552 | if (page == NULL) { |
1553 | ret = -ENOMEM; | |
1554 | goto err; | |
1555 | } | |
1556 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1557 | ret = -EIO; |
c9de560d | 1558 | goto err; |
fdf6c7a7 | 1559 | } |
2457aec6 MG |
1560 | |
1561 | /* Pages marked accessed already */ | |
c9de560d AT |
1562 | e4b->bd_bitmap_page = page; |
1563 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1564 | |
1565 | block++; | |
1566 | pnum = block / blocks_per_page; | |
1567 | poff = block % blocks_per_page; | |
1568 | ||
2457aec6 | 1569 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1570 | if (page == NULL || !PageUptodate(page)) { |
1571 | if (page) | |
09cbfeaf | 1572 | put_page(page); |
adb7ef60 | 1573 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1574 | if (page) { |
1575 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1576 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1577 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1578 | gfp); | |
fdf6c7a7 SF |
1579 | if (ret) { |
1580 | unlock_page(page); | |
1581 | goto err; | |
1582 | } | |
1583 | } | |
c9de560d AT |
1584 | unlock_page(page); |
1585 | } | |
1586 | } | |
c57ab39b YL |
1587 | if (page == NULL) { |
1588 | ret = -ENOMEM; | |
1589 | goto err; | |
1590 | } | |
1591 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1592 | ret = -EIO; |
c9de560d | 1593 | goto err; |
fdf6c7a7 | 1594 | } |
2457aec6 MG |
1595 | |
1596 | /* Pages marked accessed already */ | |
c9de560d AT |
1597 | e4b->bd_buddy_page = page; |
1598 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d | 1599 | |
c9de560d AT |
1600 | return 0; |
1601 | ||
1602 | err: | |
26626f11 | 1603 | if (page) |
09cbfeaf | 1604 | put_page(page); |
c9de560d | 1605 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1606 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1607 | if (e4b->bd_buddy_page) |
09cbfeaf | 1608 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1609 | e4b->bd_buddy = NULL; |
1610 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1611 | return ret; |
c9de560d AT |
1612 | } |
1613 | ||
adb7ef60 KK |
1614 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1615 | struct ext4_buddy *e4b) | |
1616 | { | |
1617 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1618 | } | |
1619 | ||
e39e07fd | 1620 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1621 | { |
1622 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1623 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1624 | if (e4b->bd_buddy_page) |
09cbfeaf | 1625 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1626 | } |
1627 | ||
1628 | ||
1629 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1630 | { | |
ce3cca33 | 1631 | int order = 1, max; |
c9de560d AT |
1632 | void *bb; |
1633 | ||
c5e8f3f3 | 1634 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1635 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1636 | ||
c9de560d | 1637 | while (order <= e4b->bd_blkbits + 1) { |
ce3cca33 CX |
1638 | bb = mb_find_buddy(e4b, order, &max); |
1639 | if (!mb_test_bit(block >> order, bb)) { | |
c9de560d AT |
1640 | /* this block is part of buddy of order 'order' */ |
1641 | return order; | |
1642 | } | |
c9de560d AT |
1643 | order++; |
1644 | } | |
1645 | return 0; | |
1646 | } | |
1647 | ||
955ce5f5 | 1648 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1649 | { |
1650 | __u32 *addr; | |
1651 | ||
1652 | len = cur + len; | |
1653 | while (cur < len) { | |
1654 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1655 | /* fast path: clear whole word at once */ | |
1656 | addr = bm + (cur >> 3); | |
1657 | *addr = 0; | |
1658 | cur += 32; | |
1659 | continue; | |
1660 | } | |
955ce5f5 | 1661 | mb_clear_bit(cur, bm); |
c9de560d AT |
1662 | cur++; |
1663 | } | |
1664 | } | |
1665 | ||
eabe0444 AS |
1666 | /* clear bits in given range |
1667 | * will return first found zero bit if any, -1 otherwise | |
1668 | */ | |
1669 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1670 | { | |
1671 | __u32 *addr; | |
1672 | int zero_bit = -1; | |
1673 | ||
1674 | len = cur + len; | |
1675 | while (cur < len) { | |
1676 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1677 | /* fast path: clear whole word at once */ | |
1678 | addr = bm + (cur >> 3); | |
1679 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1680 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1681 | *addr = 0; | |
1682 | cur += 32; | |
1683 | continue; | |
1684 | } | |
1685 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1686 | zero_bit = cur; | |
1687 | cur++; | |
1688 | } | |
1689 | ||
1690 | return zero_bit; | |
1691 | } | |
1692 | ||
123e3016 | 1693 | void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1694 | { |
1695 | __u32 *addr; | |
1696 | ||
1697 | len = cur + len; | |
1698 | while (cur < len) { | |
1699 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1700 | /* fast path: set whole word at once */ | |
1701 | addr = bm + (cur >> 3); | |
1702 | *addr = 0xffffffff; | |
1703 | cur += 32; | |
1704 | continue; | |
1705 | } | |
955ce5f5 | 1706 | mb_set_bit(cur, bm); |
c9de560d AT |
1707 | cur++; |
1708 | } | |
1709 | } | |
1710 | ||
eabe0444 AS |
1711 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
1712 | { | |
1713 | if (mb_test_bit(*bit + side, bitmap)) { | |
1714 | mb_clear_bit(*bit, bitmap); | |
1715 | (*bit) -= side; | |
1716 | return 1; | |
1717 | } | |
1718 | else { | |
1719 | (*bit) += side; | |
1720 | mb_set_bit(*bit, bitmap); | |
1721 | return -1; | |
1722 | } | |
1723 | } | |
1724 | ||
1725 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1726 | { | |
1727 | int max; | |
1728 | int order = 1; | |
1729 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1730 | ||
1731 | while (buddy) { | |
1732 | void *buddy2; | |
1733 | ||
1734 | /* Bits in range [first; last] are known to be set since | |
1735 | * corresponding blocks were allocated. Bits in range | |
1736 | * (first; last) will stay set because they form buddies on | |
1737 | * upper layer. We just deal with borders if they don't | |
1738 | * align with upper layer and then go up. | |
1739 | * Releasing entire group is all about clearing | |
1740 | * single bit of highest order buddy. | |
1741 | */ | |
1742 | ||
1743 | /* Example: | |
1744 | * --------------------------------- | |
1745 | * | 1 | 1 | 1 | 1 | | |
1746 | * --------------------------------- | |
1747 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1748 | * --------------------------------- | |
1749 | * 0 1 2 3 4 5 6 7 | |
1750 | * \_____________________/ | |
1751 | * | |
1752 | * Neither [1] nor [6] is aligned to above layer. | |
1753 | * Left neighbour [0] is free, so mark it busy, | |
1754 | * decrease bb_counters and extend range to | |
1755 | * [0; 6] | |
1756 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1757 | * mark [6] free, increase bb_counters and shrink range to | |
1758 | * [0; 5]. | |
1759 | * Then shift range to [0; 2], go up and do the same. | |
1760 | */ | |
1761 | ||
1762 | ||
1763 | if (first & 1) | |
1764 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1765 | if (!(last & 1)) | |
1766 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1767 | if (first > last) | |
1768 | break; | |
1769 | order++; | |
1770 | ||
1771 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1772 | mb_clear_bits(buddy, first, last - first + 1); | |
1773 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1774 | break; | |
1775 | } | |
1776 | first >>= 1; | |
1777 | last >>= 1; | |
1778 | buddy = buddy2; | |
1779 | } | |
1780 | } | |
1781 | ||
7e5a8cdd | 1782 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1783 | int first, int count) |
c9de560d | 1784 | { |
eabe0444 AS |
1785 | int left_is_free = 0; |
1786 | int right_is_free = 0; | |
1787 | int block; | |
1788 | int last = first + count - 1; | |
c9de560d AT |
1789 | struct super_block *sb = e4b->bd_sb; |
1790 | ||
c99d1e6e TT |
1791 | if (WARN_ON(count == 0)) |
1792 | return; | |
eabe0444 | 1793 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1794 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1795 | /* Don't bother if the block group is corrupt. */ |
1796 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1797 | return; | |
1798 | ||
c9de560d AT |
1799 | mb_check_buddy(e4b); |
1800 | mb_free_blocks_double(inode, e4b, first, count); | |
1801 | ||
07b5b8e1 | 1802 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1803 | e4b->bd_info->bb_free += count; |
1804 | if (first < e4b->bd_info->bb_first_free) | |
1805 | e4b->bd_info->bb_first_free = first; | |
1806 | ||
eabe0444 AS |
1807 | /* access memory sequentially: check left neighbour, |
1808 | * clear range and then check right neighbour | |
1809 | */ | |
c9de560d | 1810 | if (first != 0) |
eabe0444 AS |
1811 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1812 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1813 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1814 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1815 | ||
1816 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1817 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1818 | ext4_fsblk_t blocknr; |
1819 | ||
1820 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1821 | blocknr += EXT4_C2B(sbi, block); |
8016e29f HS |
1822 | if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { |
1823 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1824 | inode ? inode->i_ino : 0, | |
1825 | blocknr, | |
1826 | "freeing already freed block (bit %u); block bitmap corrupt.", | |
1827 | block); | |
1828 | ext4_mark_group_bitmap_corrupted( | |
1829 | sb, e4b->bd_group, | |
db79e6d1 | 1830 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
8016e29f | 1831 | } |
eabe0444 AS |
1832 | goto done; |
1833 | } | |
1834 | ||
1835 | /* let's maintain fragments counter */ | |
1836 | if (left_is_free && right_is_free) | |
c9de560d | 1837 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1838 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1839 | e4b->bd_info->bb_fragments++; |
1840 | ||
eabe0444 AS |
1841 | /* buddy[0] == bd_bitmap is a special case, so handle |
1842 | * it right away and let mb_buddy_mark_free stay free of | |
1843 | * zero order checks. | |
1844 | * Check if neighbours are to be coaleasced, | |
1845 | * adjust bitmap bb_counters and borders appropriately. | |
1846 | */ | |
1847 | if (first & 1) { | |
1848 | first += !left_is_free; | |
1849 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1850 | } | |
1851 | if (!(last & 1)) { | |
1852 | last -= !right_is_free; | |
1853 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1854 | } | |
c9de560d | 1855 | |
eabe0444 AS |
1856 | if (first <= last) |
1857 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1858 | |
eabe0444 | 1859 | done: |
8a57d9d6 | 1860 | mb_set_largest_free_order(sb, e4b->bd_info); |
196e402a | 1861 | mb_update_avg_fragment_size(sb, e4b->bd_info); |
c9de560d | 1862 | mb_check_buddy(e4b); |
c9de560d AT |
1863 | } |
1864 | ||
15c006a2 | 1865 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1866 | int needed, struct ext4_free_extent *ex) |
1867 | { | |
1868 | int next = block; | |
15c006a2 | 1869 | int max, order; |
c9de560d AT |
1870 | void *buddy; |
1871 | ||
bc8e6740 | 1872 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1873 | BUG_ON(ex == NULL); |
1874 | ||
15c006a2 | 1875 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1876 | BUG_ON(buddy == NULL); |
1877 | BUG_ON(block >= max); | |
1878 | if (mb_test_bit(block, buddy)) { | |
1879 | ex->fe_len = 0; | |
1880 | ex->fe_start = 0; | |
1881 | ex->fe_group = 0; | |
1882 | return 0; | |
1883 | } | |
1884 | ||
15c006a2 RD |
1885 | /* find actual order */ |
1886 | order = mb_find_order_for_block(e4b, block); | |
1887 | block = block >> order; | |
c9de560d AT |
1888 | |
1889 | ex->fe_len = 1 << order; | |
1890 | ex->fe_start = block << order; | |
1891 | ex->fe_group = e4b->bd_group; | |
1892 | ||
1893 | /* calc difference from given start */ | |
1894 | next = next - ex->fe_start; | |
1895 | ex->fe_len -= next; | |
1896 | ex->fe_start += next; | |
1897 | ||
1898 | while (needed > ex->fe_len && | |
d8ec0c39 | 1899 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1900 | |
1901 | if (block + 1 >= max) | |
1902 | break; | |
1903 | ||
1904 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1905 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1906 | break; |
1907 | ||
b051d8dc | 1908 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1909 | |
c9de560d AT |
1910 | block = next >> order; |
1911 | ex->fe_len += 1 << order; | |
1912 | } | |
1913 | ||
31562b95 | 1914 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1915 | /* Should never happen! (but apparently sometimes does?!?) */ |
1916 | WARN_ON(1); | |
cd84bbba SB |
1917 | ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, |
1918 | "corruption or bug in mb_find_extent " | |
1919 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1920 | block, order, needed, ex->fe_group, ex->fe_start, | |
1921 | ex->fe_len, ex->fe_logical); | |
43c73221 TT |
1922 | ex->fe_len = 0; |
1923 | ex->fe_start = 0; | |
1924 | ex->fe_group = 0; | |
1925 | } | |
c9de560d AT |
1926 | return ex->fe_len; |
1927 | } | |
1928 | ||
1929 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1930 | { | |
1931 | int ord; | |
1932 | int mlen = 0; | |
1933 | int max = 0; | |
1934 | int cur; | |
1935 | int start = ex->fe_start; | |
1936 | int len = ex->fe_len; | |
1937 | unsigned ret = 0; | |
1938 | int len0 = len; | |
1939 | void *buddy; | |
218a6944 | 1940 | bool split = false; |
c9de560d AT |
1941 | |
1942 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1943 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1944 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1945 | mb_check_buddy(e4b); |
1946 | mb_mark_used_double(e4b, start, len); | |
1947 | ||
07b5b8e1 | 1948 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1949 | e4b->bd_info->bb_free -= len; |
1950 | if (e4b->bd_info->bb_first_free == start) | |
1951 | e4b->bd_info->bb_first_free += len; | |
1952 | ||
1953 | /* let's maintain fragments counter */ | |
1954 | if (start != 0) | |
c5e8f3f3 | 1955 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1956 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1957 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1958 | if (mlen && max) |
1959 | e4b->bd_info->bb_fragments++; | |
1960 | else if (!mlen && !max) | |
1961 | e4b->bd_info->bb_fragments--; | |
1962 | ||
1963 | /* let's maintain buddy itself */ | |
1964 | while (len) { | |
218a6944 | 1965 | if (!split) |
1966 | ord = mb_find_order_for_block(e4b, start); | |
c9de560d AT |
1967 | |
1968 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1969 | /* the whole chunk may be allocated at once! */ | |
1970 | mlen = 1 << ord; | |
218a6944 | 1971 | if (!split) |
1972 | buddy = mb_find_buddy(e4b, ord, &max); | |
1973 | else | |
1974 | split = false; | |
c9de560d AT |
1975 | BUG_ON((start >> ord) >= max); |
1976 | mb_set_bit(start >> ord, buddy); | |
1977 | e4b->bd_info->bb_counters[ord]--; | |
1978 | start += mlen; | |
1979 | len -= mlen; | |
1980 | BUG_ON(len < 0); | |
1981 | continue; | |
1982 | } | |
1983 | ||
1984 | /* store for history */ | |
1985 | if (ret == 0) | |
1986 | ret = len | (ord << 16); | |
1987 | ||
1988 | /* we have to split large buddy */ | |
1989 | BUG_ON(ord <= 0); | |
1990 | buddy = mb_find_buddy(e4b, ord, &max); | |
1991 | mb_set_bit(start >> ord, buddy); | |
1992 | e4b->bd_info->bb_counters[ord]--; | |
1993 | ||
1994 | ord--; | |
1995 | cur = (start >> ord) & ~1U; | |
1996 | buddy = mb_find_buddy(e4b, ord, &max); | |
1997 | mb_clear_bit(cur, buddy); | |
1998 | mb_clear_bit(cur + 1, buddy); | |
1999 | e4b->bd_info->bb_counters[ord]++; | |
2000 | e4b->bd_info->bb_counters[ord]++; | |
218a6944 | 2001 | split = true; |
c9de560d | 2002 | } |
8a57d9d6 | 2003 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 2004 | |
196e402a | 2005 | mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); |
123e3016 | 2006 | mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
2007 | mb_check_buddy(e4b); |
2008 | ||
2009 | return ret; | |
2010 | } | |
2011 | ||
2012 | /* | |
2013 | * Must be called under group lock! | |
2014 | */ | |
2015 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
2016 | struct ext4_buddy *e4b) | |
2017 | { | |
2018 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2019 | int ret; | |
2020 | ||
2021 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
2022 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
2023 | ||
2024 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
2025 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
2026 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
2027 | ||
2028 | /* preallocation can change ac_b_ex, thus we store actually | |
2029 | * allocated blocks for history */ | |
2030 | ac->ac_f_ex = ac->ac_b_ex; | |
2031 | ||
2032 | ac->ac_status = AC_STATUS_FOUND; | |
2033 | ac->ac_tail = ret & 0xffff; | |
2034 | ac->ac_buddy = ret >> 16; | |
2035 | ||
c3a326a6 AK |
2036 | /* |
2037 | * take the page reference. We want the page to be pinned | |
2038 | * so that we don't get a ext4_mb_init_cache_call for this | |
2039 | * group until we update the bitmap. That would mean we | |
2040 | * double allocate blocks. The reference is dropped | |
2041 | * in ext4_mb_release_context | |
2042 | */ | |
c9de560d AT |
2043 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
2044 | get_page(ac->ac_bitmap_page); | |
2045 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
2046 | get_page(ac->ac_buddy_page); | |
c9de560d | 2047 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 2048 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
2049 | spin_lock(&sbi->s_md_lock); |
2050 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
2051 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
2052 | spin_unlock(&sbi->s_md_lock); | |
2053 | } | |
53f86b17 RH |
2054 | /* |
2055 | * As we've just preallocated more space than | |
2056 | * user requested originally, we store allocated | |
2057 | * space in a special descriptor. | |
2058 | */ | |
2059 | if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
2060 | ext4_mb_new_preallocation(ac); | |
2061 | ||
c9de560d AT |
2062 | } |
2063 | ||
c9de560d AT |
2064 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
2065 | struct ext4_buddy *e4b, | |
2066 | int finish_group) | |
2067 | { | |
2068 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2069 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2070 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2071 | struct ext4_free_extent ex; | |
2072 | int max; | |
2073 | ||
032115fc AK |
2074 | if (ac->ac_status == AC_STATUS_FOUND) |
2075 | return; | |
c9de560d AT |
2076 | /* |
2077 | * We don't want to scan for a whole year | |
2078 | */ | |
2079 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
2080 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2081 | ac->ac_status = AC_STATUS_BREAK; | |
2082 | return; | |
2083 | } | |
2084 | ||
2085 | /* | |
2086 | * Haven't found good chunk so far, let's continue | |
2087 | */ | |
2088 | if (bex->fe_len < gex->fe_len) | |
2089 | return; | |
2090 | ||
2091 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
2092 | && bex->fe_group == e4b->bd_group) { | |
2093 | /* recheck chunk's availability - we don't know | |
2094 | * when it was found (within this lock-unlock | |
2095 | * period or not) */ | |
15c006a2 | 2096 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
2097 | if (max >= gex->fe_len) { |
2098 | ext4_mb_use_best_found(ac, e4b); | |
2099 | return; | |
2100 | } | |
2101 | } | |
2102 | } | |
2103 | ||
2104 | /* | |
2105 | * The routine checks whether found extent is good enough. If it is, | |
2106 | * then the extent gets marked used and flag is set to the context | |
2107 | * to stop scanning. Otherwise, the extent is compared with the | |
2108 | * previous found extent and if new one is better, then it's stored | |
2109 | * in the context. Later, the best found extent will be used, if | |
2110 | * mballoc can't find good enough extent. | |
2111 | * | |
2112 | * FIXME: real allocation policy is to be designed yet! | |
2113 | */ | |
2114 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
2115 | struct ext4_free_extent *ex, | |
2116 | struct ext4_buddy *e4b) | |
2117 | { | |
2118 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2119 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2120 | ||
2121 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
2122 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
2123 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
2124 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
2125 | ||
2126 | ac->ac_found++; | |
2127 | ||
2128 | /* | |
2129 | * The special case - take what you catch first | |
2130 | */ | |
2131 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2132 | *bex = *ex; | |
2133 | ext4_mb_use_best_found(ac, e4b); | |
2134 | return; | |
2135 | } | |
2136 | ||
2137 | /* | |
2138 | * Let's check whether the chuck is good enough | |
2139 | */ | |
2140 | if (ex->fe_len == gex->fe_len) { | |
2141 | *bex = *ex; | |
2142 | ext4_mb_use_best_found(ac, e4b); | |
2143 | return; | |
2144 | } | |
2145 | ||
2146 | /* | |
2147 | * If this is first found extent, just store it in the context | |
2148 | */ | |
2149 | if (bex->fe_len == 0) { | |
2150 | *bex = *ex; | |
2151 | return; | |
2152 | } | |
2153 | ||
2154 | /* | |
2155 | * If new found extent is better, store it in the context | |
2156 | */ | |
2157 | if (bex->fe_len < gex->fe_len) { | |
2158 | /* if the request isn't satisfied, any found extent | |
2159 | * larger than previous best one is better */ | |
2160 | if (ex->fe_len > bex->fe_len) | |
2161 | *bex = *ex; | |
2162 | } else if (ex->fe_len > gex->fe_len) { | |
2163 | /* if the request is satisfied, then we try to find | |
2164 | * an extent that still satisfy the request, but is | |
2165 | * smaller than previous one */ | |
2166 | if (ex->fe_len < bex->fe_len) | |
2167 | *bex = *ex; | |
2168 | } | |
2169 | ||
2170 | ext4_mb_check_limits(ac, e4b, 0); | |
2171 | } | |
2172 | ||
089ceecc ES |
2173 | static noinline_for_stack |
2174 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
2175 | struct ext4_buddy *e4b) |
2176 | { | |
2177 | struct ext4_free_extent ex = ac->ac_b_ex; | |
2178 | ext4_group_t group = ex.fe_group; | |
2179 | int max; | |
2180 | int err; | |
2181 | ||
2182 | BUG_ON(ex.fe_len <= 0); | |
2183 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2184 | if (err) | |
2185 | return err; | |
2186 | ||
2187 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 2188 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
2189 | |
2190 | if (max > 0) { | |
2191 | ac->ac_b_ex = ex; | |
2192 | ext4_mb_use_best_found(ac, e4b); | |
2193 | } | |
2194 | ||
2195 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2196 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2197 | |
2198 | return 0; | |
2199 | } | |
2200 | ||
089ceecc ES |
2201 | static noinline_for_stack |
2202 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
2203 | struct ext4_buddy *e4b) |
2204 | { | |
2205 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
2206 | int max; | |
2207 | int err; | |
2208 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 2209 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
2210 | struct ext4_free_extent ex; |
2211 | ||
2212 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
2213 | return 0; | |
838cd0cf YY |
2214 | if (grp->bb_free == 0) |
2215 | return 0; | |
c9de560d AT |
2216 | |
2217 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2218 | if (err) | |
2219 | return err; | |
2220 | ||
163a203d DW |
2221 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
2222 | ext4_mb_unload_buddy(e4b); | |
2223 | return 0; | |
2224 | } | |
2225 | ||
c9de560d | 2226 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 2227 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 2228 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 2229 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
2230 | |
2231 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
2232 | ext4_fsblk_t start; | |
2233 | ||
5661bd68 AM |
2234 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
2235 | ex.fe_start; | |
c9de560d AT |
2236 | /* use do_div to get remainder (would be 64-bit modulo) */ |
2237 | if (do_div(start, sbi->s_stripe) == 0) { | |
2238 | ac->ac_found++; | |
2239 | ac->ac_b_ex = ex; | |
2240 | ext4_mb_use_best_found(ac, e4b); | |
2241 | } | |
2242 | } else if (max >= ac->ac_g_ex.fe_len) { | |
2243 | BUG_ON(ex.fe_len <= 0); | |
2244 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2245 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2246 | ac->ac_found++; | |
2247 | ac->ac_b_ex = ex; | |
2248 | ext4_mb_use_best_found(ac, e4b); | |
2249 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
2250 | /* Sometimes, caller may want to merge even small | |
2251 | * number of blocks to an existing extent */ | |
2252 | BUG_ON(ex.fe_len <= 0); | |
2253 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2254 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2255 | ac->ac_found++; | |
2256 | ac->ac_b_ex = ex; | |
2257 | ext4_mb_use_best_found(ac, e4b); | |
2258 | } | |
2259 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2260 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2261 | |
2262 | return 0; | |
2263 | } | |
2264 | ||
2265 | /* | |
2266 | * The routine scans buddy structures (not bitmap!) from given order | |
2267 | * to max order and tries to find big enough chunk to satisfy the req | |
2268 | */ | |
089ceecc ES |
2269 | static noinline_for_stack |
2270 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2271 | struct ext4_buddy *e4b) |
2272 | { | |
2273 | struct super_block *sb = ac->ac_sb; | |
2274 | struct ext4_group_info *grp = e4b->bd_info; | |
2275 | void *buddy; | |
2276 | int i; | |
2277 | int k; | |
2278 | int max; | |
2279 | ||
2280 | BUG_ON(ac->ac_2order <= 0); | |
4b68f6df | 2281 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { |
c9de560d AT |
2282 | if (grp->bb_counters[i] == 0) |
2283 | continue; | |
2284 | ||
2285 | buddy = mb_find_buddy(e4b, i, &max); | |
2286 | BUG_ON(buddy == NULL); | |
2287 | ||
ffad0a44 | 2288 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
2289 | if (k >= max) { |
2290 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
2291 | "%d free clusters of order %d. But found 0", | |
2292 | grp->bb_counters[i], i); | |
2293 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
2294 | e4b->bd_group, | |
2295 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
2296 | break; | |
2297 | } | |
c9de560d AT |
2298 | ac->ac_found++; |
2299 | ||
2300 | ac->ac_b_ex.fe_len = 1 << i; | |
2301 | ac->ac_b_ex.fe_start = k << i; | |
2302 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
2303 | ||
2304 | ext4_mb_use_best_found(ac, e4b); | |
2305 | ||
53f86b17 | 2306 | BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
c9de560d AT |
2307 | |
2308 | if (EXT4_SB(sb)->s_mb_stats) | |
2309 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
2310 | ||
2311 | break; | |
2312 | } | |
2313 | } | |
2314 | ||
2315 | /* | |
2316 | * The routine scans the group and measures all found extents. | |
2317 | * In order to optimize scanning, caller must pass number of | |
2318 | * free blocks in the group, so the routine can know upper limit. | |
2319 | */ | |
089ceecc ES |
2320 | static noinline_for_stack |
2321 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2322 | struct ext4_buddy *e4b) |
2323 | { | |
2324 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 2325 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2326 | struct ext4_free_extent ex; |
2327 | int i; | |
2328 | int free; | |
2329 | ||
2330 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
2331 | if (WARN_ON(free <= 0)) |
2332 | return; | |
c9de560d AT |
2333 | |
2334 | i = e4b->bd_info->bb_first_free; | |
2335 | ||
2336 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 2337 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
2338 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
2339 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 2340 | /* |
e56eb659 | 2341 | * IF we have corrupt bitmap, we won't find any |
26346ff6 | 2342 | * free blocks even though group info says we |
b483bb77 | 2343 | * have free blocks |
26346ff6 | 2344 | */ |
e29136f8 | 2345 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2346 | "%d free clusters as per " |
fde4d95a | 2347 | "group info. But bitmap says 0", |
26346ff6 | 2348 | free); |
736dedbb WS |
2349 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2350 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
2351 | break; |
2352 | } | |
2353 | ||
15c006a2 | 2354 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2355 | if (WARN_ON(ex.fe_len <= 0)) |
2356 | break; | |
26346ff6 | 2357 | if (free < ex.fe_len) { |
e29136f8 | 2358 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2359 | "%d free clusters as per " |
fde4d95a | 2360 | "group info. But got %d blocks", |
26346ff6 | 2361 | free, ex.fe_len); |
736dedbb WS |
2362 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2363 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
2364 | /* |
2365 | * The number of free blocks differs. This mostly | |
2366 | * indicate that the bitmap is corrupt. So exit | |
2367 | * without claiming the space. | |
2368 | */ | |
2369 | break; | |
26346ff6 | 2370 | } |
ab0c00fc | 2371 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2372 | ext4_mb_measure_extent(ac, &ex, e4b); |
2373 | ||
2374 | i += ex.fe_len; | |
2375 | free -= ex.fe_len; | |
2376 | } | |
2377 | ||
2378 | ext4_mb_check_limits(ac, e4b, 1); | |
2379 | } | |
2380 | ||
2381 | /* | |
2382 | * This is a special case for storages like raid5 | |
506bf2d8 | 2383 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2384 | */ |
089ceecc ES |
2385 | static noinline_for_stack |
2386 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2387 | struct ext4_buddy *e4b) |
2388 | { | |
2389 | struct super_block *sb = ac->ac_sb; | |
2390 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2391 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2392 | struct ext4_free_extent ex; |
2393 | ext4_fsblk_t first_group_block; | |
2394 | ext4_fsblk_t a; | |
2395 | ext4_grpblk_t i; | |
2396 | int max; | |
2397 | ||
2398 | BUG_ON(sbi->s_stripe == 0); | |
2399 | ||
2400 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2401 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2402 | ||
c9de560d AT |
2403 | a = first_group_block + sbi->s_stripe - 1; |
2404 | do_div(a, sbi->s_stripe); | |
2405 | i = (a * sbi->s_stripe) - first_group_block; | |
2406 | ||
7137d7a4 | 2407 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2408 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2409 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2410 | if (max >= sbi->s_stripe) { |
2411 | ac->ac_found++; | |
ab0c00fc | 2412 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2413 | ac->ac_b_ex = ex; |
2414 | ext4_mb_use_best_found(ac, e4b); | |
2415 | break; | |
2416 | } | |
2417 | } | |
2418 | i += sbi->s_stripe; | |
2419 | } | |
2420 | } | |
2421 | ||
42ac1848 | 2422 | /* |
8ef123fe | 2423 | * This is also called BEFORE we load the buddy bitmap. |
42ac1848 | 2424 | * Returns either 1 or 0 indicating that the group is either suitable |
8ef123fe | 2425 | * for the allocation or not. |
42ac1848 | 2426 | */ |
8ef123fe | 2427 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
c9de560d AT |
2428 | ext4_group_t group, int cr) |
2429 | { | |
8ef123fe | 2430 | ext4_grpblk_t free, fragments; |
a4912123 | 2431 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2432 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2433 | ||
2434 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2435 | |
dddcd2f9 | 2436 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
8ef123fe | 2437 | return false; |
01fc48e8 | 2438 | |
dddcd2f9 | 2439 | free = grp->bb_free; |
2440 | if (free == 0) | |
8ef123fe | 2441 | return false; |
c9de560d | 2442 | |
c9de560d | 2443 | fragments = grp->bb_fragments; |
c9de560d | 2444 | if (fragments == 0) |
8ef123fe | 2445 | return false; |
c9de560d AT |
2446 | |
2447 | switch (cr) { | |
2448 | case 0: | |
2449 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2450 | |
a4912123 TT |
2451 | /* Avoid using the first bg of a flexgroup for data files */ |
2452 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2453 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2454 | ((group % flex_size) == 0)) | |
8ef123fe | 2455 | return false; |
a4912123 | 2456 | |
dddcd2f9 | 2457 | if (free < ac->ac_g_ex.fe_len) |
2458 | return false; | |
2459 | ||
4b68f6df | 2460 | if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) |
8ef123fe | 2461 | return true; |
40ae3487 TT |
2462 | |
2463 | if (grp->bb_largest_free_order < ac->ac_2order) | |
8ef123fe | 2464 | return false; |
40ae3487 | 2465 | |
8ef123fe | 2466 | return true; |
c9de560d AT |
2467 | case 1: |
2468 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2469 | return true; |
c9de560d AT |
2470 | break; |
2471 | case 2: | |
2472 | if (free >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2473 | return true; |
c9de560d AT |
2474 | break; |
2475 | case 3: | |
8ef123fe | 2476 | return true; |
c9de560d AT |
2477 | default: |
2478 | BUG(); | |
2479 | } | |
2480 | ||
8ef123fe RH |
2481 | return false; |
2482 | } | |
2483 | ||
2484 | /* | |
2485 | * This could return negative error code if something goes wrong | |
2486 | * during ext4_mb_init_group(). This should not be called with | |
2487 | * ext4_lock_group() held. | |
a5fda113 TT |
2488 | * |
2489 | * Note: because we are conditionally operating with the group lock in | |
2490 | * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this | |
2491 | * function using __acquire and __release. This means we need to be | |
2492 | * super careful before messing with the error path handling via "goto | |
2493 | * out"! | |
8ef123fe RH |
2494 | */ |
2495 | static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, | |
2496 | ext4_group_t group, int cr) | |
2497 | { | |
2498 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
99377830 | 2499 | struct super_block *sb = ac->ac_sb; |
c1d2c7d4 | 2500 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
99377830 | 2501 | bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
8ef123fe RH |
2502 | ext4_grpblk_t free; |
2503 | int ret = 0; | |
2504 | ||
a6c75eaf HS |
2505 | if (sbi->s_mb_stats) |
2506 | atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); | |
a5fda113 | 2507 | if (should_lock) { |
99377830 | 2508 | ext4_lock_group(sb, group); |
a5fda113 TT |
2509 | __release(ext4_group_lock_ptr(sb, group)); |
2510 | } | |
8ef123fe RH |
2511 | free = grp->bb_free; |
2512 | if (free == 0) | |
2513 | goto out; | |
2514 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2515 | goto out; | |
2516 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
2517 | goto out; | |
a5fda113 TT |
2518 | if (should_lock) { |
2519 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2520 | ext4_unlock_group(sb, group); |
a5fda113 | 2521 | } |
8ef123fe RH |
2522 | |
2523 | /* We only do this if the grp has never been initialized */ | |
2524 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
c1d2c7d4 AZ |
2525 | struct ext4_group_desc *gdp = |
2526 | ext4_get_group_desc(sb, group, NULL); | |
2527 | int ret; | |
2528 | ||
2529 | /* cr=0/1 is a very optimistic search to find large | |
2530 | * good chunks almost for free. If buddy data is not | |
2531 | * ready, then this optimization makes no sense. But | |
2532 | * we never skip the first block group in a flex_bg, | |
2533 | * since this gets used for metadata block allocation, | |
2534 | * and we want to make sure we locate metadata blocks | |
2535 | * in the first block group in the flex_bg if possible. | |
2536 | */ | |
2537 | if (cr < 2 && | |
2538 | (!sbi->s_log_groups_per_flex || | |
2539 | ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && | |
2540 | !(ext4_has_group_desc_csum(sb) && | |
2541 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) | |
2542 | return 0; | |
2543 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); | |
8ef123fe RH |
2544 | if (ret) |
2545 | return ret; | |
2546 | } | |
2547 | ||
a5fda113 | 2548 | if (should_lock) { |
99377830 | 2549 | ext4_lock_group(sb, group); |
a5fda113 TT |
2550 | __release(ext4_group_lock_ptr(sb, group)); |
2551 | } | |
8ef123fe RH |
2552 | ret = ext4_mb_good_group(ac, group, cr); |
2553 | out: | |
a5fda113 TT |
2554 | if (should_lock) { |
2555 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2556 | ext4_unlock_group(sb, group); |
a5fda113 | 2557 | } |
8ef123fe | 2558 | return ret; |
c9de560d AT |
2559 | } |
2560 | ||
cfd73237 AZ |
2561 | /* |
2562 | * Start prefetching @nr block bitmaps starting at @group. | |
2563 | * Return the next group which needs to be prefetched. | |
2564 | */ | |
3d392b26 TT |
2565 | ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
2566 | unsigned int nr, int *cnt) | |
cfd73237 AZ |
2567 | { |
2568 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2569 | struct buffer_head *bh; | |
2570 | struct blk_plug plug; | |
2571 | ||
2572 | blk_start_plug(&plug); | |
2573 | while (nr-- > 0) { | |
2574 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2575 | NULL); | |
2576 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2577 | ||
2578 | /* | |
2579 | * Prefetch block groups with free blocks; but don't | |
2580 | * bother if it is marked uninitialized on disk, since | |
2581 | * it won't require I/O to read. Also only try to | |
2582 | * prefetch once, so we avoid getblk() call, which can | |
2583 | * be expensive. | |
2584 | */ | |
2585 | if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) && | |
2586 | EXT4_MB_GRP_NEED_INIT(grp) && | |
2587 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2588 | !(ext4_has_group_desc_csum(sb) && | |
2589 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2590 | bh = ext4_read_block_bitmap_nowait(sb, group, true); | |
2591 | if (bh && !IS_ERR(bh)) { | |
2592 | if (!buffer_uptodate(bh) && cnt) | |
2593 | (*cnt)++; | |
2594 | brelse(bh); | |
2595 | } | |
2596 | } | |
2597 | if (++group >= ngroups) | |
2598 | group = 0; | |
2599 | } | |
2600 | blk_finish_plug(&plug); | |
2601 | return group; | |
2602 | } | |
2603 | ||
2604 | /* | |
2605 | * Prefetching reads the block bitmap into the buffer cache; but we | |
2606 | * need to make sure that the buddy bitmap in the page cache has been | |
2607 | * initialized. Note that ext4_mb_init_group() will block if the I/O | |
2608 | * is not yet completed, or indeed if it was not initiated by | |
2609 | * ext4_mb_prefetch did not start the I/O. | |
2610 | * | |
2611 | * TODO: We should actually kick off the buddy bitmap setup in a work | |
2612 | * queue when the buffer I/O is completed, so that we don't block | |
2613 | * waiting for the block allocation bitmap read to finish when | |
2614 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). | |
2615 | */ | |
3d392b26 TT |
2616 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
2617 | unsigned int nr) | |
cfd73237 AZ |
2618 | { |
2619 | while (nr-- > 0) { | |
2620 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2621 | NULL); | |
2622 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2623 | ||
2624 | if (!group) | |
2625 | group = ext4_get_groups_count(sb); | |
2626 | group--; | |
2627 | grp = ext4_get_group_info(sb, group); | |
2628 | ||
2629 | if (EXT4_MB_GRP_NEED_INIT(grp) && | |
2630 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2631 | !(ext4_has_group_desc_csum(sb) && | |
2632 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2633 | if (ext4_mb_init_group(sb, group, GFP_NOFS)) | |
2634 | break; | |
2635 | } | |
2636 | } | |
2637 | } | |
2638 | ||
4ddfef7b ES |
2639 | static noinline_for_stack int |
2640 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2641 | { |
cfd73237 | 2642 | ext4_group_t prefetch_grp = 0, ngroups, group, i; |
4fca50d4 | 2643 | int cr = -1, new_cr; |
42ac1848 | 2644 | int err = 0, first_err = 0; |
cfd73237 | 2645 | unsigned int nr = 0, prefetch_ios = 0; |
c9de560d AT |
2646 | struct ext4_sb_info *sbi; |
2647 | struct super_block *sb; | |
2648 | struct ext4_buddy e4b; | |
66d5e027 | 2649 | int lost; |
c9de560d AT |
2650 | |
2651 | sb = ac->ac_sb; | |
2652 | sbi = EXT4_SB(sb); | |
8df9675f | 2653 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2654 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2655 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2656 | ngroups = sbi->s_blockfile_groups; |
2657 | ||
c9de560d AT |
2658 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2659 | ||
2660 | /* first, try the goal */ | |
2661 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2662 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2663 | goto out; | |
2664 | ||
2665 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2666 | goto out; | |
2667 | ||
2668 | /* | |
e9a3cd48 | 2669 | * ac->ac_2order is set only if the fe_len is a power of 2 |
2670 | * if ac->ac_2order is set we also set criteria to 0 so that we | |
c9de560d AT |
2671 | * try exact allocation using buddy. |
2672 | */ | |
2673 | i = fls(ac->ac_g_ex.fe_len); | |
2674 | ac->ac_2order = 0; | |
2675 | /* | |
2676 | * We search using buddy data only if the order of the request | |
2677 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2678 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2679 | * We also support searching for power-of-two requests only for |
2680 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2681 | */ |
4b68f6df | 2682 | if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { |
c9de560d AT |
2683 | /* |
2684 | * This should tell if fe_len is exactly power of 2 | |
2685 | */ | |
2686 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d | 2687 | ac->ac_2order = array_index_nospec(i - 1, |
4b68f6df | 2688 | MB_NUM_ORDERS(sb)); |
c9de560d AT |
2689 | } |
2690 | ||
4ba74d00 TT |
2691 | /* if stream allocation is enabled, use global goal */ |
2692 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2693 | /* TBD: may be hot point */ |
2694 | spin_lock(&sbi->s_md_lock); | |
2695 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2696 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2697 | spin_unlock(&sbi->s_md_lock); | |
2698 | } | |
4ba74d00 | 2699 | |
c9de560d AT |
2700 | /* Let's just scan groups to find more-less suitable blocks */ |
2701 | cr = ac->ac_2order ? 0 : 1; | |
2702 | /* | |
2703 | * cr == 0 try to get exact allocation, | |
2704 | * cr == 3 try to get anything | |
2705 | */ | |
2706 | repeat: | |
2707 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2708 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2709 | /* |
2710 | * searching for the right group start | |
2711 | * from the goal value specified | |
2712 | */ | |
2713 | group = ac->ac_g_ex.fe_group; | |
196e402a HS |
2714 | ac->ac_last_optimal_group = group; |
2715 | ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; | |
cfd73237 | 2716 | prefetch_grp = group; |
ed8f9c75 | 2717 | |
4fca50d4 JK |
2718 | for (i = 0, new_cr = cr; i < ngroups; i++, |
2719 | ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { | |
2720 | int ret = 0; | |
196e402a | 2721 | |
2ed5724d | 2722 | cond_resched(); |
196e402a HS |
2723 | if (new_cr != cr) { |
2724 | cr = new_cr; | |
2725 | goto repeat; | |
2726 | } | |
c9de560d | 2727 | |
cfd73237 AZ |
2728 | /* |
2729 | * Batch reads of the block allocation bitmaps | |
2730 | * to get multiple READs in flight; limit | |
2731 | * prefetching at cr=0/1, otherwise mballoc can | |
2732 | * spend a lot of time loading imperfect groups | |
2733 | */ | |
2734 | if ((prefetch_grp == group) && | |
2735 | (cr > 1 || | |
2736 | prefetch_ios < sbi->s_mb_prefetch_limit)) { | |
2737 | unsigned int curr_ios = prefetch_ios; | |
2738 | ||
2739 | nr = sbi->s_mb_prefetch; | |
2740 | if (ext4_has_feature_flex_bg(sb)) { | |
82ef1370 CX |
2741 | nr = 1 << sbi->s_log_groups_per_flex; |
2742 | nr -= group & (nr - 1); | |
2743 | nr = min(nr, sbi->s_mb_prefetch); | |
cfd73237 AZ |
2744 | } |
2745 | prefetch_grp = ext4_mb_prefetch(sb, group, | |
2746 | nr, &prefetch_ios); | |
2747 | if (prefetch_ios == curr_ios) | |
2748 | nr = 0; | |
2749 | } | |
2750 | ||
8a57d9d6 | 2751 | /* This now checks without needing the buddy page */ |
8ef123fe | 2752 | ret = ext4_mb_good_group_nolock(ac, group, cr); |
42ac1848 LC |
2753 | if (ret <= 0) { |
2754 | if (!first_err) | |
2755 | first_err = ret; | |
c9de560d | 2756 | continue; |
42ac1848 | 2757 | } |
c9de560d | 2758 | |
c9de560d AT |
2759 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2760 | if (err) | |
2761 | goto out; | |
2762 | ||
2763 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2764 | |
2765 | /* | |
2766 | * We need to check again after locking the | |
2767 | * block group | |
2768 | */ | |
42ac1848 | 2769 | ret = ext4_mb_good_group(ac, group, cr); |
8ef123fe | 2770 | if (ret == 0) { |
c9de560d | 2771 | ext4_unlock_group(sb, group); |
e39e07fd | 2772 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2773 | continue; |
2774 | } | |
2775 | ||
2776 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2777 | if (cr == 0) |
c9de560d | 2778 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2779 | else if (cr == 1 && sbi->s_stripe && |
2780 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2781 | ext4_mb_scan_aligned(ac, &e4b); |
2782 | else | |
2783 | ext4_mb_complex_scan_group(ac, &e4b); | |
2784 | ||
2785 | ext4_unlock_group(sb, group); | |
e39e07fd | 2786 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2787 | |
2788 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2789 | break; | |
2790 | } | |
a6c75eaf HS |
2791 | /* Processed all groups and haven't found blocks */ |
2792 | if (sbi->s_mb_stats && i == ngroups) | |
2793 | atomic64_inc(&sbi->s_bal_cX_failed[cr]); | |
c9de560d AT |
2794 | } |
2795 | ||
2796 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2797 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2798 | /* | |
2799 | * We've been searching too long. Let's try to allocate | |
2800 | * the best chunk we've found so far | |
2801 | */ | |
c9de560d AT |
2802 | ext4_mb_try_best_found(ac, &e4b); |
2803 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2804 | /* | |
2805 | * Someone more lucky has already allocated it. | |
2806 | * The only thing we can do is just take first | |
2807 | * found block(s) | |
c9de560d | 2808 | */ |
66d5e027 | 2809 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
2810 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", | |
c55ee7d2 | 2811 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
2812 | ac->ac_b_ex.fe_len, lost); | |
2813 | ||
c9de560d AT |
2814 | ac->ac_b_ex.fe_group = 0; |
2815 | ac->ac_b_ex.fe_start = 0; | |
2816 | ac->ac_b_ex.fe_len = 0; | |
2817 | ac->ac_status = AC_STATUS_CONTINUE; | |
2818 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2819 | cr = 3; | |
c9de560d AT |
2820 | goto repeat; |
2821 | } | |
2822 | } | |
a6c75eaf HS |
2823 | |
2824 | if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) | |
2825 | atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); | |
c9de560d | 2826 | out: |
42ac1848 LC |
2827 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2828 | err = first_err; | |
bbc4ec77 | 2829 | |
d3df1453 | 2830 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
bbc4ec77 RH |
2831 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
2832 | ac->ac_flags, cr, err); | |
cfd73237 AZ |
2833 | |
2834 | if (nr) | |
2835 | ext4_mb_prefetch_fini(sb, prefetch_grp, nr); | |
2836 | ||
c9de560d AT |
2837 | return err; |
2838 | } | |
2839 | ||
c9de560d AT |
2840 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2841 | { | |
359745d7 | 2842 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
2843 | ext4_group_t group; |
2844 | ||
8df9675f | 2845 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2846 | return NULL; |
c9de560d | 2847 | group = *pos + 1; |
a9df9a49 | 2848 | return (void *) ((unsigned long) group); |
c9de560d AT |
2849 | } |
2850 | ||
2851 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2852 | { | |
359745d7 | 2853 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
2854 | ext4_group_t group; |
2855 | ||
2856 | ++*pos; | |
8df9675f | 2857 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2858 | return NULL; |
2859 | group = *pos + 1; | |
a9df9a49 | 2860 | return (void *) ((unsigned long) group); |
c9de560d AT |
2861 | } |
2862 | ||
2863 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2864 | { | |
359745d7 | 2865 | struct super_block *sb = pde_data(file_inode(seq->file)); |
a9df9a49 | 2866 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2867 | int i; |
1c8457ca | 2868 | int err, buddy_loaded = 0; |
c9de560d | 2869 | struct ext4_buddy e4b; |
1c8457ca | 2870 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2871 | unsigned char blocksize_bits = min_t(unsigned char, |
2872 | sb->s_blocksize_bits, | |
2873 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2874 | struct sg { |
2875 | struct ext4_group_info info; | |
b80b32b6 | 2876 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2877 | } sg; |
2878 | ||
2879 | group--; | |
2880 | if (group == 0) | |
97b4af2f RV |
2881 | seq_puts(seq, "#group: free frags first [" |
2882 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2883 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2884 | |
b80b32b6 TT |
2885 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2886 | sizeof(struct ext4_group_info); | |
2887 | ||
1c8457ca AK |
2888 | grinfo = ext4_get_group_info(sb, group); |
2889 | /* Load the group info in memory only if not already loaded. */ | |
2890 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2891 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2892 | if (err) { | |
2893 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2894 | return 0; | |
2895 | } | |
2896 | buddy_loaded = 1; | |
c9de560d | 2897 | } |
1c8457ca | 2898 | |
b80b32b6 | 2899 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2900 | |
2901 | if (buddy_loaded) | |
2902 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2903 | |
a9df9a49 | 2904 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2905 | sg.info.bb_fragments, sg.info.bb_first_free); |
2906 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2907 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d | 2908 | sg.info.bb_counters[i] : 0); |
e0d438c7 | 2909 | seq_puts(seq, " ]\n"); |
c9de560d AT |
2910 | |
2911 | return 0; | |
2912 | } | |
2913 | ||
2914 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2915 | { | |
2916 | } | |
2917 | ||
247dbed8 | 2918 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2919 | .start = ext4_mb_seq_groups_start, |
2920 | .next = ext4_mb_seq_groups_next, | |
2921 | .stop = ext4_mb_seq_groups_stop, | |
2922 | .show = ext4_mb_seq_groups_show, | |
2923 | }; | |
2924 | ||
a6c75eaf HS |
2925 | int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) |
2926 | { | |
c30365b9 | 2927 | struct super_block *sb = seq->private; |
a6c75eaf HS |
2928 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
2929 | ||
2930 | seq_puts(seq, "mballoc:\n"); | |
2931 | if (!sbi->s_mb_stats) { | |
2932 | seq_puts(seq, "\tmb stats collection turned off.\n"); | |
2933 | seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); | |
2934 | return 0; | |
2935 | } | |
2936 | seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); | |
2937 | seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); | |
2938 | ||
2939 | seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned)); | |
2940 | ||
2941 | seq_puts(seq, "\tcr0_stats:\n"); | |
2942 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0])); | |
2943 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2944 | atomic64_read(&sbi->s_bal_cX_groups_considered[0])); | |
2945 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2946 | atomic64_read(&sbi->s_bal_cX_failed[0])); | |
196e402a HS |
2947 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2948 | atomic_read(&sbi->s_bal_cr0_bad_suggestions)); | |
a6c75eaf HS |
2949 | |
2950 | seq_puts(seq, "\tcr1_stats:\n"); | |
2951 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1])); | |
2952 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2953 | atomic64_read(&sbi->s_bal_cX_groups_considered[1])); | |
2954 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2955 | atomic64_read(&sbi->s_bal_cX_failed[1])); | |
196e402a HS |
2956 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2957 | atomic_read(&sbi->s_bal_cr1_bad_suggestions)); | |
a6c75eaf HS |
2958 | |
2959 | seq_puts(seq, "\tcr2_stats:\n"); | |
2960 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2])); | |
2961 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2962 | atomic64_read(&sbi->s_bal_cX_groups_considered[2])); | |
2963 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2964 | atomic64_read(&sbi->s_bal_cX_failed[2])); | |
2965 | ||
2966 | seq_puts(seq, "\tcr3_stats:\n"); | |
2967 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3])); | |
2968 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2969 | atomic64_read(&sbi->s_bal_cX_groups_considered[3])); | |
2970 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2971 | atomic64_read(&sbi->s_bal_cX_failed[3])); | |
2972 | seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned)); | |
2973 | seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); | |
2974 | seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); | |
2975 | seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); | |
2976 | seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); | |
2977 | ||
2978 | seq_printf(seq, "\tbuddies_generated: %u/%u\n", | |
2979 | atomic_read(&sbi->s_mb_buddies_generated), | |
2980 | ext4_get_groups_count(sb)); | |
2981 | seq_printf(seq, "\tbuddies_time_used: %llu\n", | |
2982 | atomic64_read(&sbi->s_mb_generation_time)); | |
2983 | seq_printf(seq, "\tpreallocated: %u\n", | |
2984 | atomic_read(&sbi->s_mb_preallocated)); | |
2985 | seq_printf(seq, "\tdiscarded: %u\n", | |
2986 | atomic_read(&sbi->s_mb_discarded)); | |
2987 | return 0; | |
2988 | } | |
2989 | ||
f68f4063 | 2990 | static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) |
a5fda113 | 2991 | __acquires(&EXT4_SB(sb)->s_mb_rb_lock) |
f68f4063 | 2992 | { |
359745d7 | 2993 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
2994 | unsigned long position; |
2995 | ||
2996 | read_lock(&EXT4_SB(sb)->s_mb_rb_lock); | |
2997 | ||
2998 | if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1) | |
2999 | return NULL; | |
3000 | position = *pos + 1; | |
3001 | return (void *) ((unsigned long) position); | |
3002 | } | |
3003 | ||
3004 | static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) | |
3005 | { | |
359745d7 | 3006 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
3007 | unsigned long position; |
3008 | ||
3009 | ++*pos; | |
3010 | if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1) | |
3011 | return NULL; | |
3012 | position = *pos + 1; | |
3013 | return (void *) ((unsigned long) position); | |
3014 | } | |
3015 | ||
3016 | static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) | |
3017 | { | |
359745d7 | 3018 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
3019 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3020 | unsigned long position = ((unsigned long) v); | |
3021 | struct ext4_group_info *grp; | |
3022 | struct rb_node *n; | |
3023 | unsigned int count, min, max; | |
3024 | ||
3025 | position--; | |
3026 | if (position >= MB_NUM_ORDERS(sb)) { | |
3027 | seq_puts(seq, "fragment_size_tree:\n"); | |
3028 | n = rb_first(&sbi->s_mb_avg_fragment_size_root); | |
3029 | if (!n) { | |
3030 | seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n"); | |
3031 | return 0; | |
3032 | } | |
3033 | grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb); | |
3034 | min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0; | |
3035 | count = 1; | |
3036 | while (rb_next(n)) { | |
3037 | count++; | |
3038 | n = rb_next(n); | |
3039 | } | |
3040 | grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb); | |
3041 | max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0; | |
3042 | ||
3043 | seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n", | |
3044 | min, max, count); | |
3045 | return 0; | |
3046 | } | |
3047 | ||
3048 | if (position == 0) { | |
3049 | seq_printf(seq, "optimize_scan: %d\n", | |
3050 | test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); | |
3051 | seq_puts(seq, "max_free_order_lists:\n"); | |
3052 | } | |
3053 | count = 0; | |
3054 | list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], | |
3055 | bb_largest_free_order_node) | |
3056 | count++; | |
3057 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", | |
3058 | (unsigned int)position, count); | |
3059 | ||
3060 | return 0; | |
3061 | } | |
3062 | ||
3063 | static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) | |
a5fda113 | 3064 | __releases(&EXT4_SB(sb)->s_mb_rb_lock) |
f68f4063 | 3065 | { |
359745d7 | 3066 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
3067 | |
3068 | read_unlock(&EXT4_SB(sb)->s_mb_rb_lock); | |
3069 | } | |
3070 | ||
3071 | const struct seq_operations ext4_mb_seq_structs_summary_ops = { | |
3072 | .start = ext4_mb_seq_structs_summary_start, | |
3073 | .next = ext4_mb_seq_structs_summary_next, | |
3074 | .stop = ext4_mb_seq_structs_summary_stop, | |
3075 | .show = ext4_mb_seq_structs_summary_show, | |
3076 | }; | |
3077 | ||
fb1813f4 CW |
3078 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
3079 | { | |
3080 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3081 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
3082 | ||
3083 | BUG_ON(!cachep); | |
3084 | return cachep; | |
3085 | } | |
5f21b0e6 | 3086 | |
28623c2f TT |
3087 | /* |
3088 | * Allocate the top-level s_group_info array for the specified number | |
3089 | * of groups | |
3090 | */ | |
3091 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
3092 | { | |
3093 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3094 | unsigned size; | |
df3da4ea | 3095 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
3096 | |
3097 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
3098 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
3099 | if (size <= sbi->s_group_info_size) | |
3100 | return 0; | |
3101 | ||
3102 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 3103 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
3104 | if (!new_groupinfo) { |
3105 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
3106 | return -ENOMEM; | |
3107 | } | |
df3da4ea SJS |
3108 | rcu_read_lock(); |
3109 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
3110 | if (old_groupinfo) | |
3111 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 3112 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
3113 | rcu_read_unlock(); |
3114 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 3115 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
3116 | if (old_groupinfo) |
3117 | ext4_kvfree_array_rcu(old_groupinfo); | |
666245d9 | 3118 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
28623c2f TT |
3119 | sbi->s_group_info_size); |
3120 | return 0; | |
3121 | } | |
3122 | ||
5f21b0e6 | 3123 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 3124 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
3125 | struct ext4_group_desc *desc) |
3126 | { | |
fb1813f4 | 3127 | int i; |
5f21b0e6 | 3128 | int metalen = 0; |
df3da4ea | 3129 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
3130 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3131 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 3132 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
3133 | |
3134 | /* | |
3135 | * First check if this group is the first of a reserved block. | |
3136 | * If it's true, we have to allocate a new table of pointers | |
3137 | * to ext4_group_info structures | |
3138 | */ | |
3139 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
3140 | metalen = sizeof(*meta_group_info) << | |
3141 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 3142 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 3143 | if (meta_group_info == NULL) { |
7f6a11e7 | 3144 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 3145 | "for a buddy group"); |
5f21b0e6 FB |
3146 | goto exit_meta_group_info; |
3147 | } | |
df3da4ea SJS |
3148 | rcu_read_lock(); |
3149 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
3150 | rcu_read_unlock(); | |
5f21b0e6 FB |
3151 | } |
3152 | ||
df3da4ea | 3153 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
3154 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
3155 | ||
4fdb5543 | 3156 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 3157 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 3158 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
3159 | goto exit_group_info; |
3160 | } | |
3161 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
3162 | &(meta_group_info[i]->bb_state)); | |
3163 | ||
3164 | /* | |
3165 | * initialize bb_free to be able to skip | |
3166 | * empty groups without initialization | |
3167 | */ | |
8844618d TT |
3168 | if (ext4_has_group_desc_csum(sb) && |
3169 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 3170 | meta_group_info[i]->bb_free = |
cff1dfd7 | 3171 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
3172 | } else { |
3173 | meta_group_info[i]->bb_free = | |
021b65bb | 3174 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
3175 | } |
3176 | ||
3177 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 3178 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 3179 | meta_group_info[i]->bb_free_root = RB_ROOT; |
196e402a HS |
3180 | INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); |
3181 | RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb); | |
8a57d9d6 | 3182 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
196e402a | 3183 | meta_group_info[i]->bb_group = group; |
5f21b0e6 | 3184 | |
a3450215 | 3185 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
3186 | return 0; |
3187 | ||
3188 | exit_group_info: | |
3189 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 3190 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
3191 | struct ext4_group_info ***group_info; |
3192 | ||
3193 | rcu_read_lock(); | |
3194 | group_info = rcu_dereference(sbi->s_group_info); | |
3195 | kfree(group_info[idx]); | |
3196 | group_info[idx] = NULL; | |
3197 | rcu_read_unlock(); | |
caaf7a29 | 3198 | } |
5f21b0e6 FB |
3199 | exit_meta_group_info: |
3200 | return -ENOMEM; | |
3201 | } /* ext4_mb_add_groupinfo */ | |
3202 | ||
c9de560d AT |
3203 | static int ext4_mb_init_backend(struct super_block *sb) |
3204 | { | |
8df9675f | 3205 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 3206 | ext4_group_t i; |
c9de560d | 3207 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 3208 | int err; |
5f21b0e6 | 3209 | struct ext4_group_desc *desc; |
df3da4ea | 3210 | struct ext4_group_info ***group_info; |
fb1813f4 | 3211 | struct kmem_cache *cachep; |
5f21b0e6 | 3212 | |
28623c2f TT |
3213 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
3214 | if (err) | |
3215 | return err; | |
c9de560d | 3216 | |
c9de560d AT |
3217 | sbi->s_buddy_cache = new_inode(sb); |
3218 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 3219 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
3220 | goto err_freesgi; |
3221 | } | |
48e6061b YJ |
3222 | /* To avoid potentially colliding with an valid on-disk inode number, |
3223 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
3224 | * not in the inode hash, so it should never be found by iget(), but | |
3225 | * this will avoid confusion if it ever shows up during debugging. */ | |
3226 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 3227 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 3228 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3229 | cond_resched(); |
c9de560d AT |
3230 | desc = ext4_get_group_desc(sb, i, NULL); |
3231 | if (desc == NULL) { | |
9d8b9ec4 | 3232 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
3233 | goto err_freebuddy; |
3234 | } | |
5f21b0e6 FB |
3235 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
3236 | goto err_freebuddy; | |
c9de560d AT |
3237 | } |
3238 | ||
cfd73237 | 3239 | if (ext4_has_feature_flex_bg(sb)) { |
f91436d5 ST |
3240 | /* a single flex group is supposed to be read by a single IO. |
3241 | * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is | |
3242 | * unsigned integer, so the maximum shift is 32. | |
3243 | */ | |
3244 | if (sbi->s_es->s_log_groups_per_flex >= 32) { | |
3245 | ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); | |
a8867f4e | 3246 | goto err_freebuddy; |
f91436d5 ST |
3247 | } |
3248 | sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, | |
82ef1370 | 3249 | BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); |
cfd73237 AZ |
3250 | sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ |
3251 | } else { | |
3252 | sbi->s_mb_prefetch = 32; | |
3253 | } | |
3254 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) | |
3255 | sbi->s_mb_prefetch = ext4_get_groups_count(sb); | |
3256 | /* now many real IOs to prefetch within a single allocation at cr=0 | |
3257 | * given cr=0 is an CPU-related optimization we shouldn't try to | |
3258 | * load too many groups, at some point we should start to use what | |
3259 | * we've got in memory. | |
3260 | * with an average random access time 5ms, it'd take a second to get | |
3261 | * 200 groups (* N with flex_bg), so let's make this limit 4 | |
3262 | */ | |
3263 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; | |
3264 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) | |
3265 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); | |
3266 | ||
c9de560d AT |
3267 | return 0; |
3268 | ||
3269 | err_freebuddy: | |
fb1813f4 | 3270 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 3271 | while (i-- > 0) |
fb1813f4 | 3272 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 3273 | i = sbi->s_group_info_size; |
df3da4ea SJS |
3274 | rcu_read_lock(); |
3275 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 3276 | while (i-- > 0) |
df3da4ea SJS |
3277 | kfree(group_info[i]); |
3278 | rcu_read_unlock(); | |
c9de560d AT |
3279 | iput(sbi->s_buddy_cache); |
3280 | err_freesgi: | |
df3da4ea SJS |
3281 | rcu_read_lock(); |
3282 | kvfree(rcu_dereference(sbi->s_group_info)); | |
3283 | rcu_read_unlock(); | |
c9de560d AT |
3284 | return -ENOMEM; |
3285 | } | |
3286 | ||
2892c15d ES |
3287 | static void ext4_groupinfo_destroy_slabs(void) |
3288 | { | |
3289 | int i; | |
3290 | ||
3291 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 3292 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
3293 | ext4_groupinfo_caches[i] = NULL; |
3294 | } | |
3295 | } | |
3296 | ||
3297 | static int ext4_groupinfo_create_slab(size_t size) | |
3298 | { | |
3299 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
3300 | int slab_size; | |
3301 | int blocksize_bits = order_base_2(size); | |
3302 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3303 | struct kmem_cache *cachep; | |
3304 | ||
3305 | if (cache_index >= NR_GRPINFO_CACHES) | |
3306 | return -EINVAL; | |
3307 | ||
3308 | if (unlikely(cache_index < 0)) | |
3309 | cache_index = 0; | |
3310 | ||
3311 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
3312 | if (ext4_groupinfo_caches[cache_index]) { | |
3313 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
3314 | return 0; /* Already created */ | |
3315 | } | |
3316 | ||
3317 | slab_size = offsetof(struct ext4_group_info, | |
3318 | bb_counters[blocksize_bits + 2]); | |
3319 | ||
3320 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
3321 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
3322 | NULL); | |
3323 | ||
823ba01f TM |
3324 | ext4_groupinfo_caches[cache_index] = cachep; |
3325 | ||
2892c15d ES |
3326 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
3327 | if (!cachep) { | |
9d8b9ec4 TT |
3328 | printk(KERN_EMERG |
3329 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
3330 | return -ENOMEM; |
3331 | } | |
3332 | ||
2892c15d ES |
3333 | return 0; |
3334 | } | |
3335 | ||
55cdd0af WJ |
3336 | static void ext4_discard_work(struct work_struct *work) |
3337 | { | |
3338 | struct ext4_sb_info *sbi = container_of(work, | |
3339 | struct ext4_sb_info, s_discard_work); | |
3340 | struct super_block *sb = sbi->s_sb; | |
3341 | struct ext4_free_data *fd, *nfd; | |
3342 | struct ext4_buddy e4b; | |
3343 | struct list_head discard_list; | |
3344 | ext4_group_t grp, load_grp; | |
3345 | int err = 0; | |
3346 | ||
3347 | INIT_LIST_HEAD(&discard_list); | |
3348 | spin_lock(&sbi->s_md_lock); | |
3349 | list_splice_init(&sbi->s_discard_list, &discard_list); | |
3350 | spin_unlock(&sbi->s_md_lock); | |
3351 | ||
3352 | load_grp = UINT_MAX; | |
3353 | list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { | |
3354 | /* | |
5036ab8d WJ |
3355 | * If filesystem is umounting or no memory or suffering |
3356 | * from no space, give up the discard | |
55cdd0af | 3357 | */ |
5036ab8d WJ |
3358 | if ((sb->s_flags & SB_ACTIVE) && !err && |
3359 | !atomic_read(&sbi->s_retry_alloc_pending)) { | |
55cdd0af WJ |
3360 | grp = fd->efd_group; |
3361 | if (grp != load_grp) { | |
3362 | if (load_grp != UINT_MAX) | |
3363 | ext4_mb_unload_buddy(&e4b); | |
3364 | ||
3365 | err = ext4_mb_load_buddy(sb, grp, &e4b); | |
3366 | if (err) { | |
3367 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3368 | load_grp = UINT_MAX; | |
3369 | continue; | |
3370 | } else { | |
3371 | load_grp = grp; | |
3372 | } | |
3373 | } | |
3374 | ||
3375 | ext4_lock_group(sb, grp); | |
3376 | ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, | |
3377 | fd->efd_start_cluster + fd->efd_count - 1, 1); | |
3378 | ext4_unlock_group(sb, grp); | |
3379 | } | |
3380 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3381 | } | |
3382 | ||
3383 | if (load_grp != UINT_MAX) | |
3384 | ext4_mb_unload_buddy(&e4b); | |
3385 | } | |
3386 | ||
9d99012f | 3387 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
3388 | { |
3389 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 3390 | unsigned i, j; |
935244cd | 3391 | unsigned offset, offset_incr; |
c9de560d | 3392 | unsigned max; |
74767c5a | 3393 | int ret; |
c9de560d | 3394 | |
4b68f6df | 3395 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
3396 | |
3397 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
3398 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
3399 | ret = -ENOMEM; |
3400 | goto out; | |
c9de560d | 3401 | } |
ff7ef329 | 3402 | |
4b68f6df | 3403 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
3404 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
3405 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
3406 | ret = -ENOMEM; |
3407 | goto out; | |
3408 | } | |
3409 | ||
2892c15d ES |
3410 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
3411 | if (ret < 0) | |
3412 | goto out; | |
c9de560d AT |
3413 | |
3414 | /* order 0 is regular bitmap */ | |
3415 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
3416 | sbi->s_mb_offsets[0] = 0; | |
3417 | ||
3418 | i = 1; | |
3419 | offset = 0; | |
935244cd | 3420 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
3421 | max = sb->s_blocksize << 2; |
3422 | do { | |
3423 | sbi->s_mb_offsets[i] = offset; | |
3424 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
3425 | offset += offset_incr; |
3426 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
3427 | max = max >> 1; |
3428 | i++; | |
4b68f6df HS |
3429 | } while (i < MB_NUM_ORDERS(sb)); |
3430 | ||
196e402a HS |
3431 | sbi->s_mb_avg_fragment_size_root = RB_ROOT; |
3432 | sbi->s_mb_largest_free_orders = | |
3433 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3434 | GFP_KERNEL); | |
3435 | if (!sbi->s_mb_largest_free_orders) { | |
3436 | ret = -ENOMEM; | |
3437 | goto out; | |
3438 | } | |
3439 | sbi->s_mb_largest_free_orders_locks = | |
3440 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3441 | GFP_KERNEL); | |
3442 | if (!sbi->s_mb_largest_free_orders_locks) { | |
3443 | ret = -ENOMEM; | |
3444 | goto out; | |
3445 | } | |
3446 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3447 | INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); | |
3448 | rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); | |
3449 | } | |
3450 | rwlock_init(&sbi->s_mb_rb_lock); | |
c9de560d | 3451 | |
c9de560d | 3452 | spin_lock_init(&sbi->s_md_lock); |
d08854f5 | 3453 | sbi->s_mb_free_pending = 0; |
a0154344 | 3454 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
55cdd0af WJ |
3455 | INIT_LIST_HEAD(&sbi->s_discard_list); |
3456 | INIT_WORK(&sbi->s_discard_work, ext4_discard_work); | |
5036ab8d | 3457 | atomic_set(&sbi->s_retry_alloc_pending, 0); |
c9de560d AT |
3458 | |
3459 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
3460 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
3461 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
3462 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
3463 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27bc446e | 3464 | sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC; |
27baebb8 TT |
3465 | /* |
3466 | * The default group preallocation is 512, which for 4k block | |
3467 | * sizes translates to 2 megabytes. However for bigalloc file | |
3468 | * systems, this is probably too big (i.e, if the cluster size | |
3469 | * is 1 megabyte, then group preallocation size becomes half a | |
3470 | * gigabyte!). As a default, we will keep a two megabyte | |
3471 | * group pralloc size for cluster sizes up to 64k, and after | |
3472 | * that, we will force a minimum group preallocation size of | |
3473 | * 32 clusters. This translates to 8 megs when the cluster | |
3474 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
3475 | * which seems reasonable as a default. | |
3476 | */ | |
3477 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
3478 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
3479 | /* |
3480 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
3481 | * to the lowest multiple of s_stripe which is bigger than | |
3482 | * the s_mb_group_prealloc as determined above. We want | |
3483 | * the preallocation size to be an exact multiple of the | |
3484 | * RAID stripe size so that preallocations don't fragment | |
3485 | * the stripes. | |
3486 | */ | |
3487 | if (sbi->s_stripe > 1) { | |
3488 | sbi->s_mb_group_prealloc = roundup( | |
3489 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
3490 | } | |
c9de560d | 3491 | |
730c213c | 3492 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 3493 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 3494 | ret = -ENOMEM; |
029b10c5 | 3495 | goto out; |
c9de560d | 3496 | } |
730c213c | 3497 | for_each_possible_cpu(i) { |
c9de560d | 3498 | struct ext4_locality_group *lg; |
730c213c | 3499 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 3500 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
3501 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
3502 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
3503 | spin_lock_init(&lg->lg_prealloc_lock); |
3504 | } | |
3505 | ||
10f0d2a5 | 3506 | if (bdev_nonrot(sb->s_bdev)) |
196e402a HS |
3507 | sbi->s_mb_max_linear_groups = 0; |
3508 | else | |
3509 | sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; | |
79a77c5a YJ |
3510 | /* init file for buddy data */ |
3511 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
3512 | if (ret != 0) |
3513 | goto out_free_locality_groups; | |
79a77c5a | 3514 | |
7aa0baea TM |
3515 | return 0; |
3516 | ||
3517 | out_free_locality_groups: | |
3518 | free_percpu(sbi->s_locality_groups); | |
3519 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 3520 | out: |
196e402a HS |
3521 | kfree(sbi->s_mb_largest_free_orders); |
3522 | kfree(sbi->s_mb_largest_free_orders_locks); | |
7aa0baea TM |
3523 | kfree(sbi->s_mb_offsets); |
3524 | sbi->s_mb_offsets = NULL; | |
3525 | kfree(sbi->s_mb_maxs); | |
3526 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 3527 | return ret; |
c9de560d AT |
3528 | } |
3529 | ||
955ce5f5 | 3530 | /* need to called with the ext4 group lock held */ |
d3df1453 | 3531 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
c9de560d AT |
3532 | { |
3533 | struct ext4_prealloc_space *pa; | |
3534 | struct list_head *cur, *tmp; | |
3535 | int count = 0; | |
3536 | ||
3537 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
3538 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3539 | list_del(&pa->pa_group_list); | |
3540 | count++; | |
688f05a0 | 3541 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d | 3542 | } |
d3df1453 | 3543 | return count; |
c9de560d AT |
3544 | } |
3545 | ||
3546 | int ext4_mb_release(struct super_block *sb) | |
3547 | { | |
8df9675f | 3548 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
3549 | ext4_group_t i; |
3550 | int num_meta_group_infos; | |
df3da4ea | 3551 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 3552 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 3553 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
d3df1453 | 3554 | int count; |
c9de560d | 3555 | |
55cdd0af WJ |
3556 | if (test_opt(sb, DISCARD)) { |
3557 | /* | |
3558 | * wait the discard work to drain all of ext4_free_data | |
3559 | */ | |
3560 | flush_work(&sbi->s_discard_work); | |
3561 | WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); | |
3562 | } | |
3563 | ||
c9de560d | 3564 | if (sbi->s_group_info) { |
8df9675f | 3565 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3566 | cond_resched(); |
c9de560d | 3567 | grinfo = ext4_get_group_info(sb, i); |
a3450215 | 3568 | mb_group_bb_bitmap_free(grinfo); |
c9de560d | 3569 | ext4_lock_group(sb, i); |
d3df1453 RH |
3570 | count = ext4_mb_cleanup_pa(grinfo); |
3571 | if (count) | |
3572 | mb_debug(sb, "mballoc: %d PAs left\n", | |
3573 | count); | |
c9de560d | 3574 | ext4_unlock_group(sb, i); |
fb1813f4 | 3575 | kmem_cache_free(cachep, grinfo); |
c9de560d | 3576 | } |
8df9675f | 3577 | num_meta_group_infos = (ngroups + |
c9de560d AT |
3578 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
3579 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
3580 | rcu_read_lock(); |
3581 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 3582 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
3583 | kfree(group_info[i]); |
3584 | kvfree(group_info); | |
3585 | rcu_read_unlock(); | |
c9de560d | 3586 | } |
196e402a HS |
3587 | kfree(sbi->s_mb_largest_free_orders); |
3588 | kfree(sbi->s_mb_largest_free_orders_locks); | |
c9de560d AT |
3589 | kfree(sbi->s_mb_offsets); |
3590 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 3591 | iput(sbi->s_buddy_cache); |
c9de560d | 3592 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
3593 | ext4_msg(sb, KERN_INFO, |
3594 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
3595 | atomic_read(&sbi->s_bal_allocated), |
3596 | atomic_read(&sbi->s_bal_reqs), | |
3597 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 | 3598 | ext4_msg(sb, KERN_INFO, |
a6c75eaf | 3599 | "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " |
9d8b9ec4 | 3600 | "%u 2^N hits, %u breaks, %u lost", |
c9de560d | 3601 | atomic_read(&sbi->s_bal_ex_scanned), |
a6c75eaf | 3602 | atomic_read(&sbi->s_bal_groups_scanned), |
c9de560d AT |
3603 | atomic_read(&sbi->s_bal_goals), |
3604 | atomic_read(&sbi->s_bal_2orders), | |
3605 | atomic_read(&sbi->s_bal_breaks), | |
3606 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 | 3607 | ext4_msg(sb, KERN_INFO, |
67d25186 HS |
3608 | "mballoc: %u generated and it took %llu", |
3609 | atomic_read(&sbi->s_mb_buddies_generated), | |
3610 | atomic64_read(&sbi->s_mb_generation_time)); | |
9d8b9ec4 TT |
3611 | ext4_msg(sb, KERN_INFO, |
3612 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
3613 | atomic_read(&sbi->s_mb_preallocated), |
3614 | atomic_read(&sbi->s_mb_discarded)); | |
3615 | } | |
3616 | ||
730c213c | 3617 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
3618 | |
3619 | return 0; | |
3620 | } | |
3621 | ||
77ca6cdf | 3622 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
3623 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
3624 | struct bio **biop) | |
5c521830 | 3625 | { |
5c521830 JZ |
3626 | ext4_fsblk_t discard_block; |
3627 | ||
84130193 TT |
3628 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
3629 | ext4_group_first_block_no(sb, block_group)); | |
3630 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
3631 | trace_ext4_discard_blocks(sb, |
3632 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
3633 | if (biop) { |
3634 | return __blkdev_issue_discard(sb->s_bdev, | |
3635 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
3636 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
44abff2c | 3637 | GFP_NOFS, biop); |
a0154344 DJ |
3638 | } else |
3639 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
3640 | } |
3641 | ||
a0154344 DJ |
3642 | static void ext4_free_data_in_buddy(struct super_block *sb, |
3643 | struct ext4_free_data *entry) | |
c9de560d | 3644 | { |
c9de560d | 3645 | struct ext4_buddy e4b; |
c894058d | 3646 | struct ext4_group_info *db; |
d9f34504 | 3647 | int err, count = 0, count2 = 0; |
c9de560d | 3648 | |
d3df1453 | 3649 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
18aadd47 | 3650 | entry->efd_count, entry->efd_group, entry); |
c9de560d | 3651 | |
18aadd47 BJ |
3652 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
3653 | /* we expect to find existing buddy because it's pinned */ | |
3654 | BUG_ON(err != 0); | |
b90f6870 | 3655 | |
d08854f5 TT |
3656 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
3657 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
3658 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 3659 | |
18aadd47 BJ |
3660 | db = e4b.bd_info; |
3661 | /* there are blocks to put in buddy to make them really free */ | |
3662 | count += entry->efd_count; | |
3663 | count2++; | |
3664 | ext4_lock_group(sb, entry->efd_group); | |
3665 | /* Take it out of per group rb tree */ | |
3666 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
3667 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 3668 | |
18aadd47 BJ |
3669 | /* |
3670 | * Clear the trimmed flag for the group so that the next | |
3671 | * ext4_trim_fs can trim it. | |
3672 | * If the volume is mounted with -o discard, online discard | |
3673 | * is supported and the free blocks will be trimmed online. | |
3674 | */ | |
3675 | if (!test_opt(sb, DISCARD)) | |
3676 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 3677 | |
18aadd47 BJ |
3678 | if (!db->bb_free_root.rb_node) { |
3679 | /* No more items in the per group rb tree | |
3680 | * balance refcounts from ext4_mb_free_metadata() | |
3681 | */ | |
09cbfeaf KS |
3682 | put_page(e4b.bd_buddy_page); |
3683 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 3684 | } |
18aadd47 | 3685 | ext4_unlock_group(sb, entry->efd_group); |
18aadd47 | 3686 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 3687 | |
d3df1453 RH |
3688 | mb_debug(sb, "freed %d blocks in %d structures\n", count, |
3689 | count2); | |
c9de560d AT |
3690 | } |
3691 | ||
a0154344 DJ |
3692 | /* |
3693 | * This function is called by the jbd2 layer once the commit has finished, | |
3694 | * so we know we can free the blocks that were released with that commit. | |
3695 | */ | |
3696 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
3697 | { | |
3698 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3699 | struct ext4_free_data *entry, *tmp; | |
a0154344 DJ |
3700 | struct list_head freed_data_list; |
3701 | struct list_head *cut_pos = NULL; | |
55cdd0af | 3702 | bool wake; |
a0154344 DJ |
3703 | |
3704 | INIT_LIST_HEAD(&freed_data_list); | |
3705 | ||
3706 | spin_lock(&sbi->s_md_lock); | |
3707 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
3708 | if (entry->efd_tid != commit_tid) | |
3709 | break; | |
3710 | cut_pos = &entry->efd_list; | |
3711 | } | |
3712 | if (cut_pos) | |
3713 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
3714 | cut_pos); | |
3715 | spin_unlock(&sbi->s_md_lock); | |
3716 | ||
55cdd0af WJ |
3717 | list_for_each_entry(entry, &freed_data_list, efd_list) |
3718 | ext4_free_data_in_buddy(sb, entry); | |
a0154344 | 3719 | |
55cdd0af WJ |
3720 | if (test_opt(sb, DISCARD)) { |
3721 | spin_lock(&sbi->s_md_lock); | |
3722 | wake = list_empty(&sbi->s_discard_list); | |
3723 | list_splice_tail(&freed_data_list, &sbi->s_discard_list); | |
3724 | spin_unlock(&sbi->s_md_lock); | |
3725 | if (wake) | |
3726 | queue_work(system_unbound_wq, &sbi->s_discard_work); | |
3727 | } else { | |
3728 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
3729 | kmem_cache_free(ext4_free_data_cachep, entry); | |
a0154344 | 3730 | } |
a0154344 DJ |
3731 | } |
3732 | ||
5dabfc78 | 3733 | int __init ext4_init_mballoc(void) |
c9de560d | 3734 | { |
16828088 TT |
3735 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
3736 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 3737 | if (ext4_pspace_cachep == NULL) |
f283529a | 3738 | goto out; |
c9de560d | 3739 | |
16828088 TT |
3740 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
3741 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3742 | if (ext4_ac_cachep == NULL) |
3743 | goto out_pa_free; | |
c894058d | 3744 | |
18aadd47 BJ |
3745 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
3746 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3747 | if (ext4_free_data_cachep == NULL) |
3748 | goto out_ac_free; | |
3749 | ||
c9de560d | 3750 | return 0; |
f283529a RH |
3751 | |
3752 | out_ac_free: | |
3753 | kmem_cache_destroy(ext4_ac_cachep); | |
3754 | out_pa_free: | |
3755 | kmem_cache_destroy(ext4_pspace_cachep); | |
3756 | out: | |
3757 | return -ENOMEM; | |
c9de560d AT |
3758 | } |
3759 | ||
5dabfc78 | 3760 | void ext4_exit_mballoc(void) |
c9de560d | 3761 | { |
60e6679e | 3762 | /* |
3e03f9ca JDB |
3763 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
3764 | * before destroying the slab cache. | |
3765 | */ | |
3766 | rcu_barrier(); | |
c9de560d | 3767 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 3768 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 3769 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 3770 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
3771 | } |
3772 | ||
3773 | ||
3774 | /* | |
73b2c716 | 3775 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
3776 | * Returns 0 if success or error code |
3777 | */ | |
4ddfef7b ES |
3778 | static noinline_for_stack int |
3779 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 3780 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
3781 | { |
3782 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
3783 | struct ext4_group_desc *gdp; |
3784 | struct buffer_head *gdp_bh; | |
3785 | struct ext4_sb_info *sbi; | |
3786 | struct super_block *sb; | |
3787 | ext4_fsblk_t block; | |
519deca0 | 3788 | int err, len; |
c9de560d AT |
3789 | |
3790 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3791 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
3792 | ||
3793 | sb = ac->ac_sb; | |
3794 | sbi = EXT4_SB(sb); | |
c9de560d | 3795 | |
574ca174 | 3796 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
3797 | if (IS_ERR(bitmap_bh)) { |
3798 | err = PTR_ERR(bitmap_bh); | |
3799 | bitmap_bh = NULL; | |
c9de560d | 3800 | goto out_err; |
9008a58e | 3801 | } |
c9de560d | 3802 | |
5d601255 | 3803 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
3804 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
3805 | EXT4_JTR_NONE); | |
c9de560d AT |
3806 | if (err) |
3807 | goto out_err; | |
3808 | ||
3809 | err = -EIO; | |
3810 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3811 | if (!gdp) | |
3812 | goto out_err; | |
3813 | ||
a9df9a49 | 3814 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3815 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3816 | |
5d601255 | 3817 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
188c299e | 3818 | err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE); |
c9de560d AT |
3819 | if (err) |
3820 | goto out_err; | |
3821 | ||
bda00de7 | 3822 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3823 | |
53accfa9 | 3824 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
ce9f24cc | 3825 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
12062ddd | 3826 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3827 | "fs metadata", block, block+len); |
519deca0 | 3828 | /* File system mounted not to panic on error |
554a5ccc | 3829 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3830 | * We leak some of the blocks here. |
3831 | */ | |
955ce5f5 | 3832 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
123e3016 | 3833 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
c3e94d1d | 3834 | ac->ac_b_ex.fe_len); |
955ce5f5 | 3835 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3836 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3837 | if (!err) |
554a5ccc | 3838 | err = -EFSCORRUPTED; |
519deca0 | 3839 | goto out_err; |
c9de560d | 3840 | } |
955ce5f5 AK |
3841 | |
3842 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3843 | #ifdef AGGRESSIVE_CHECK |
3844 | { | |
3845 | int i; | |
3846 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3847 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3848 | bitmap_bh->b_data)); | |
3849 | } | |
3850 | } | |
3851 | #endif | |
123e3016 | 3852 | mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
c3e94d1d | 3853 | ac->ac_b_ex.fe_len); |
8844618d TT |
3854 | if (ext4_has_group_desc_csum(sb) && |
3855 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3856 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3857 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3858 | ext4_free_clusters_after_init(sb, |
021b65bb | 3859 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3860 | } |
021b65bb TT |
3861 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3862 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3863 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3864 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3865 | |
3866 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3867 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3868 | /* |
6bc6e63f | 3869 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3870 | */ |
6bc6e63f AK |
3871 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3872 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3873 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3874 | reserv_clstrs); | |
c9de560d | 3875 | |
772cb7c8 JS |
3876 | if (sbi->s_log_groups_per_flex) { |
3877 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3878 | ac->ac_b_ex.fe_group); | |
90ba983f | 3879 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3880 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3881 | flex_group)->free_clusters); | |
772cb7c8 JS |
3882 | } |
3883 | ||
0390131b | 3884 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3885 | if (err) |
3886 | goto out_err; | |
0390131b | 3887 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3888 | |
3889 | out_err: | |
42a10add | 3890 | brelse(bitmap_bh); |
c9de560d AT |
3891 | return err; |
3892 | } | |
3893 | ||
8016e29f HS |
3894 | /* |
3895 | * Idempotent helper for Ext4 fast commit replay path to set the state of | |
3896 | * blocks in bitmaps and update counters. | |
3897 | */ | |
3898 | void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, | |
3899 | int len, int state) | |
3900 | { | |
3901 | struct buffer_head *bitmap_bh = NULL; | |
3902 | struct ext4_group_desc *gdp; | |
3903 | struct buffer_head *gdp_bh; | |
3904 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3905 | ext4_group_t group; | |
3906 | ext4_grpblk_t blkoff; | |
a5c0e2fd | 3907 | int i, err; |
8016e29f | 3908 | int already; |
bfdc502a | 3909 | unsigned int clen, clen_changed, thisgrp_len; |
8016e29f | 3910 | |
bfdc502a RH |
3911 | while (len > 0) { |
3912 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
8016e29f | 3913 | |
bfdc502a RH |
3914 | /* |
3915 | * Check to see if we are freeing blocks across a group | |
3916 | * boundary. | |
3917 | * In case of flex_bg, this can happen that (block, len) may | |
3918 | * span across more than one group. In that case we need to | |
3919 | * get the corresponding group metadata to work with. | |
3920 | * For this we have goto again loop. | |
3921 | */ | |
3922 | thisgrp_len = min_t(unsigned int, (unsigned int)len, | |
3923 | EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); | |
3924 | clen = EXT4_NUM_B2C(sbi, thisgrp_len); | |
3925 | ||
8c91c579 RH |
3926 | if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { |
3927 | ext4_error(sb, "Marking blocks in system zone - " | |
3928 | "Block = %llu, len = %u", | |
3929 | block, thisgrp_len); | |
3930 | bitmap_bh = NULL; | |
3931 | break; | |
3932 | } | |
3933 | ||
bfdc502a RH |
3934 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
3935 | if (IS_ERR(bitmap_bh)) { | |
3936 | err = PTR_ERR(bitmap_bh); | |
3937 | bitmap_bh = NULL; | |
3938 | break; | |
3939 | } | |
8016e29f | 3940 | |
bfdc502a RH |
3941 | err = -EIO; |
3942 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
3943 | if (!gdp) | |
3944 | break; | |
8016e29f | 3945 | |
bfdc502a RH |
3946 | ext4_lock_group(sb, group); |
3947 | already = 0; | |
3948 | for (i = 0; i < clen; i++) | |
3949 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == | |
3950 | !state) | |
3951 | already++; | |
8016e29f | 3952 | |
bfdc502a RH |
3953 | clen_changed = clen - already; |
3954 | if (state) | |
123e3016 | 3955 | mb_set_bits(bitmap_bh->b_data, blkoff, clen); |
bfdc502a | 3956 | else |
bd8247ee | 3957 | mb_clear_bits(bitmap_bh->b_data, blkoff, clen); |
bfdc502a RH |
3958 | if (ext4_has_group_desc_csum(sb) && |
3959 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
3960 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
3961 | ext4_free_group_clusters_set(sb, gdp, | |
3962 | ext4_free_clusters_after_init(sb, group, gdp)); | |
3963 | } | |
3964 | if (state) | |
3965 | clen = ext4_free_group_clusters(sb, gdp) - clen_changed; | |
3966 | else | |
3967 | clen = ext4_free_group_clusters(sb, gdp) + clen_changed; | |
8016e29f | 3968 | |
bfdc502a RH |
3969 | ext4_free_group_clusters_set(sb, gdp, clen); |
3970 | ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh); | |
3971 | ext4_group_desc_csum_set(sb, group, gdp); | |
8016e29f | 3972 | |
bfdc502a | 3973 | ext4_unlock_group(sb, group); |
8016e29f | 3974 | |
bfdc502a RH |
3975 | if (sbi->s_log_groups_per_flex) { |
3976 | ext4_group_t flex_group = ext4_flex_group(sbi, group); | |
3977 | struct flex_groups *fg = sbi_array_rcu_deref(sbi, | |
3978 | s_flex_groups, flex_group); | |
8016e29f | 3979 | |
bfdc502a RH |
3980 | if (state) |
3981 | atomic64_sub(clen_changed, &fg->free_clusters); | |
3982 | else | |
3983 | atomic64_add(clen_changed, &fg->free_clusters); | |
3984 | ||
3985 | } | |
3986 | ||
3987 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
3988 | if (err) | |
3989 | break; | |
3990 | sync_dirty_buffer(bitmap_bh); | |
3991 | err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
3992 | sync_dirty_buffer(gdp_bh); | |
3993 | if (err) | |
3994 | break; | |
3995 | ||
3996 | block += thisgrp_len; | |
3997 | len -= thisgrp_len; | |
3998 | brelse(bitmap_bh); | |
3999 | BUG_ON(len < 0); | |
8016e29f HS |
4000 | } |
4001 | ||
8016e29f | 4002 | if (err) |
bfdc502a | 4003 | brelse(bitmap_bh); |
8016e29f HS |
4004 | } |
4005 | ||
c9de560d AT |
4006 | /* |
4007 | * here we normalize request for locality group | |
d7a1fee1 DE |
4008 | * Group request are normalized to s_mb_group_prealloc, which goes to |
4009 | * s_strip if we set the same via mount option. | |
4010 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 4011 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
4012 | * |
4013 | * XXX: should we try to preallocate more than the group has now? | |
4014 | */ | |
4015 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
4016 | { | |
4017 | struct super_block *sb = ac->ac_sb; | |
4018 | struct ext4_locality_group *lg = ac->ac_lg; | |
4019 | ||
4020 | BUG_ON(lg == NULL); | |
d7a1fee1 | 4021 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
d3df1453 | 4022 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
c9de560d AT |
4023 | } |
4024 | ||
4025 | /* | |
4026 | * Normalization means making request better in terms of | |
4027 | * size and alignment | |
4028 | */ | |
4ddfef7b ES |
4029 | static noinline_for_stack void |
4030 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
4031 | struct ext4_allocation_request *ar) |
4032 | { | |
53accfa9 | 4033 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
4034 | int bsbits, max; |
4035 | ext4_lblk_t end; | |
1592d2c5 CW |
4036 | loff_t size, start_off; |
4037 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 4038 | ext4_lblk_t start; |
c9de560d | 4039 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 4040 | struct ext4_prealloc_space *pa; |
c9de560d AT |
4041 | |
4042 | /* do normalize only data requests, metadata requests | |
4043 | do not need preallocation */ | |
4044 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4045 | return; | |
4046 | ||
4047 | /* sometime caller may want exact blocks */ | |
4048 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4049 | return; | |
4050 | ||
4051 | /* caller may indicate that preallocation isn't | |
4052 | * required (it's a tail, for example) */ | |
4053 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
4054 | return; | |
4055 | ||
4056 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
4057 | ext4_mb_normalize_group_request(ac); | |
4058 | return ; | |
4059 | } | |
4060 | ||
4061 | bsbits = ac->ac_sb->s_blocksize_bits; | |
4062 | ||
4063 | /* first, let's learn actual file size | |
4064 | * given current request is allocated */ | |
53accfa9 | 4065 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
4066 | size = size << bsbits; |
4067 | if (size < i_size_read(ac->ac_inode)) | |
4068 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 4069 | orig_size = size; |
c9de560d | 4070 | |
1930479c VC |
4071 | /* max size of free chunks */ |
4072 | max = 2 << bsbits; | |
c9de560d | 4073 | |
1930479c VC |
4074 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
4075 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
4076 | |
4077 | /* first, try to predict filesize */ | |
4078 | /* XXX: should this table be tunable? */ | |
4079 | start_off = 0; | |
4080 | if (size <= 16 * 1024) { | |
4081 | size = 16 * 1024; | |
4082 | } else if (size <= 32 * 1024) { | |
4083 | size = 32 * 1024; | |
4084 | } else if (size <= 64 * 1024) { | |
4085 | size = 64 * 1024; | |
4086 | } else if (size <= 128 * 1024) { | |
4087 | size = 128 * 1024; | |
4088 | } else if (size <= 256 * 1024) { | |
4089 | size = 256 * 1024; | |
4090 | } else if (size <= 512 * 1024) { | |
4091 | size = 512 * 1024; | |
4092 | } else if (size <= 1024 * 1024) { | |
4093 | size = 1024 * 1024; | |
1930479c | 4094 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 4095 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
4096 | (21 - bsbits)) << 21; |
4097 | size = 2 * 1024 * 1024; | |
4098 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
4099 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4100 | (22 - bsbits)) << 22; | |
4101 | size = 4 * 1024 * 1024; | |
4102 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 4103 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
4104 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4105 | (23 - bsbits)) << 23; | |
4106 | size = 8 * 1024 * 1024; | |
4107 | } else { | |
b27b1535 XW |
4108 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
4109 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
4110 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 4111 | } |
5a0790c2 AK |
4112 | size = size >> bsbits; |
4113 | start = start_off >> bsbits; | |
c9de560d | 4114 | |
a08f789d BL |
4115 | /* |
4116 | * For tiny groups (smaller than 8MB) the chosen allocation | |
4117 | * alignment may be larger than group size. Make sure the | |
4118 | * alignment does not move allocation to a different group which | |
4119 | * makes mballoc fail assertions later. | |
4120 | */ | |
4121 | start = max(start, rounddown(ac->ac_o_ex.fe_logical, | |
4122 | (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); | |
4123 | ||
c9de560d AT |
4124 | /* don't cover already allocated blocks in selected range */ |
4125 | if (ar->pleft && start <= ar->lleft) { | |
4126 | size -= ar->lleft + 1 - start; | |
4127 | start = ar->lleft + 1; | |
4128 | } | |
4129 | if (ar->pright && start + size - 1 >= ar->lright) | |
4130 | size -= start + size - ar->lright; | |
4131 | ||
cd648b8a JK |
4132 | /* |
4133 | * Trim allocation request for filesystems with artificially small | |
4134 | * groups. | |
4135 | */ | |
4136 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
4137 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
4138 | ||
c9de560d AT |
4139 | end = start + size; |
4140 | ||
4141 | /* check we don't cross already preallocated blocks */ | |
4142 | rcu_read_lock(); | |
9a0762c5 | 4143 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 4144 | ext4_lblk_t pa_end; |
c9de560d | 4145 | |
c9de560d AT |
4146 | if (pa->pa_deleted) |
4147 | continue; | |
4148 | spin_lock(&pa->pa_lock); | |
4149 | if (pa->pa_deleted) { | |
4150 | spin_unlock(&pa->pa_lock); | |
4151 | continue; | |
4152 | } | |
4153 | ||
53accfa9 TT |
4154 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
4155 | pa->pa_len); | |
c9de560d AT |
4156 | |
4157 | /* PA must not overlap original request */ | |
4158 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
4159 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
4160 | ||
38877f4e ES |
4161 | /* skip PAs this normalized request doesn't overlap with */ |
4162 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
4163 | spin_unlock(&pa->pa_lock); |
4164 | continue; | |
4165 | } | |
4166 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
4167 | ||
38877f4e | 4168 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
4169 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
4170 | BUG_ON(pa_end < start); | |
4171 | start = pa_end; | |
38877f4e | 4172 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
4173 | BUG_ON(pa->pa_lstart > end); |
4174 | end = pa->pa_lstart; | |
4175 | } | |
4176 | spin_unlock(&pa->pa_lock); | |
4177 | } | |
4178 | rcu_read_unlock(); | |
4179 | size = end - start; | |
4180 | ||
4181 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
4182 | rcu_read_lock(); | |
9a0762c5 | 4183 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 4184 | ext4_lblk_t pa_end; |
53accfa9 | 4185 | |
c9de560d AT |
4186 | spin_lock(&pa->pa_lock); |
4187 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
4188 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
4189 | pa->pa_len); | |
c9de560d AT |
4190 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
4191 | } | |
4192 | spin_unlock(&pa->pa_lock); | |
4193 | } | |
4194 | rcu_read_unlock(); | |
4195 | ||
cf4ff938 BL |
4196 | /* |
4197 | * In this function "start" and "size" are normalized for better | |
4198 | * alignment and length such that we could preallocate more blocks. | |
4199 | * This normalization is done such that original request of | |
4200 | * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and | |
4201 | * "size" boundaries. | |
4202 | * (Note fe_len can be relaxed since FS block allocation API does not | |
4203 | * provide gurantee on number of contiguous blocks allocation since that | |
4204 | * depends upon free space left, etc). | |
4205 | * In case of inode pa, later we use the allocated blocks | |
4206 | * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated | |
4207 | * range of goal/best blocks [start, size] to put it at the | |
4208 | * ac_o_ex.fe_logical extent of this inode. | |
4209 | * (See ext4_mb_use_inode_pa() for more details) | |
4210 | */ | |
4211 | if (start + size <= ac->ac_o_ex.fe_logical || | |
c9de560d | 4212 | start > ac->ac_o_ex.fe_logical) { |
9d8b9ec4 TT |
4213 | ext4_msg(ac->ac_sb, KERN_ERR, |
4214 | "start %lu, size %lu, fe_logical %lu", | |
4215 | (unsigned long) start, (unsigned long) size, | |
4216 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 4217 | BUG(); |
c9de560d | 4218 | } |
b5b60778 | 4219 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
4220 | |
4221 | /* now prepare goal request */ | |
4222 | ||
4223 | /* XXX: is it better to align blocks WRT to logical | |
4224 | * placement or satisfy big request as is */ | |
4225 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 4226 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
4227 | |
4228 | /* define goal start in order to merge */ | |
4229 | if (ar->pright && (ar->lright == (start + size))) { | |
4230 | /* merge to the right */ | |
4231 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
4232 | &ac->ac_f_ex.fe_group, | |
4233 | &ac->ac_f_ex.fe_start); | |
4234 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
4235 | } | |
4236 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
4237 | /* merge to the left */ | |
4238 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
4239 | &ac->ac_f_ex.fe_group, | |
4240 | &ac->ac_f_ex.fe_start); | |
4241 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
4242 | } | |
4243 | ||
d3df1453 RH |
4244 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
4245 | orig_size, start); | |
c9de560d AT |
4246 | } |
4247 | ||
4248 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
4249 | { | |
4250 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4251 | ||
a6c75eaf | 4252 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { |
c9de560d AT |
4253 | atomic_inc(&sbi->s_bal_reqs); |
4254 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 4255 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
4256 | atomic_inc(&sbi->s_bal_success); |
4257 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
a6c75eaf | 4258 | atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); |
c9de560d AT |
4259 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && |
4260 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
4261 | atomic_inc(&sbi->s_bal_goals); | |
4262 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
4263 | atomic_inc(&sbi->s_bal_breaks); | |
4264 | } | |
4265 | ||
296c355c TT |
4266 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
4267 | trace_ext4_mballoc_alloc(ac); | |
4268 | else | |
4269 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
4270 | } |
4271 | ||
b844167e CW |
4272 | /* |
4273 | * Called on failure; free up any blocks from the inode PA for this | |
4274 | * context. We don't need this for MB_GROUP_PA because we only change | |
4275 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
4276 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
4277 | */ | |
4278 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
4279 | { | |
4280 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
4281 | struct ext4_buddy e4b; |
4282 | int err; | |
b844167e | 4283 | |
86f0afd4 | 4284 | if (pa == NULL) { |
c99d1e6e TT |
4285 | if (ac->ac_f_ex.fe_len == 0) |
4286 | return; | |
86f0afd4 TT |
4287 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
4288 | if (err) { | |
4289 | /* | |
4290 | * This should never happen since we pin the | |
4291 | * pages in the ext4_allocation_context so | |
4292 | * ext4_mb_load_buddy() should never fail. | |
4293 | */ | |
4294 | WARN(1, "mb_load_buddy failed (%d)", err); | |
4295 | return; | |
4296 | } | |
4297 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
4298 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
4299 | ac->ac_f_ex.fe_len); | |
4300 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 4301 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
4302 | return; |
4303 | } | |
4304 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 4305 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
4306 | } |
4307 | ||
c9de560d AT |
4308 | /* |
4309 | * use blocks preallocated to inode | |
4310 | */ | |
4311 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
4312 | struct ext4_prealloc_space *pa) | |
4313 | { | |
53accfa9 | 4314 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
4315 | ext4_fsblk_t start; |
4316 | ext4_fsblk_t end; | |
4317 | int len; | |
4318 | ||
4319 | /* found preallocated blocks, use them */ | |
4320 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
4321 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
4322 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
4323 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
4324 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
4325 | &ac->ac_b_ex.fe_start); | |
4326 | ac->ac_b_ex.fe_len = len; | |
4327 | ac->ac_status = AC_STATUS_FOUND; | |
4328 | ac->ac_pa = pa; | |
4329 | ||
4330 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 4331 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
4332 | BUG_ON(pa->pa_free < len); |
4333 | pa->pa_free -= len; | |
4334 | ||
d3df1453 | 4335 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
4336 | } |
4337 | ||
4338 | /* | |
4339 | * use blocks preallocated to locality group | |
4340 | */ | |
4341 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
4342 | struct ext4_prealloc_space *pa) | |
4343 | { | |
03cddb80 | 4344 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 4345 | |
c9de560d AT |
4346 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
4347 | &ac->ac_b_ex.fe_group, | |
4348 | &ac->ac_b_ex.fe_start); | |
4349 | ac->ac_b_ex.fe_len = len; | |
4350 | ac->ac_status = AC_STATUS_FOUND; | |
4351 | ac->ac_pa = pa; | |
4352 | ||
4353 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 4354 | * possible race when the group is being loaded concurrently |
c9de560d | 4355 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
4356 | * in on-disk bitmap -- see ext4_mb_release_context() |
4357 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 4358 | */ |
d3df1453 RH |
4359 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
4360 | pa->pa_lstart-len, len, pa); | |
c9de560d AT |
4361 | } |
4362 | ||
5e745b04 AK |
4363 | /* |
4364 | * Return the prealloc space that have minimal distance | |
4365 | * from the goal block. @cpa is the prealloc | |
4366 | * space that is having currently known minimal distance | |
4367 | * from the goal block. | |
4368 | */ | |
4369 | static struct ext4_prealloc_space * | |
4370 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
4371 | struct ext4_prealloc_space *pa, | |
4372 | struct ext4_prealloc_space *cpa) | |
4373 | { | |
4374 | ext4_fsblk_t cur_distance, new_distance; | |
4375 | ||
4376 | if (cpa == NULL) { | |
4377 | atomic_inc(&pa->pa_count); | |
4378 | return pa; | |
4379 | } | |
79211c8e AM |
4380 | cur_distance = abs(goal_block - cpa->pa_pstart); |
4381 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 4382 | |
5a54b2f1 | 4383 | if (cur_distance <= new_distance) |
5e745b04 AK |
4384 | return cpa; |
4385 | ||
4386 | /* drop the previous reference */ | |
4387 | atomic_dec(&cpa->pa_count); | |
4388 | atomic_inc(&pa->pa_count); | |
4389 | return pa; | |
4390 | } | |
4391 | ||
c9de560d AT |
4392 | /* |
4393 | * search goal blocks in preallocated space | |
4394 | */ | |
4fca8f07 | 4395 | static noinline_for_stack bool |
4ddfef7b | 4396 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 4397 | { |
53accfa9 | 4398 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 4399 | int order, i; |
c9de560d AT |
4400 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
4401 | struct ext4_locality_group *lg; | |
5e745b04 AK |
4402 | struct ext4_prealloc_space *pa, *cpa = NULL; |
4403 | ext4_fsblk_t goal_block; | |
c9de560d AT |
4404 | |
4405 | /* only data can be preallocated */ | |
4406 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 4407 | return false; |
c9de560d AT |
4408 | |
4409 | /* first, try per-file preallocation */ | |
4410 | rcu_read_lock(); | |
9a0762c5 | 4411 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
4412 | |
4413 | /* all fields in this condition don't change, | |
4414 | * so we can skip locking for them */ | |
4415 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
4416 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
4417 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
4418 | continue; |
4419 | ||
fb0a387d | 4420 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 4421 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
4422 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
4423 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
4424 | continue; |
4425 | ||
c9de560d AT |
4426 | /* found preallocated blocks, use them */ |
4427 | spin_lock(&pa->pa_lock); | |
4428 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
4429 | atomic_inc(&pa->pa_count); | |
4430 | ext4_mb_use_inode_pa(ac, pa); | |
4431 | spin_unlock(&pa->pa_lock); | |
4432 | ac->ac_criteria = 10; | |
4433 | rcu_read_unlock(); | |
4fca8f07 | 4434 | return true; |
c9de560d AT |
4435 | } |
4436 | spin_unlock(&pa->pa_lock); | |
4437 | } | |
4438 | rcu_read_unlock(); | |
4439 | ||
4440 | /* can we use group allocation? */ | |
4441 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 4442 | return false; |
c9de560d AT |
4443 | |
4444 | /* inode may have no locality group for some reason */ | |
4445 | lg = ac->ac_lg; | |
4446 | if (lg == NULL) | |
4fca8f07 | 4447 | return false; |
6be2ded1 AK |
4448 | order = fls(ac->ac_o_ex.fe_len) - 1; |
4449 | if (order > PREALLOC_TB_SIZE - 1) | |
4450 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4451 | order = PREALLOC_TB_SIZE - 1; | |
4452 | ||
bda00de7 | 4453 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
4454 | /* |
4455 | * search for the prealloc space that is having | |
4456 | * minimal distance from the goal block. | |
4457 | */ | |
6be2ded1 AK |
4458 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
4459 | rcu_read_lock(); | |
4460 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
4461 | pa_inode_list) { | |
4462 | spin_lock(&pa->pa_lock); | |
4463 | if (pa->pa_deleted == 0 && | |
4464 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
4465 | |
4466 | cpa = ext4_mb_check_group_pa(goal_block, | |
4467 | pa, cpa); | |
6be2ded1 | 4468 | } |
c9de560d | 4469 | spin_unlock(&pa->pa_lock); |
c9de560d | 4470 | } |
6be2ded1 | 4471 | rcu_read_unlock(); |
c9de560d | 4472 | } |
5e745b04 AK |
4473 | if (cpa) { |
4474 | ext4_mb_use_group_pa(ac, cpa); | |
4475 | ac->ac_criteria = 20; | |
4fca8f07 | 4476 | return true; |
5e745b04 | 4477 | } |
4fca8f07 | 4478 | return false; |
c9de560d AT |
4479 | } |
4480 | ||
7a2fcbf7 AK |
4481 | /* |
4482 | * the function goes through all block freed in the group | |
4483 | * but not yet committed and marks them used in in-core bitmap. | |
4484 | * buddy must be generated from this bitmap | |
955ce5f5 | 4485 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
4486 | */ |
4487 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
4488 | ext4_group_t group) | |
4489 | { | |
4490 | struct rb_node *n; | |
4491 | struct ext4_group_info *grp; | |
4492 | struct ext4_free_data *entry; | |
4493 | ||
4494 | grp = ext4_get_group_info(sb, group); | |
4495 | n = rb_first(&(grp->bb_free_root)); | |
4496 | ||
4497 | while (n) { | |
18aadd47 | 4498 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
123e3016 | 4499 | mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); |
7a2fcbf7 AK |
4500 | n = rb_next(n); |
4501 | } | |
4502 | return; | |
4503 | } | |
4504 | ||
c9de560d AT |
4505 | /* |
4506 | * the function goes through all preallocation in this group and marks them | |
4507 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 4508 | * Need to be called with ext4 group lock held |
c9de560d | 4509 | */ |
089ceecc ES |
4510 | static noinline_for_stack |
4511 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
4512 | ext4_group_t group) |
4513 | { | |
4514 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4515 | struct ext4_prealloc_space *pa; | |
4516 | struct list_head *cur; | |
4517 | ext4_group_t groupnr; | |
4518 | ext4_grpblk_t start; | |
4519 | int preallocated = 0; | |
c9de560d AT |
4520 | int len; |
4521 | ||
4522 | /* all form of preallocation discards first load group, | |
4523 | * so the only competing code is preallocation use. | |
4524 | * we don't need any locking here | |
4525 | * notice we do NOT ignore preallocations with pa_deleted | |
4526 | * otherwise we could leave used blocks available for | |
4527 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
4528 | * is dropping preallocation | |
4529 | */ | |
4530 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4531 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
4532 | spin_lock(&pa->pa_lock); | |
4533 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4534 | &groupnr, &start); | |
4535 | len = pa->pa_len; | |
4536 | spin_unlock(&pa->pa_lock); | |
4537 | if (unlikely(len == 0)) | |
4538 | continue; | |
4539 | BUG_ON(groupnr != group); | |
123e3016 | 4540 | mb_set_bits(bitmap, start, len); |
c9de560d | 4541 | preallocated += len; |
c9de560d | 4542 | } |
d3df1453 | 4543 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
4544 | } |
4545 | ||
27bc446e | 4546 | static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
4547 | struct ext4_prealloc_space *pa) | |
4548 | { | |
4549 | struct ext4_inode_info *ei; | |
4550 | ||
4551 | if (pa->pa_deleted) { | |
4552 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", | |
4553 | pa->pa_type, pa->pa_pstart, pa->pa_lstart, | |
4554 | pa->pa_len); | |
4555 | return; | |
4556 | } | |
4557 | ||
4558 | pa->pa_deleted = 1; | |
4559 | ||
4560 | if (pa->pa_type == MB_INODE_PA) { | |
4561 | ei = EXT4_I(pa->pa_inode); | |
4562 | atomic_dec(&ei->i_prealloc_active); | |
4563 | } | |
4564 | } | |
4565 | ||
c9de560d AT |
4566 | static void ext4_mb_pa_callback(struct rcu_head *head) |
4567 | { | |
4568 | struct ext4_prealloc_space *pa; | |
4569 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
4570 | |
4571 | BUG_ON(atomic_read(&pa->pa_count)); | |
4572 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
4573 | kmem_cache_free(ext4_pspace_cachep, pa); |
4574 | } | |
4575 | ||
4576 | /* | |
4577 | * drops a reference to preallocated space descriptor | |
4578 | * if this was the last reference and the space is consumed | |
4579 | */ | |
4580 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
4581 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
4582 | { | |
a9df9a49 | 4583 | ext4_group_t grp; |
d33a1976 | 4584 | ext4_fsblk_t grp_blk; |
c9de560d | 4585 | |
c9de560d AT |
4586 | /* in this short window concurrent discard can set pa_deleted */ |
4587 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
4588 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
4589 | spin_unlock(&pa->pa_lock); | |
4590 | return; | |
4591 | } | |
4592 | ||
c9de560d AT |
4593 | if (pa->pa_deleted == 1) { |
4594 | spin_unlock(&pa->pa_lock); | |
4595 | return; | |
4596 | } | |
4597 | ||
27bc446e | 4598 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4599 | spin_unlock(&pa->pa_lock); |
4600 | ||
d33a1976 | 4601 | grp_blk = pa->pa_pstart; |
60e6679e | 4602 | /* |
cc0fb9ad AK |
4603 | * If doing group-based preallocation, pa_pstart may be in the |
4604 | * next group when pa is used up | |
4605 | */ | |
4606 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
4607 | grp_blk--; |
4608 | ||
bd86298e | 4609 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
4610 | |
4611 | /* | |
4612 | * possible race: | |
4613 | * | |
4614 | * P1 (buddy init) P2 (regular allocation) | |
4615 | * find block B in PA | |
4616 | * copy on-disk bitmap to buddy | |
4617 | * mark B in on-disk bitmap | |
4618 | * drop PA from group | |
4619 | * mark all PAs in buddy | |
4620 | * | |
4621 | * thus, P1 initializes buddy with B available. to prevent this | |
4622 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
4623 | * against that pair | |
4624 | */ | |
4625 | ext4_lock_group(sb, grp); | |
4626 | list_del(&pa->pa_group_list); | |
4627 | ext4_unlock_group(sb, grp); | |
4628 | ||
4629 | spin_lock(pa->pa_obj_lock); | |
4630 | list_del_rcu(&pa->pa_inode_list); | |
4631 | spin_unlock(pa->pa_obj_lock); | |
4632 | ||
4633 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4634 | } | |
4635 | ||
4636 | /* | |
4637 | * creates new preallocated space for given inode | |
4638 | */ | |
53f86b17 | 4639 | static noinline_for_stack void |
4ddfef7b | 4640 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4641 | { |
4642 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 4643 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
4644 | struct ext4_prealloc_space *pa; |
4645 | struct ext4_group_info *grp; | |
4646 | struct ext4_inode_info *ei; | |
4647 | ||
4648 | /* preallocate only when found space is larger then requested */ | |
4649 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4650 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4651 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4652 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4653 | |
53f86b17 | 4654 | pa = ac->ac_pa; |
c9de560d AT |
4655 | |
4656 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
4657 | int winl; | |
4658 | int wins; | |
4659 | int win; | |
4660 | int offs; | |
4661 | ||
4662 | /* we can't allocate as much as normalizer wants. | |
4663 | * so, found space must get proper lstart | |
4664 | * to cover original request */ | |
4665 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
4666 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
4667 | ||
4668 | /* we're limited by original request in that | |
4669 | * logical block must be covered any way | |
4670 | * winl is window we can move our chunk within */ | |
4671 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
4672 | ||
4673 | /* also, we should cover whole original request */ | |
53accfa9 | 4674 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
4675 | |
4676 | /* the smallest one defines real window */ | |
4677 | win = min(winl, wins); | |
4678 | ||
53accfa9 TT |
4679 | offs = ac->ac_o_ex.fe_logical % |
4680 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
4681 | if (offs && offs < win) |
4682 | win = offs; | |
4683 | ||
53accfa9 | 4684 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 4685 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
4686 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
4687 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
4688 | } | |
4689 | ||
4690 | /* preallocation can change ac_b_ex, thus we store actually | |
4691 | * allocated blocks for history */ | |
4692 | ac->ac_f_ex = ac->ac_b_ex; | |
4693 | ||
4694 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
4695 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4696 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4697 | pa->pa_free = pa->pa_len; | |
c9de560d | 4698 | spin_lock_init(&pa->pa_lock); |
d794bf8e AK |
4699 | INIT_LIST_HEAD(&pa->pa_inode_list); |
4700 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 4701 | pa->pa_deleted = 0; |
cc0fb9ad | 4702 | pa->pa_type = MB_INODE_PA; |
c9de560d | 4703 | |
d3df1453 RH |
4704 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4705 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4706 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
4707 | |
4708 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 4709 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
4710 | |
4711 | ei = EXT4_I(ac->ac_inode); | |
4712 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4713 | ||
4714 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
4715 | pa->pa_inode = ac->ac_inode; | |
4716 | ||
c9de560d | 4717 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d AT |
4718 | |
4719 | spin_lock(pa->pa_obj_lock); | |
4720 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
4721 | spin_unlock(pa->pa_obj_lock); | |
27bc446e | 4722 | atomic_inc(&ei->i_prealloc_active); |
c9de560d AT |
4723 | } |
4724 | ||
4725 | /* | |
4726 | * creates new preallocated space for locality group inodes belongs to | |
4727 | */ | |
53f86b17 | 4728 | static noinline_for_stack void |
4ddfef7b | 4729 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4730 | { |
4731 | struct super_block *sb = ac->ac_sb; | |
4732 | struct ext4_locality_group *lg; | |
4733 | struct ext4_prealloc_space *pa; | |
4734 | struct ext4_group_info *grp; | |
4735 | ||
4736 | /* preallocate only when found space is larger then requested */ | |
4737 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4738 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4739 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4740 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4741 | |
53f86b17 | 4742 | pa = ac->ac_pa; |
c9de560d AT |
4743 | |
4744 | /* preallocation can change ac_b_ex, thus we store actually | |
4745 | * allocated blocks for history */ | |
4746 | ac->ac_f_ex = ac->ac_b_ex; | |
4747 | ||
4748 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4749 | pa->pa_lstart = pa->pa_pstart; | |
4750 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4751 | pa->pa_free = pa->pa_len; | |
c9de560d | 4752 | spin_lock_init(&pa->pa_lock); |
6be2ded1 | 4753 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 4754 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 4755 | pa->pa_deleted = 0; |
cc0fb9ad | 4756 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 4757 | |
d3df1453 RH |
4758 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4759 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4760 | trace_ext4_mb_new_group_pa(ac, pa); |
c9de560d AT |
4761 | |
4762 | ext4_mb_use_group_pa(ac, pa); | |
4763 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
4764 | ||
4765 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4766 | lg = ac->ac_lg; | |
4767 | BUG_ON(lg == NULL); | |
4768 | ||
4769 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
4770 | pa->pa_inode = NULL; | |
4771 | ||
c9de560d | 4772 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 4773 | |
6be2ded1 AK |
4774 | /* |
4775 | * We will later add the new pa to the right bucket | |
4776 | * after updating the pa_free in ext4_mb_release_context | |
4777 | */ | |
c9de560d AT |
4778 | } |
4779 | ||
53f86b17 | 4780 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
c9de560d | 4781 | { |
c9de560d | 4782 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
53f86b17 | 4783 | ext4_mb_new_group_pa(ac); |
c9de560d | 4784 | else |
53f86b17 | 4785 | ext4_mb_new_inode_pa(ac); |
c9de560d AT |
4786 | } |
4787 | ||
4788 | /* | |
4789 | * finds all unused blocks in on-disk bitmap, frees them in | |
4790 | * in-core bitmap and buddy. | |
4791 | * @pa must be unlinked from inode and group lists, so that | |
4792 | * nobody else can find/use it. | |
4793 | * the caller MUST hold group/inode locks. | |
4794 | * TODO: optimize the case when there are no in-core structures yet | |
4795 | */ | |
4ddfef7b ES |
4796 | static noinline_for_stack int |
4797 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 4798 | struct ext4_prealloc_space *pa) |
c9de560d | 4799 | { |
c9de560d AT |
4800 | struct super_block *sb = e4b->bd_sb; |
4801 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
4802 | unsigned int end; |
4803 | unsigned int next; | |
c9de560d AT |
4804 | ext4_group_t group; |
4805 | ext4_grpblk_t bit; | |
ba80b101 | 4806 | unsigned long long grp_blk_start; |
c9de560d AT |
4807 | int free = 0; |
4808 | ||
4809 | BUG_ON(pa->pa_deleted == 0); | |
4810 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 4811 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
4812 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
4813 | end = bit + pa->pa_len; | |
4814 | ||
c9de560d | 4815 | while (bit < end) { |
ffad0a44 | 4816 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
4817 | if (bit >= end) |
4818 | break; | |
ffad0a44 | 4819 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
d3df1453 | 4820 | mb_debug(sb, "free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
4821 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
4822 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
4823 | free += next - bit; |
4824 | ||
3e1e5f50 | 4825 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
4826 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
4827 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 4828 | next - bit); |
c9de560d AT |
4829 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
4830 | bit = next + 1; | |
4831 | } | |
4832 | if (free != pa->pa_free) { | |
9d8b9ec4 | 4833 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 4834 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
4835 | pa, (unsigned long) pa->pa_lstart, |
4836 | (unsigned long) pa->pa_pstart, | |
36bad423 | 4837 | pa->pa_len); |
e29136f8 | 4838 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 4839 | free, pa->pa_free); |
e56eb659 AK |
4840 | /* |
4841 | * pa is already deleted so we use the value obtained | |
4842 | * from the bitmap and continue. | |
4843 | */ | |
c9de560d | 4844 | } |
c9de560d AT |
4845 | atomic_add(free, &sbi->s_mb_discarded); |
4846 | ||
863c37fc | 4847 | return 0; |
c9de560d AT |
4848 | } |
4849 | ||
4ddfef7b ES |
4850 | static noinline_for_stack int |
4851 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 4852 | struct ext4_prealloc_space *pa) |
c9de560d | 4853 | { |
c9de560d AT |
4854 | struct super_block *sb = e4b->bd_sb; |
4855 | ext4_group_t group; | |
4856 | ext4_grpblk_t bit; | |
4857 | ||
60e07cf5 | 4858 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
4859 | BUG_ON(pa->pa_deleted == 0); |
4860 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
4861 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
4862 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
4863 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 4864 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
4865 | |
4866 | return 0; | |
4867 | } | |
4868 | ||
4869 | /* | |
4870 | * releases all preallocations in given group | |
4871 | * | |
4872 | * first, we need to decide discard policy: | |
4873 | * - when do we discard | |
4874 | * 1) ENOSPC | |
4875 | * - how many do we discard | |
4876 | * 1) how many requested | |
4877 | */ | |
4ddfef7b ES |
4878 | static noinline_for_stack int |
4879 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
8c80fb31 | 4880 | ext4_group_t group, int *busy) |
c9de560d AT |
4881 | { |
4882 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4883 | struct buffer_head *bitmap_bh = NULL; | |
4884 | struct ext4_prealloc_space *pa, *tmp; | |
4885 | struct list_head list; | |
4886 | struct ext4_buddy e4b; | |
4887 | int err; | |
8c80fb31 | 4888 | int free = 0; |
c9de560d | 4889 | |
d3df1453 | 4890 | mb_debug(sb, "discard preallocation for group %u\n", group); |
c9de560d | 4891 | if (list_empty(&grp->bb_prealloc_list)) |
bbc4ec77 | 4892 | goto out_dbg; |
c9de560d | 4893 | |
574ca174 | 4894 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4895 | if (IS_ERR(bitmap_bh)) { |
4896 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4897 | ext4_error_err(sb, -err, |
4898 | "Error %d reading block bitmap for %u", | |
4899 | err, group); | |
bbc4ec77 | 4900 | goto out_dbg; |
c9de560d AT |
4901 | } |
4902 | ||
4903 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 4904 | if (err) { |
9651e6b2 KK |
4905 | ext4_warning(sb, "Error %d loading buddy information for %u", |
4906 | err, group); | |
ce89f46c | 4907 | put_bh(bitmap_bh); |
bbc4ec77 | 4908 | goto out_dbg; |
ce89f46c | 4909 | } |
c9de560d | 4910 | |
c9de560d | 4911 | INIT_LIST_HEAD(&list); |
c9de560d AT |
4912 | ext4_lock_group(sb, group); |
4913 | list_for_each_entry_safe(pa, tmp, | |
4914 | &grp->bb_prealloc_list, pa_group_list) { | |
4915 | spin_lock(&pa->pa_lock); | |
4916 | if (atomic_read(&pa->pa_count)) { | |
4917 | spin_unlock(&pa->pa_lock); | |
8c80fb31 | 4918 | *busy = 1; |
c9de560d AT |
4919 | continue; |
4920 | } | |
4921 | if (pa->pa_deleted) { | |
4922 | spin_unlock(&pa->pa_lock); | |
4923 | continue; | |
4924 | } | |
4925 | ||
4926 | /* seems this one can be freed ... */ | |
27bc446e | 4927 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 4928 | |
70022da8 YB |
4929 | if (!free) |
4930 | this_cpu_inc(discard_pa_seq); | |
4931 | ||
c9de560d AT |
4932 | /* we can trust pa_free ... */ |
4933 | free += pa->pa_free; | |
4934 | ||
4935 | spin_unlock(&pa->pa_lock); | |
4936 | ||
4937 | list_del(&pa->pa_group_list); | |
4938 | list_add(&pa->u.pa_tmp_list, &list); | |
4939 | } | |
4940 | ||
c9de560d AT |
4941 | /* now free all selected PAs */ |
4942 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
4943 | ||
4944 | /* remove from object (inode or locality group) */ | |
4945 | spin_lock(pa->pa_obj_lock); | |
4946 | list_del_rcu(&pa->pa_inode_list); | |
4947 | spin_unlock(pa->pa_obj_lock); | |
4948 | ||
cc0fb9ad | 4949 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 4950 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 4951 | else |
3e1e5f50 | 4952 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4953 | |
4954 | list_del(&pa->u.pa_tmp_list); | |
4955 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4956 | } | |
4957 | ||
c9de560d | 4958 | ext4_unlock_group(sb, group); |
e39e07fd | 4959 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4960 | put_bh(bitmap_bh); |
bbc4ec77 | 4961 | out_dbg: |
d3df1453 | 4962 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
8c80fb31 CX |
4963 | free, group, grp->bb_free); |
4964 | return free; | |
c9de560d AT |
4965 | } |
4966 | ||
4967 | /* | |
4968 | * releases all non-used preallocated blocks for given inode | |
4969 | * | |
4970 | * It's important to discard preallocations under i_data_sem | |
4971 | * We don't want another block to be served from the prealloc | |
4972 | * space when we are discarding the inode prealloc space. | |
4973 | * | |
4974 | * FIXME!! Make sure it is valid at all the call sites | |
4975 | */ | |
27bc446e | 4976 | void ext4_discard_preallocations(struct inode *inode, unsigned int needed) |
c9de560d AT |
4977 | { |
4978 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4979 | struct super_block *sb = inode->i_sb; | |
4980 | struct buffer_head *bitmap_bh = NULL; | |
4981 | struct ext4_prealloc_space *pa, *tmp; | |
4982 | ext4_group_t group = 0; | |
4983 | struct list_head list; | |
4984 | struct ext4_buddy e4b; | |
4985 | int err; | |
4986 | ||
c2ea3fde | 4987 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4988 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4989 | return; | |
4990 | } | |
4991 | ||
8016e29f HS |
4992 | if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) |
4993 | return; | |
4994 | ||
d3df1453 RH |
4995 | mb_debug(sb, "discard preallocation for inode %lu\n", |
4996 | inode->i_ino); | |
27bc446e | 4997 | trace_ext4_discard_preallocations(inode, |
4998 | atomic_read(&ei->i_prealloc_active), needed); | |
c9de560d AT |
4999 | |
5000 | INIT_LIST_HEAD(&list); | |
5001 | ||
27bc446e | 5002 | if (needed == 0) |
5003 | needed = UINT_MAX; | |
5004 | ||
c9de560d AT |
5005 | repeat: |
5006 | /* first, collect all pa's in the inode */ | |
5007 | spin_lock(&ei->i_prealloc_lock); | |
27bc446e | 5008 | while (!list_empty(&ei->i_prealloc_list) && needed) { |
5009 | pa = list_entry(ei->i_prealloc_list.prev, | |
c9de560d AT |
5010 | struct ext4_prealloc_space, pa_inode_list); |
5011 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
5012 | spin_lock(&pa->pa_lock); | |
5013 | if (atomic_read(&pa->pa_count)) { | |
5014 | /* this shouldn't happen often - nobody should | |
5015 | * use preallocation while we're discarding it */ | |
5016 | spin_unlock(&pa->pa_lock); | |
5017 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
5018 | ext4_msg(sb, KERN_ERR, |
5019 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
5020 | WARN_ON(1); |
5021 | schedule_timeout_uninterruptible(HZ); | |
5022 | goto repeat; | |
5023 | ||
5024 | } | |
5025 | if (pa->pa_deleted == 0) { | |
27bc446e | 5026 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
5027 | spin_unlock(&pa->pa_lock); |
5028 | list_del_rcu(&pa->pa_inode_list); | |
5029 | list_add(&pa->u.pa_tmp_list, &list); | |
27bc446e | 5030 | needed--; |
c9de560d AT |
5031 | continue; |
5032 | } | |
5033 | ||
5034 | /* someone is deleting pa right now */ | |
5035 | spin_unlock(&pa->pa_lock); | |
5036 | spin_unlock(&ei->i_prealloc_lock); | |
5037 | ||
5038 | /* we have to wait here because pa_deleted | |
5039 | * doesn't mean pa is already unlinked from | |
5040 | * the list. as we might be called from | |
5041 | * ->clear_inode() the inode will get freed | |
5042 | * and concurrent thread which is unlinking | |
5043 | * pa from inode's list may access already | |
5044 | * freed memory, bad-bad-bad */ | |
5045 | ||
5046 | /* XXX: if this happens too often, we can | |
5047 | * add a flag to force wait only in case | |
5048 | * of ->clear_inode(), but not in case of | |
5049 | * regular truncate */ | |
5050 | schedule_timeout_uninterruptible(HZ); | |
5051 | goto repeat; | |
5052 | } | |
5053 | spin_unlock(&ei->i_prealloc_lock); | |
5054 | ||
5055 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 5056 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 5057 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 5058 | |
9651e6b2 KK |
5059 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5060 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 5061 | if (err) { |
54d3adbc TT |
5062 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5063 | err, group); | |
ce89f46c AK |
5064 | continue; |
5065 | } | |
c9de560d | 5066 | |
574ca174 | 5067 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5068 | if (IS_ERR(bitmap_bh)) { |
5069 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5070 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
5071 | err, group); | |
e39e07fd | 5072 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 5073 | continue; |
c9de560d AT |
5074 | } |
5075 | ||
5076 | ext4_lock_group(sb, group); | |
5077 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5078 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
5079 | ext4_unlock_group(sb, group); |
5080 | ||
e39e07fd | 5081 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
5082 | put_bh(bitmap_bh); |
5083 | ||
5084 | list_del(&pa->u.pa_tmp_list); | |
5085 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5086 | } | |
5087 | } | |
5088 | ||
53f86b17 RH |
5089 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
5090 | { | |
5091 | struct ext4_prealloc_space *pa; | |
5092 | ||
5093 | BUG_ON(ext4_pspace_cachep == NULL); | |
5094 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); | |
5095 | if (!pa) | |
5096 | return -ENOMEM; | |
5097 | atomic_set(&pa->pa_count, 1); | |
5098 | ac->ac_pa = pa; | |
5099 | return 0; | |
5100 | } | |
5101 | ||
5102 | static void ext4_mb_pa_free(struct ext4_allocation_context *ac) | |
5103 | { | |
5104 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
5105 | ||
5106 | BUG_ON(!pa); | |
5107 | ac->ac_pa = NULL; | |
5108 | WARN_ON(!atomic_dec_and_test(&pa->pa_count)); | |
5109 | kmem_cache_free(ext4_pspace_cachep, pa); | |
5110 | } | |
5111 | ||
6ba495e9 | 5112 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 5113 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 5114 | { |
e68cf40c | 5115 | ext4_group_t i, ngroups; |
c9de560d | 5116 | |
9b5f6c9b | 5117 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e3570639 ES |
5118 | return; |
5119 | ||
8df9675f | 5120 | ngroups = ext4_get_groups_count(sb); |
d3df1453 | 5121 | mb_debug(sb, "groups: "); |
8df9675f | 5122 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
5123 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
5124 | struct ext4_prealloc_space *pa; | |
5125 | ext4_grpblk_t start; | |
5126 | struct list_head *cur; | |
5127 | ext4_lock_group(sb, i); | |
5128 | list_for_each(cur, &grp->bb_prealloc_list) { | |
5129 | pa = list_entry(cur, struct ext4_prealloc_space, | |
5130 | pa_group_list); | |
5131 | spin_lock(&pa->pa_lock); | |
5132 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
5133 | NULL, &start); | |
5134 | spin_unlock(&pa->pa_lock); | |
d3df1453 RH |
5135 | mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
5136 | pa->pa_len); | |
c9de560d | 5137 | } |
60bd63d1 | 5138 | ext4_unlock_group(sb, i); |
d3df1453 RH |
5139 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
5140 | grp->bb_fragments); | |
c9de560d | 5141 | } |
c9de560d | 5142 | } |
e68cf40c RH |
5143 | |
5144 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
5145 | { | |
5146 | struct super_block *sb = ac->ac_sb; | |
5147 | ||
9b5f6c9b | 5148 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e68cf40c RH |
5149 | return; |
5150 | ||
d3df1453 | 5151 | mb_debug(sb, "Can't allocate:" |
e68cf40c | 5152 | " Allocation context details:"); |
d3df1453 | 5153 | mb_debug(sb, "status %u flags 0x%x", |
e68cf40c | 5154 | ac->ac_status, ac->ac_flags); |
d3df1453 | 5155 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
e68cf40c RH |
5156 | "goal %lu/%lu/%lu@%lu, " |
5157 | "best %lu/%lu/%lu@%lu cr %d", | |
5158 | (unsigned long)ac->ac_o_ex.fe_group, | |
5159 | (unsigned long)ac->ac_o_ex.fe_start, | |
5160 | (unsigned long)ac->ac_o_ex.fe_len, | |
5161 | (unsigned long)ac->ac_o_ex.fe_logical, | |
5162 | (unsigned long)ac->ac_g_ex.fe_group, | |
5163 | (unsigned long)ac->ac_g_ex.fe_start, | |
5164 | (unsigned long)ac->ac_g_ex.fe_len, | |
5165 | (unsigned long)ac->ac_g_ex.fe_logical, | |
5166 | (unsigned long)ac->ac_b_ex.fe_group, | |
5167 | (unsigned long)ac->ac_b_ex.fe_start, | |
5168 | (unsigned long)ac->ac_b_ex.fe_len, | |
5169 | (unsigned long)ac->ac_b_ex.fe_logical, | |
5170 | (int)ac->ac_criteria); | |
d3df1453 | 5171 | mb_debug(sb, "%u found", ac->ac_found); |
e68cf40c RH |
5172 | ext4_mb_show_pa(sb); |
5173 | } | |
c9de560d | 5174 | #else |
e68cf40c RH |
5175 | static inline void ext4_mb_show_pa(struct super_block *sb) |
5176 | { | |
5177 | return; | |
5178 | } | |
c9de560d AT |
5179 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
5180 | { | |
e68cf40c | 5181 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
5182 | return; |
5183 | } | |
5184 | #endif | |
5185 | ||
5186 | /* | |
5187 | * We use locality group preallocation for small size file. The size of the | |
5188 | * file is determined by the current size or the resulting size after | |
5189 | * allocation which ever is larger | |
5190 | * | |
b713a5ec | 5191 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
5192 | */ |
5193 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
5194 | { | |
5195 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
5196 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
5197 | loff_t size, isize; | |
a9f2a293 | 5198 | bool inode_pa_eligible, group_pa_eligible; |
c9de560d AT |
5199 | |
5200 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
5201 | return; | |
5202 | ||
4ba74d00 TT |
5203 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
5204 | return; | |
5205 | ||
a9f2a293 JK |
5206 | group_pa_eligible = sbi->s_mb_group_prealloc > 0; |
5207 | inode_pa_eligible = true; | |
53accfa9 | 5208 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
5209 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
5210 | >> bsbits; | |
c9de560d | 5211 | |
a9f2a293 | 5212 | /* No point in using inode preallocation for closed files */ |
82dd124c | 5213 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
a9f2a293 JK |
5214 | !inode_is_open_for_write(ac->ac_inode)) |
5215 | inode_pa_eligible = false; | |
50797481 | 5216 | |
71780577 | 5217 | size = max(size, isize); |
a9f2a293 JK |
5218 | /* Don't use group allocation for large files */ |
5219 | if (size > sbi->s_mb_stream_request) | |
5220 | group_pa_eligible = false; | |
5221 | ||
5222 | if (!group_pa_eligible) { | |
5223 | if (inode_pa_eligible) | |
5224 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
5225 | else | |
5226 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
c9de560d | 5227 | return; |
4ba74d00 | 5228 | } |
c9de560d AT |
5229 | |
5230 | BUG_ON(ac->ac_lg != NULL); | |
5231 | /* | |
5232 | * locality group prealloc space are per cpu. The reason for having | |
5233 | * per cpu locality group is to reduce the contention between block | |
5234 | * request from multiple CPUs. | |
5235 | */ | |
a0b6bc63 | 5236 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
5237 | |
5238 | /* we're going to use group allocation */ | |
5239 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
5240 | ||
5241 | /* serialize all allocations in the group */ | |
5242 | mutex_lock(&ac->ac_lg->lg_mutex); | |
5243 | } | |
5244 | ||
4ddfef7b ES |
5245 | static noinline_for_stack int |
5246 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
5247 | struct ext4_allocation_request *ar) |
5248 | { | |
5249 | struct super_block *sb = ar->inode->i_sb; | |
5250 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
5251 | struct ext4_super_block *es = sbi->s_es; | |
5252 | ext4_group_t group; | |
498e5f24 TT |
5253 | unsigned int len; |
5254 | ext4_fsblk_t goal; | |
c9de560d AT |
5255 | ext4_grpblk_t block; |
5256 | ||
5257 | /* we can't allocate > group size */ | |
5258 | len = ar->len; | |
5259 | ||
5260 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
5261 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
5262 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
5263 | |
5264 | /* start searching from the goal */ | |
5265 | goal = ar->goal; | |
5266 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5267 | goal >= ext4_blocks_count(es)) | |
5268 | goal = le32_to_cpu(es->s_first_data_block); | |
5269 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
5270 | ||
5271 | /* set up allocation goals */ | |
f5a44db5 | 5272 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 5273 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
5274 | ac->ac_sb = sb; |
5275 | ac->ac_inode = ar->inode; | |
53accfa9 | 5276 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
5277 | ac->ac_o_ex.fe_group = group; |
5278 | ac->ac_o_ex.fe_start = block; | |
5279 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 5280 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 5281 | ac->ac_flags = ar->flags; |
c9de560d | 5282 | |
3cb77bd2 | 5283 | /* we have to define context: we'll work with a file or |
c9de560d AT |
5284 | * locality group. this is a policy, actually */ |
5285 | ext4_mb_group_or_file(ac); | |
5286 | ||
d3df1453 | 5287 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
c9de560d AT |
5288 | "left: %u/%u, right %u/%u to %swritable\n", |
5289 | (unsigned) ar->len, (unsigned) ar->logical, | |
5290 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
5291 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
5292 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 5293 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
5294 | return 0; |
5295 | ||
5296 | } | |
5297 | ||
6be2ded1 AK |
5298 | static noinline_for_stack void |
5299 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
5300 | struct ext4_locality_group *lg, | |
5301 | int order, int total_entries) | |
5302 | { | |
5303 | ext4_group_t group = 0; | |
5304 | struct ext4_buddy e4b; | |
5305 | struct list_head discard_list; | |
5306 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 5307 | |
d3df1453 | 5308 | mb_debug(sb, "discard locality group preallocation\n"); |
6be2ded1 AK |
5309 | |
5310 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
5311 | |
5312 | spin_lock(&lg->lg_prealloc_lock); | |
5313 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
92e9c58c MB |
5314 | pa_inode_list, |
5315 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
5316 | spin_lock(&pa->pa_lock); |
5317 | if (atomic_read(&pa->pa_count)) { | |
5318 | /* | |
5319 | * This is the pa that we just used | |
5320 | * for block allocation. So don't | |
5321 | * free that | |
5322 | */ | |
5323 | spin_unlock(&pa->pa_lock); | |
5324 | continue; | |
5325 | } | |
5326 | if (pa->pa_deleted) { | |
5327 | spin_unlock(&pa->pa_lock); | |
5328 | continue; | |
5329 | } | |
5330 | /* only lg prealloc space */ | |
cc0fb9ad | 5331 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
5332 | |
5333 | /* seems this one can be freed ... */ | |
27bc446e | 5334 | ext4_mb_mark_pa_deleted(sb, pa); |
6be2ded1 AK |
5335 | spin_unlock(&pa->pa_lock); |
5336 | ||
5337 | list_del_rcu(&pa->pa_inode_list); | |
5338 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
5339 | ||
5340 | total_entries--; | |
5341 | if (total_entries <= 5) { | |
5342 | /* | |
5343 | * we want to keep only 5 entries | |
5344 | * allowing it to grow to 8. This | |
5345 | * mak sure we don't call discard | |
5346 | * soon for this list. | |
5347 | */ | |
5348 | break; | |
5349 | } | |
5350 | } | |
5351 | spin_unlock(&lg->lg_prealloc_lock); | |
5352 | ||
5353 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 5354 | int err; |
6be2ded1 | 5355 | |
bd86298e | 5356 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
5357 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5358 | GFP_NOFS|__GFP_NOFAIL); | |
5359 | if (err) { | |
54d3adbc TT |
5360 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5361 | err, group); | |
6be2ded1 AK |
5362 | continue; |
5363 | } | |
5364 | ext4_lock_group(sb, group); | |
5365 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5366 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
5367 | ext4_unlock_group(sb, group); |
5368 | ||
e39e07fd | 5369 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
5370 | list_del(&pa->u.pa_tmp_list); |
5371 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5372 | } | |
6be2ded1 AK |
5373 | } |
5374 | ||
5375 | /* | |
5376 | * We have incremented pa_count. So it cannot be freed at this | |
5377 | * point. Also we hold lg_mutex. So no parallel allocation is | |
5378 | * possible from this lg. That means pa_free cannot be updated. | |
5379 | * | |
5380 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
5381 | * which can cause the lg_prealloc_list to be updated. | |
5382 | */ | |
5383 | ||
5384 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
5385 | { | |
5386 | int order, added = 0, lg_prealloc_count = 1; | |
5387 | struct super_block *sb = ac->ac_sb; | |
5388 | struct ext4_locality_group *lg = ac->ac_lg; | |
5389 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
5390 | ||
5391 | order = fls(pa->pa_free) - 1; | |
5392 | if (order > PREALLOC_TB_SIZE - 1) | |
5393 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
5394 | order = PREALLOC_TB_SIZE - 1; | |
5395 | /* Add the prealloc space to lg */ | |
f1167009 | 5396 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 5397 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
92e9c58c MB |
5398 | pa_inode_list, |
5399 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
5400 | spin_lock(&tmp_pa->pa_lock); |
5401 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 5402 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
5403 | continue; |
5404 | } | |
5405 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
5406 | /* Add to the tail of the previous entry */ | |
5407 | list_add_tail_rcu(&pa->pa_inode_list, | |
5408 | &tmp_pa->pa_inode_list); | |
5409 | added = 1; | |
5410 | /* | |
5411 | * we want to count the total | |
5412 | * number of entries in the list | |
5413 | */ | |
5414 | } | |
5415 | spin_unlock(&tmp_pa->pa_lock); | |
5416 | lg_prealloc_count++; | |
5417 | } | |
5418 | if (!added) | |
5419 | list_add_tail_rcu(&pa->pa_inode_list, | |
5420 | &lg->lg_prealloc_list[order]); | |
f1167009 | 5421 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
5422 | |
5423 | /* Now trim the list to be not more than 8 elements */ | |
5424 | if (lg_prealloc_count > 8) { | |
5425 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 5426 | order, lg_prealloc_count); |
6be2ded1 AK |
5427 | return; |
5428 | } | |
5429 | return ; | |
5430 | } | |
5431 | ||
27bc446e | 5432 | /* |
5433 | * if per-inode prealloc list is too long, trim some PA | |
5434 | */ | |
5435 | static void ext4_mb_trim_inode_pa(struct inode *inode) | |
5436 | { | |
5437 | struct ext4_inode_info *ei = EXT4_I(inode); | |
5438 | struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); | |
5439 | int count, delta; | |
5440 | ||
5441 | count = atomic_read(&ei->i_prealloc_active); | |
5442 | delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1; | |
5443 | if (count > sbi->s_mb_max_inode_prealloc + delta) { | |
5444 | count -= sbi->s_mb_max_inode_prealloc; | |
5445 | ext4_discard_preallocations(inode, count); | |
5446 | } | |
5447 | } | |
5448 | ||
c9de560d AT |
5449 | /* |
5450 | * release all resource we used in allocation | |
5451 | */ | |
5452 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
5453 | { | |
27bc446e | 5454 | struct inode *inode = ac->ac_inode; |
5455 | struct ext4_inode_info *ei = EXT4_I(inode); | |
53accfa9 | 5456 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
5457 | struct ext4_prealloc_space *pa = ac->ac_pa; |
5458 | if (pa) { | |
cc0fb9ad | 5459 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 5460 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 5461 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
5462 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
5463 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
5464 | pa->pa_free -= ac->ac_b_ex.fe_len; |
5465 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
5466 | spin_unlock(&pa->pa_lock); | |
66d5e027 | 5467 | |
5468 | /* | |
5469 | * We want to add the pa to the right bucket. | |
5470 | * Remove it from the list and while adding | |
5471 | * make sure the list to which we are adding | |
5472 | * doesn't grow big. | |
5473 | */ | |
5474 | if (likely(pa->pa_free)) { | |
5475 | spin_lock(pa->pa_obj_lock); | |
5476 | list_del_rcu(&pa->pa_inode_list); | |
5477 | spin_unlock(pa->pa_obj_lock); | |
5478 | ext4_mb_add_n_trim(ac); | |
5479 | } | |
ba443916 | 5480 | } |
27bc446e | 5481 | |
5482 | if (pa->pa_type == MB_INODE_PA) { | |
5483 | /* | |
5484 | * treat per-inode prealloc list as a lru list, then try | |
5485 | * to trim the least recently used PA. | |
5486 | */ | |
5487 | spin_lock(pa->pa_obj_lock); | |
5488 | list_move(&pa->pa_inode_list, &ei->i_prealloc_list); | |
5489 | spin_unlock(pa->pa_obj_lock); | |
5490 | } | |
5491 | ||
ba443916 AK |
5492 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
5493 | } | |
c9de560d | 5494 | if (ac->ac_bitmap_page) |
09cbfeaf | 5495 | put_page(ac->ac_bitmap_page); |
c9de560d | 5496 | if (ac->ac_buddy_page) |
09cbfeaf | 5497 | put_page(ac->ac_buddy_page); |
c9de560d AT |
5498 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
5499 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
5500 | ext4_mb_collect_stats(ac); | |
27bc446e | 5501 | ext4_mb_trim_inode_pa(inode); |
c9de560d AT |
5502 | return 0; |
5503 | } | |
5504 | ||
5505 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
5506 | { | |
8df9675f | 5507 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d | 5508 | int ret; |
8c80fb31 CX |
5509 | int freed = 0, busy = 0; |
5510 | int retry = 0; | |
c9de560d | 5511 | |
9bffad1e | 5512 | trace_ext4_mb_discard_preallocations(sb, needed); |
8c80fb31 CX |
5513 | |
5514 | if (needed == 0) | |
5515 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; | |
5516 | repeat: | |
8df9675f | 5517 | for (i = 0; i < ngroups && needed > 0; i++) { |
8c80fb31 | 5518 | ret = ext4_mb_discard_group_preallocations(sb, i, &busy); |
c9de560d AT |
5519 | freed += ret; |
5520 | needed -= ret; | |
8c80fb31 CX |
5521 | cond_resched(); |
5522 | } | |
5523 | ||
5524 | if (needed > 0 && busy && ++retry < 3) { | |
5525 | busy = 0; | |
5526 | goto repeat; | |
c9de560d AT |
5527 | } |
5528 | ||
5529 | return freed; | |
5530 | } | |
5531 | ||
cf5e2ca6 | 5532 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
07b5b8e1 | 5533 | struct ext4_allocation_context *ac, u64 *seq) |
cf5e2ca6 RH |
5534 | { |
5535 | int freed; | |
07b5b8e1 RH |
5536 | u64 seq_retry = 0; |
5537 | bool ret = false; | |
cf5e2ca6 RH |
5538 | |
5539 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
07b5b8e1 RH |
5540 | if (freed) { |
5541 | ret = true; | |
5542 | goto out_dbg; | |
5543 | } | |
5544 | seq_retry = ext4_get_discard_pa_seq_sum(); | |
99377830 RH |
5545 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
5546 | ac->ac_flags |= EXT4_MB_STRICT_CHECK; | |
07b5b8e1 RH |
5547 | *seq = seq_retry; |
5548 | ret = true; | |
5549 | } | |
5550 | ||
5551 | out_dbg: | |
5552 | mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); | |
5553 | return ret; | |
cf5e2ca6 RH |
5554 | } |
5555 | ||
8016e29f HS |
5556 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, |
5557 | struct ext4_allocation_request *ar, int *errp); | |
5558 | ||
c9de560d AT |
5559 | /* |
5560 | * Main entry point into mballoc to allocate blocks | |
5561 | * it tries to use preallocation first, then falls back | |
5562 | * to usual allocation | |
5563 | */ | |
5564 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 5565 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 5566 | { |
256bdb49 | 5567 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
5568 | struct ext4_sb_info *sbi; |
5569 | struct super_block *sb; | |
5570 | ext4_fsblk_t block = 0; | |
60e58e0f | 5571 | unsigned int inquota = 0; |
53accfa9 | 5572 | unsigned int reserv_clstrs = 0; |
07b5b8e1 | 5573 | u64 seq; |
c9de560d | 5574 | |
b10a44c3 | 5575 | might_sleep(); |
c9de560d AT |
5576 | sb = ar->inode->i_sb; |
5577 | sbi = EXT4_SB(sb); | |
5578 | ||
9bffad1e | 5579 | trace_ext4_request_blocks(ar); |
8016e29f HS |
5580 | if (sbi->s_mount_state & EXT4_FC_REPLAY) |
5581 | return ext4_mb_new_blocks_simple(handle, ar, errp); | |
ba80b101 | 5582 | |
45dc63e7 | 5583 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 5584 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
5585 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
5586 | ||
e3cf5d5d | 5587 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
5588 | /* Without delayed allocation we need to verify |
5589 | * there is enough free blocks to do block allocation | |
5590 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 5591 | */ |
55f020db | 5592 | while (ar->len && |
e7d5f315 | 5593 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 5594 | |
030ba6bc | 5595 | /* let others to free the space */ |
bb8b20ed | 5596 | cond_resched(); |
030ba6bc AK |
5597 | ar->len = ar->len >> 1; |
5598 | } | |
5599 | if (!ar->len) { | |
bbc4ec77 | 5600 | ext4_mb_show_pa(sb); |
a30d542a AK |
5601 | *errp = -ENOSPC; |
5602 | return 0; | |
5603 | } | |
53accfa9 | 5604 | reserv_clstrs = ar->len; |
55f020db | 5605 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
5606 | dquot_alloc_block_nofail(ar->inode, |
5607 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
5608 | } else { |
5609 | while (ar->len && | |
53accfa9 TT |
5610 | dquot_alloc_block(ar->inode, |
5611 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
5612 | |
5613 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
5614 | ar->len--; | |
5615 | } | |
60e58e0f MC |
5616 | } |
5617 | inquota = ar->len; | |
5618 | if (ar->len == 0) { | |
5619 | *errp = -EDQUOT; | |
6c7a120a | 5620 | goto out; |
60e58e0f | 5621 | } |
07031431 | 5622 | } |
d2a17637 | 5623 | |
85556c9a | 5624 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 5625 | if (!ac) { |
363d4251 | 5626 | ar->len = 0; |
256bdb49 | 5627 | *errp = -ENOMEM; |
6c7a120a | 5628 | goto out; |
256bdb49 ES |
5629 | } |
5630 | ||
256bdb49 | 5631 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
5632 | if (*errp) { |
5633 | ar->len = 0; | |
6c7a120a | 5634 | goto out; |
c9de560d AT |
5635 | } |
5636 | ||
256bdb49 | 5637 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
81198536 | 5638 | seq = this_cpu_read(discard_pa_seq); |
256bdb49 | 5639 | if (!ext4_mb_use_preallocated(ac)) { |
256bdb49 ES |
5640 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
5641 | ext4_mb_normalize_request(ac, ar); | |
53f86b17 RH |
5642 | |
5643 | *errp = ext4_mb_pa_alloc(ac); | |
5644 | if (*errp) | |
5645 | goto errout; | |
c9de560d AT |
5646 | repeat: |
5647 | /* allocate space in core */ | |
6c7a120a | 5648 | *errp = ext4_mb_regular_allocator(ac); |
53f86b17 RH |
5649 | /* |
5650 | * pa allocated above is added to grp->bb_prealloc_list only | |
5651 | * when we were able to allocate some block i.e. when | |
5652 | * ac->ac_status == AC_STATUS_FOUND. | |
5653 | * And error from above mean ac->ac_status != AC_STATUS_FOUND | |
5654 | * So we have to free this pa here itself. | |
5655 | */ | |
2c00ef3e | 5656 | if (*errp) { |
53f86b17 | 5657 | ext4_mb_pa_free(ac); |
2c00ef3e AK |
5658 | ext4_discard_allocated_blocks(ac); |
5659 | goto errout; | |
5660 | } | |
53f86b17 RH |
5661 | if (ac->ac_status == AC_STATUS_FOUND && |
5662 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) | |
5663 | ext4_mb_pa_free(ac); | |
c9de560d | 5664 | } |
256bdb49 | 5665 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 5666 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 5667 | if (*errp) { |
b844167e | 5668 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
5669 | goto errout; |
5670 | } else { | |
519deca0 AK |
5671 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
5672 | ar->len = ac->ac_b_ex.fe_len; | |
5673 | } | |
c9de560d | 5674 | } else { |
07b5b8e1 | 5675 | if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) |
c9de560d | 5676 | goto repeat; |
53f86b17 RH |
5677 | /* |
5678 | * If block allocation fails then the pa allocated above | |
5679 | * needs to be freed here itself. | |
5680 | */ | |
5681 | ext4_mb_pa_free(ac); | |
c9de560d | 5682 | *errp = -ENOSPC; |
6c7a120a AK |
5683 | } |
5684 | ||
6d138ced | 5685 | errout: |
6c7a120a | 5686 | if (*errp) { |
256bdb49 | 5687 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 5688 | ar->len = 0; |
256bdb49 | 5689 | ext4_mb_show_ac(ac); |
c9de560d | 5690 | } |
256bdb49 | 5691 | ext4_mb_release_context(ac); |
6c7a120a AK |
5692 | out: |
5693 | if (ac) | |
5694 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 5695 | if (inquota && ar->len < inquota) |
53accfa9 | 5696 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 5697 | if (!ar->len) { |
e3cf5d5d | 5698 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 5699 | /* release all the reserved blocks if non delalloc */ |
57042651 | 5700 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 5701 | reserv_clstrs); |
0087d9fb | 5702 | } |
c9de560d | 5703 | |
9bffad1e | 5704 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 5705 | |
c9de560d AT |
5706 | return block; |
5707 | } | |
c9de560d | 5708 | |
c894058d AK |
5709 | /* |
5710 | * We can merge two free data extents only if the physical blocks | |
5711 | * are contiguous, AND the extents were freed by the same transaction, | |
5712 | * AND the blocks are associated with the same group. | |
5713 | */ | |
a0154344 DJ |
5714 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
5715 | struct ext4_free_data *entry, | |
5716 | struct ext4_free_data *new_entry, | |
5717 | struct rb_root *entry_rb_root) | |
c894058d | 5718 | { |
a0154344 DJ |
5719 | if ((entry->efd_tid != new_entry->efd_tid) || |
5720 | (entry->efd_group != new_entry->efd_group)) | |
5721 | return; | |
5722 | if (entry->efd_start_cluster + entry->efd_count == | |
5723 | new_entry->efd_start_cluster) { | |
5724 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
5725 | new_entry->efd_count += entry->efd_count; | |
5726 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
5727 | entry->efd_start_cluster) { | |
5728 | new_entry->efd_count += entry->efd_count; | |
5729 | } else | |
5730 | return; | |
5731 | spin_lock(&sbi->s_md_lock); | |
5732 | list_del(&entry->efd_list); | |
5733 | spin_unlock(&sbi->s_md_lock); | |
5734 | rb_erase(&entry->efd_node, entry_rb_root); | |
5735 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
5736 | } |
5737 | ||
4ddfef7b ES |
5738 | static noinline_for_stack int |
5739 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 5740 | struct ext4_free_data *new_entry) |
c9de560d | 5741 | { |
e29136f8 | 5742 | ext4_group_t group = e4b->bd_group; |
84130193 | 5743 | ext4_grpblk_t cluster; |
d08854f5 | 5744 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 5745 | struct ext4_free_data *entry; |
c9de560d AT |
5746 | struct ext4_group_info *db = e4b->bd_info; |
5747 | struct super_block *sb = e4b->bd_sb; | |
5748 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
5749 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
5750 | struct rb_node *parent = NULL, *new_node; | |
5751 | ||
0390131b | 5752 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
5753 | BUG_ON(e4b->bd_bitmap_page == NULL); |
5754 | BUG_ON(e4b->bd_buddy_page == NULL); | |
5755 | ||
18aadd47 BJ |
5756 | new_node = &new_entry->efd_node; |
5757 | cluster = new_entry->efd_start_cluster; | |
c894058d | 5758 | |
c894058d AK |
5759 | if (!*n) { |
5760 | /* first free block exent. We need to | |
5761 | protect buddy cache from being freed, | |
5762 | * otherwise we'll refresh it from | |
5763 | * on-disk bitmap and lose not-yet-available | |
5764 | * blocks */ | |
09cbfeaf KS |
5765 | get_page(e4b->bd_buddy_page); |
5766 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
5767 | } |
5768 | while (*n) { | |
5769 | parent = *n; | |
18aadd47 BJ |
5770 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
5771 | if (cluster < entry->efd_start_cluster) | |
c894058d | 5772 | n = &(*n)->rb_left; |
18aadd47 | 5773 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
5774 | n = &(*n)->rb_right; |
5775 | else { | |
e29136f8 | 5776 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
5777 | ext4_group_first_block_no(sb, group) + |
5778 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 5779 | "Block already on to-be-freed list"); |
cca41553 | 5780 | kmem_cache_free(ext4_free_data_cachep, new_entry); |
c894058d | 5781 | return 0; |
c9de560d | 5782 | } |
c894058d | 5783 | } |
c9de560d | 5784 | |
c894058d AK |
5785 | rb_link_node(new_node, parent, n); |
5786 | rb_insert_color(new_node, &db->bb_free_root); | |
5787 | ||
5788 | /* Now try to see the extent can be merged to left and right */ | |
5789 | node = rb_prev(new_node); | |
5790 | if (node) { | |
18aadd47 | 5791 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5792 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5793 | &(db->bb_free_root)); | |
c894058d | 5794 | } |
c9de560d | 5795 | |
c894058d AK |
5796 | node = rb_next(new_node); |
5797 | if (node) { | |
18aadd47 | 5798 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5799 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5800 | &(db->bb_free_root)); | |
c9de560d | 5801 | } |
a0154344 | 5802 | |
d08854f5 | 5803 | spin_lock(&sbi->s_md_lock); |
a0154344 | 5804 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
5805 | sbi->s_mb_free_pending += clusters; |
5806 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
5807 | return 0; |
5808 | } | |
5809 | ||
8016e29f HS |
5810 | /* |
5811 | * Simple allocator for Ext4 fast commit replay path. It searches for blocks | |
5812 | * linearly starting at the goal block and also excludes the blocks which | |
5813 | * are going to be in use after fast commit replay. | |
5814 | */ | |
5815 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, | |
5816 | struct ext4_allocation_request *ar, int *errp) | |
5817 | { | |
5818 | struct buffer_head *bitmap_bh; | |
5819 | struct super_block *sb = ar->inode->i_sb; | |
5820 | ext4_group_t group; | |
5821 | ext4_grpblk_t blkoff; | |
31a074a0 XY |
5822 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
5823 | ext4_grpblk_t i = 0; | |
8016e29f HS |
5824 | ext4_fsblk_t goal, block; |
5825 | struct ext4_super_block *es = EXT4_SB(sb)->s_es; | |
5826 | ||
5827 | goal = ar->goal; | |
5828 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5829 | goal >= ext4_blocks_count(es)) | |
5830 | goal = le32_to_cpu(es->s_first_data_block); | |
5831 | ||
5832 | ar->len = 0; | |
5833 | ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); | |
5834 | for (; group < ext4_get_groups_count(sb); group++) { | |
5835 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
5836 | if (IS_ERR(bitmap_bh)) { | |
5837 | *errp = PTR_ERR(bitmap_bh); | |
5838 | pr_warn("Failed to read block bitmap\n"); | |
5839 | return 0; | |
5840 | } | |
5841 | ||
5842 | ext4_get_group_no_and_offset(sb, | |
5843 | max(ext4_group_first_block_no(sb, group), goal), | |
5844 | NULL, &blkoff); | |
31a074a0 XY |
5845 | while (1) { |
5846 | i = mb_find_next_zero_bit(bitmap_bh->b_data, max, | |
8016e29f | 5847 | blkoff); |
31a074a0 XY |
5848 | if (i >= max) |
5849 | break; | |
5850 | if (ext4_fc_replay_check_excluded(sb, | |
5851 | ext4_group_first_block_no(sb, group) + i)) { | |
5852 | blkoff = i + 1; | |
5853 | } else | |
5854 | break; | |
5855 | } | |
8016e29f | 5856 | brelse(bitmap_bh); |
31a074a0 XY |
5857 | if (i < max) |
5858 | break; | |
8016e29f HS |
5859 | } |
5860 | ||
31a074a0 XY |
5861 | if (group >= ext4_get_groups_count(sb) || i >= max) { |
5862 | *errp = -ENOSPC; | |
8016e29f | 5863 | return 0; |
31a074a0 | 5864 | } |
8016e29f HS |
5865 | |
5866 | block = ext4_group_first_block_no(sb, group) + i; | |
5867 | ext4_mb_mark_bb(sb, block, 1, 1); | |
5868 | ar->len = 1; | |
5869 | ||
5870 | return block; | |
5871 | } | |
5872 | ||
5873 | static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, | |
5874 | unsigned long count) | |
5875 | { | |
5876 | struct buffer_head *bitmap_bh; | |
5877 | struct super_block *sb = inode->i_sb; | |
5878 | struct ext4_group_desc *gdp; | |
5879 | struct buffer_head *gdp_bh; | |
5880 | ext4_group_t group; | |
5881 | ext4_grpblk_t blkoff; | |
5882 | int already_freed = 0, err, i; | |
5883 | ||
5884 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
5885 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
5886 | if (IS_ERR(bitmap_bh)) { | |
5887 | err = PTR_ERR(bitmap_bh); | |
5888 | pr_warn("Failed to read block bitmap\n"); | |
5889 | return; | |
5890 | } | |
5891 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
5892 | if (!gdp) | |
5893 | return; | |
5894 | ||
5895 | for (i = 0; i < count; i++) { | |
5896 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data)) | |
5897 | already_freed++; | |
5898 | } | |
5899 | mb_clear_bits(bitmap_bh->b_data, blkoff, count); | |
5900 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
5901 | if (err) | |
5902 | return; | |
5903 | ext4_free_group_clusters_set( | |
5904 | sb, gdp, ext4_free_group_clusters(sb, gdp) + | |
5905 | count - already_freed); | |
5906 | ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh); | |
5907 | ext4_group_desc_csum_set(sb, group, gdp); | |
5908 | ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
5909 | sync_dirty_buffer(bitmap_bh); | |
5910 | sync_dirty_buffer(gdp_bh); | |
5911 | brelse(bitmap_bh); | |
5912 | } | |
5913 | ||
44338711 | 5914 | /** |
8ac3939d RH |
5915 | * ext4_mb_clear_bb() -- helper function for freeing blocks. |
5916 | * Used by ext4_free_blocks() | |
44338711 TT |
5917 | * @handle: handle for this transaction |
5918 | * @inode: inode | |
c60990b3 TT |
5919 | * @block: starting physical block to be freed |
5920 | * @count: number of blocks to be freed | |
5def1360 | 5921 | * @flags: flags used by ext4_free_blocks |
c9de560d | 5922 | */ |
8ac3939d RH |
5923 | static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, |
5924 | ext4_fsblk_t block, unsigned long count, | |
5925 | int flags) | |
c9de560d | 5926 | { |
26346ff6 | 5927 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 5928 | struct super_block *sb = inode->i_sb; |
c9de560d | 5929 | struct ext4_group_desc *gdp; |
498e5f24 | 5930 | unsigned int overflow; |
c9de560d AT |
5931 | ext4_grpblk_t bit; |
5932 | struct buffer_head *gd_bh; | |
5933 | ext4_group_t block_group; | |
5934 | struct ext4_sb_info *sbi; | |
5935 | struct ext4_buddy e4b; | |
84130193 | 5936 | unsigned int count_clusters; |
c9de560d AT |
5937 | int err = 0; |
5938 | int ret; | |
5939 | ||
8016e29f HS |
5940 | sbi = EXT4_SB(sb); |
5941 | ||
1e1c2b86 LC |
5942 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
5943 | !ext4_inode_block_valid(inode, block, count)) { | |
5944 | ext4_error(sb, "Freeing blocks in system zone - " | |
5945 | "Block = %llu, count = %lu", block, count); | |
5946 | /* err = 0. ext4_std_error should be a no op */ | |
5947 | goto error_return; | |
5948 | } | |
5949 | flags |= EXT4_FREE_BLOCKS_VALIDATED; | |
5950 | ||
c9de560d AT |
5951 | do_more: |
5952 | overflow = 0; | |
5953 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
5954 | ||
163a203d DW |
5955 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
5956 | ext4_get_group_info(sb, block_group)))) | |
5957 | return; | |
5958 | ||
c9de560d AT |
5959 | /* |
5960 | * Check to see if we are freeing blocks across a group | |
5961 | * boundary. | |
5962 | */ | |
84130193 TT |
5963 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
5964 | overflow = EXT4_C2B(sbi, bit) + count - | |
5965 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d | 5966 | count -= overflow; |
1e1c2b86 LC |
5967 | /* The range changed so it's no longer validated */ |
5968 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d | 5969 | } |
810da240 | 5970 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 5971 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5972 | if (IS_ERR(bitmap_bh)) { |
5973 | err = PTR_ERR(bitmap_bh); | |
5974 | bitmap_bh = NULL; | |
c9de560d | 5975 | goto error_return; |
ce89f46c | 5976 | } |
c9de560d | 5977 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
5978 | if (!gdp) { |
5979 | err = -EIO; | |
c9de560d | 5980 | goto error_return; |
ce89f46c | 5981 | } |
c9de560d | 5982 | |
1e1c2b86 LC |
5983 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
5984 | !ext4_inode_block_valid(inode, block, count)) { | |
12062ddd | 5985 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 5986 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
5987 | /* err = 0. ext4_std_error should be a no op */ |
5988 | goto error_return; | |
c9de560d AT |
5989 | } |
5990 | ||
5991 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
188c299e JK |
5992 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
5993 | EXT4_JTR_NONE); | |
c9de560d AT |
5994 | if (err) |
5995 | goto error_return; | |
5996 | ||
5997 | /* | |
5998 | * We are about to modify some metadata. Call the journal APIs | |
5999 | * to unshare ->b_data if a currently-committing transaction is | |
6000 | * using it | |
6001 | */ | |
6002 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 6003 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
c9de560d AT |
6004 | if (err) |
6005 | goto error_return; | |
c9de560d AT |
6006 | #ifdef AGGRESSIVE_CHECK |
6007 | { | |
6008 | int i; | |
84130193 | 6009 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
6010 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
6011 | } | |
6012 | #endif | |
84130193 | 6013 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 6014 | |
adb7ef60 KK |
6015 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
6016 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
6017 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
6018 | if (err) |
6019 | goto error_return; | |
e6362609 | 6020 | |
f96c450d DJ |
6021 | /* |
6022 | * We need to make sure we don't reuse the freed block until after the | |
6023 | * transaction is committed. We make an exception if the inode is to be | |
6024 | * written in writeback mode since writeback mode has weak data | |
6025 | * consistency guarantees. | |
6026 | */ | |
6027 | if (ext4_handle_valid(handle) && | |
6028 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
6029 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
6030 | struct ext4_free_data *new_entry; |
6031 | /* | |
7444a072 MH |
6032 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
6033 | * to fail. | |
7a2fcbf7 | 6034 | */ |
7444a072 MH |
6035 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
6036 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
6037 | new_entry->efd_start_cluster = bit; |
6038 | new_entry->efd_group = block_group; | |
6039 | new_entry->efd_count = count_clusters; | |
6040 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 6041 | |
7a2fcbf7 | 6042 | ext4_lock_group(sb, block_group); |
84130193 | 6043 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 6044 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 6045 | } else { |
7a2fcbf7 AK |
6046 | /* need to update group_info->bb_free and bitmap |
6047 | * with group lock held. generate_buddy look at | |
6048 | * them with group lock_held | |
6049 | */ | |
d71c1ae2 | 6050 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
6051 | err = ext4_issue_discard(sb, block_group, bit, count, |
6052 | NULL); | |
d71c1ae2 LC |
6053 | if (err && err != -EOPNOTSUPP) |
6054 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
a00b482b | 6055 | " group:%u block:%d count:%lu failed" |
d71c1ae2 LC |
6056 | " with %d", block_group, bit, count, |
6057 | err); | |
8f9ff189 LC |
6058 | } else |
6059 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 6060 | |
955ce5f5 | 6061 | ext4_lock_group(sb, block_group); |
84130193 TT |
6062 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
6063 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
6064 | } |
6065 | ||
021b65bb TT |
6066 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
6067 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 6068 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 6069 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 6070 | ext4_unlock_group(sb, block_group); |
c9de560d | 6071 | |
772cb7c8 JS |
6072 | if (sbi->s_log_groups_per_flex) { |
6073 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 6074 | atomic64_add(count_clusters, |
7c990728 SJS |
6075 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6076 | flex_group)->free_clusters); | |
772cb7c8 JS |
6077 | } |
6078 | ||
9fe67149 EW |
6079 | /* |
6080 | * on a bigalloc file system, defer the s_freeclusters_counter | |
6081 | * update to the caller (ext4_remove_space and friends) so they | |
6082 | * can determine if a cluster freed here should be rereserved | |
6083 | */ | |
6084 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
6085 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
6086 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
6087 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
6088 | count_clusters); | |
6089 | } | |
7d734532 JK |
6090 | |
6091 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 6092 | |
7a2fcbf7 AK |
6093 | /* We dirtied the bitmap block */ |
6094 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6095 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6096 | ||
c9de560d AT |
6097 | /* And the group descriptor block */ |
6098 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 6099 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
6100 | if (!err) |
6101 | err = ret; | |
6102 | ||
6103 | if (overflow && !err) { | |
6104 | block += count; | |
6105 | count = overflow; | |
6106 | put_bh(bitmap_bh); | |
1e1c2b86 LC |
6107 | /* The range changed so it's no longer validated */ |
6108 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d AT |
6109 | goto do_more; |
6110 | } | |
c9de560d AT |
6111 | error_return: |
6112 | brelse(bitmap_bh); | |
6113 | ext4_std_error(sb, err); | |
6114 | return; | |
6115 | } | |
7360d173 | 6116 | |
8ac3939d RH |
6117 | /** |
6118 | * ext4_free_blocks() -- Free given blocks and update quota | |
6119 | * @handle: handle for this transaction | |
6120 | * @inode: inode | |
6121 | * @bh: optional buffer of the block to be freed | |
6122 | * @block: starting physical block to be freed | |
6123 | * @count: number of blocks to be freed | |
6124 | * @flags: flags used by ext4_free_blocks | |
6125 | */ | |
6126 | void ext4_free_blocks(handle_t *handle, struct inode *inode, | |
6127 | struct buffer_head *bh, ext4_fsblk_t block, | |
6128 | unsigned long count, int flags) | |
6129 | { | |
6130 | struct super_block *sb = inode->i_sb; | |
6131 | unsigned int overflow; | |
6132 | struct ext4_sb_info *sbi; | |
6133 | ||
6134 | sbi = EXT4_SB(sb); | |
6135 | ||
6136 | if (sbi->s_mount_state & EXT4_FC_REPLAY) { | |
6137 | ext4_free_blocks_simple(inode, block, count); | |
6138 | return; | |
6139 | } | |
6140 | ||
6141 | might_sleep(); | |
6142 | if (bh) { | |
6143 | if (block) | |
6144 | BUG_ON(block != bh->b_blocknr); | |
6145 | else | |
6146 | block = bh->b_blocknr; | |
6147 | } | |
6148 | ||
6149 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && | |
6150 | !ext4_inode_block_valid(inode, block, count)) { | |
6151 | ext4_error(sb, "Freeing blocks not in datazone - " | |
6152 | "block = %llu, count = %lu", block, count); | |
6153 | return; | |
6154 | } | |
1e1c2b86 | 6155 | flags |= EXT4_FREE_BLOCKS_VALIDATED; |
8ac3939d RH |
6156 | |
6157 | ext4_debug("freeing block %llu\n", block); | |
6158 | trace_ext4_free_blocks(inode, block, count, flags); | |
6159 | ||
6160 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6161 | BUG_ON(count > 1); | |
6162 | ||
6163 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
6164 | inode, bh, block); | |
6165 | } | |
6166 | ||
6167 | /* | |
6168 | * If the extent to be freed does not begin on a cluster | |
6169 | * boundary, we need to deal with partial clusters at the | |
6170 | * beginning and end of the extent. Normally we will free | |
6171 | * blocks at the beginning or the end unless we are explicitly | |
6172 | * requested to avoid doing so. | |
6173 | */ | |
6174 | overflow = EXT4_PBLK_COFF(sbi, block); | |
6175 | if (overflow) { | |
6176 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
6177 | overflow = sbi->s_cluster_ratio - overflow; | |
6178 | block += overflow; | |
6179 | if (count > overflow) | |
6180 | count -= overflow; | |
6181 | else | |
6182 | return; | |
6183 | } else { | |
6184 | block -= overflow; | |
6185 | count += overflow; | |
6186 | } | |
1e1c2b86 LC |
6187 | /* The range changed so it's no longer validated */ |
6188 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6189 | } |
6190 | overflow = EXT4_LBLK_COFF(sbi, count); | |
6191 | if (overflow) { | |
6192 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
6193 | if (count > overflow) | |
6194 | count -= overflow; | |
6195 | else | |
6196 | return; | |
6197 | } else | |
6198 | count += sbi->s_cluster_ratio - overflow; | |
1e1c2b86 LC |
6199 | /* The range changed so it's no longer validated */ |
6200 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6201 | } |
6202 | ||
6203 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6204 | int i; | |
6205 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; | |
6206 | ||
6207 | for (i = 0; i < count; i++) { | |
6208 | cond_resched(); | |
6209 | if (is_metadata) | |
6210 | bh = sb_find_get_block(inode->i_sb, block + i); | |
6211 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
6212 | } | |
6213 | } | |
6214 | ||
6215 | ext4_mb_clear_bb(handle, inode, block, count, flags); | |
6216 | return; | |
6217 | } | |
6218 | ||
2846e820 | 6219 | /** |
0529155e | 6220 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
6221 | * @handle: handle to this transaction |
6222 | * @sb: super block | |
4907cb7b | 6223 | * @block: start physical block to add to the block group |
2846e820 AG |
6224 | * @count: number of blocks to free |
6225 | * | |
e73a347b | 6226 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 6227 | */ |
cc7365df | 6228 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
6229 | ext4_fsblk_t block, unsigned long count) |
6230 | { | |
6231 | struct buffer_head *bitmap_bh = NULL; | |
6232 | struct buffer_head *gd_bh; | |
6233 | ext4_group_t block_group; | |
6234 | ext4_grpblk_t bit; | |
6235 | unsigned int i; | |
6236 | struct ext4_group_desc *desc; | |
6237 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 6238 | struct ext4_buddy e4b; |
d77147ff | 6239 | int err = 0, ret, free_clusters_count; |
6240 | ext4_grpblk_t clusters_freed; | |
6241 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
6242 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
6243 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
6244 | |
6245 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
6246 | ||
4740b830 YY |
6247 | if (count == 0) |
6248 | return 0; | |
6249 | ||
2846e820 | 6250 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
6251 | /* |
6252 | * Check to see if we are freeing blocks across a group | |
6253 | * boundary. | |
6254 | */ | |
d77147ff | 6255 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
6256 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
6257 | block_group); |
6258 | err = -EINVAL; | |
2846e820 | 6259 | goto error_return; |
cc7365df | 6260 | } |
2cd05cc3 | 6261 | |
2846e820 | 6262 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
6263 | if (IS_ERR(bitmap_bh)) { |
6264 | err = PTR_ERR(bitmap_bh); | |
6265 | bitmap_bh = NULL; | |
2846e820 | 6266 | goto error_return; |
cc7365df YY |
6267 | } |
6268 | ||
2846e820 | 6269 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
6270 | if (!desc) { |
6271 | err = -EIO; | |
2846e820 | 6272 | goto error_return; |
cc7365df | 6273 | } |
2846e820 | 6274 | |
a00b482b | 6275 | if (!ext4_sb_block_valid(sb, NULL, block, count)) { |
2846e820 AG |
6276 | ext4_error(sb, "Adding blocks in system zones - " |
6277 | "Block = %llu, count = %lu", | |
6278 | block, count); | |
cc7365df | 6279 | err = -EINVAL; |
2846e820 AG |
6280 | goto error_return; |
6281 | } | |
6282 | ||
2cd05cc3 | 6283 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
6284 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
6285 | EXT4_JTR_NONE); | |
2846e820 AG |
6286 | if (err) |
6287 | goto error_return; | |
6288 | ||
6289 | /* | |
6290 | * We are about to modify some metadata. Call the journal APIs | |
6291 | * to unshare ->b_data if a currently-committing transaction is | |
6292 | * using it | |
6293 | */ | |
6294 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 6295 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
2846e820 AG |
6296 | if (err) |
6297 | goto error_return; | |
e73a347b | 6298 | |
d77147ff | 6299 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 6300 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 6301 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
6302 | ext4_error(sb, "bit already cleared for block %llu", |
6303 | (ext4_fsblk_t)(block + i)); | |
6304 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
6305 | } else { | |
d77147ff | 6306 | clusters_freed++; |
2846e820 AG |
6307 | } |
6308 | } | |
e73a347b AG |
6309 | |
6310 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
6311 | if (err) | |
6312 | goto error_return; | |
6313 | ||
6314 | /* | |
6315 | * need to update group_info->bb_free and bitmap | |
6316 | * with group lock held. generate_buddy look at | |
6317 | * them with group lock_held | |
6318 | */ | |
2846e820 | 6319 | ext4_lock_group(sb, block_group); |
d77147ff | 6320 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
6321 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
6322 | free_clusters_count = clusters_freed + | |
6323 | ext4_free_group_clusters(sb, desc); | |
6324 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
79f1ba49 | 6325 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 6326 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 6327 | ext4_unlock_group(sb, block_group); |
57042651 | 6328 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 6329 | clusters_freed); |
2846e820 AG |
6330 | |
6331 | if (sbi->s_log_groups_per_flex) { | |
6332 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 6333 | atomic64_add(clusters_freed, |
7c990728 SJS |
6334 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6335 | flex_group)->free_clusters); | |
2846e820 | 6336 | } |
e73a347b AG |
6337 | |
6338 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
6339 | |
6340 | /* We dirtied the bitmap block */ | |
6341 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6342 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6343 | ||
6344 | /* And the group descriptor block */ | |
6345 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
6346 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
6347 | if (!err) | |
6348 | err = ret; | |
6349 | ||
6350 | error_return: | |
6351 | brelse(bitmap_bh); | |
6352 | ext4_std_error(sb, err); | |
cc7365df | 6353 | return err; |
2846e820 AG |
6354 | } |
6355 | ||
7360d173 LC |
6356 | /** |
6357 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
6358 | * @sb: super block for the file system | |
6359 | * @start: starting block of the free extent in the alloc. group | |
6360 | * @count: number of blocks to TRIM | |
7360d173 LC |
6361 | * @e4b: ext4 buddy for the group |
6362 | * | |
6363 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
6364 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
6365 | * be called with under the group lock. | |
6366 | */ | |
bd2eea8d WJ |
6367 | static int ext4_trim_extent(struct super_block *sb, |
6368 | int start, int count, struct ext4_buddy *e4b) | |
e2cbd587 | 6369 | __releases(bitlock) |
6370 | __acquires(bitlock) | |
7360d173 LC |
6371 | { |
6372 | struct ext4_free_extent ex; | |
bd2eea8d | 6373 | ext4_group_t group = e4b->bd_group; |
d71c1ae2 | 6374 | int ret = 0; |
7360d173 | 6375 | |
b3d4c2b1 TM |
6376 | trace_ext4_trim_extent(sb, group, start, count); |
6377 | ||
7360d173 LC |
6378 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
6379 | ||
6380 | ex.fe_start = start; | |
6381 | ex.fe_group = group; | |
6382 | ex.fe_len = count; | |
6383 | ||
6384 | /* | |
6385 | * Mark blocks used, so no one can reuse them while | |
6386 | * being trimmed. | |
6387 | */ | |
6388 | mb_mark_used(e4b, &ex); | |
6389 | ext4_unlock_group(sb, group); | |
a0154344 | 6390 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
6391 | ext4_lock_group(sb, group); |
6392 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 6393 | return ret; |
7360d173 LC |
6394 | } |
6395 | ||
6920b391 WJ |
6396 | static int ext4_try_to_trim_range(struct super_block *sb, |
6397 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
6398 | ext4_grpblk_t max, ext4_grpblk_t minblocks) | |
a5fda113 TT |
6399 | __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) |
6400 | __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) | |
6920b391 WJ |
6401 | { |
6402 | ext4_grpblk_t next, count, free_count; | |
6403 | void *bitmap; | |
6920b391 WJ |
6404 | |
6405 | bitmap = e4b->bd_bitmap; | |
6406 | start = (e4b->bd_info->bb_first_free > start) ? | |
6407 | e4b->bd_info->bb_first_free : start; | |
6408 | count = 0; | |
6409 | free_count = 0; | |
6410 | ||
6411 | while (start <= max) { | |
6412 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
6413 | if (start > max) | |
6414 | break; | |
6415 | next = mb_find_next_bit(bitmap, max + 1, start); | |
6416 | ||
6417 | if ((next - start) >= minblocks) { | |
afcc4e32 LB |
6418 | int ret = ext4_trim_extent(sb, start, next - start, e4b); |
6419 | ||
6920b391 WJ |
6420 | if (ret && ret != -EOPNOTSUPP) |
6421 | break; | |
6920b391 WJ |
6422 | count += next - start; |
6423 | } | |
6424 | free_count += next - start; | |
6425 | start = next + 1; | |
6426 | ||
6427 | if (fatal_signal_pending(current)) { | |
6428 | count = -ERESTARTSYS; | |
6429 | break; | |
6430 | } | |
6431 | ||
6432 | if (need_resched()) { | |
6433 | ext4_unlock_group(sb, e4b->bd_group); | |
6434 | cond_resched(); | |
6435 | ext4_lock_group(sb, e4b->bd_group); | |
6436 | } | |
6437 | ||
6438 | if ((e4b->bd_info->bb_free - free_count) < minblocks) | |
6439 | break; | |
6440 | } | |
6441 | ||
6442 | return count; | |
6443 | } | |
6444 | ||
7360d173 LC |
6445 | /** |
6446 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
6447 | * @sb: super block for file system | |
22612283 | 6448 | * @group: group to be trimmed |
7360d173 LC |
6449 | * @start: first group block to examine |
6450 | * @max: last group block to examine | |
6451 | * @minblocks: minimum extent block count | |
d63c00ea | 6452 | * @set_trimmed: set the trimmed flag if at least one block is trimmed |
7360d173 | 6453 | * |
7360d173 LC |
6454 | * ext4_trim_all_free walks through group's block bitmap searching for free |
6455 | * extents. When the free extent is found, mark it as used in group buddy | |
6456 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
b6f5558c | 6457 | * the group buddy bitmap. |
7360d173 | 6458 | */ |
0b75a840 | 6459 | static ext4_grpblk_t |
78944086 LC |
6460 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
6461 | ext4_grpblk_t start, ext4_grpblk_t max, | |
d63c00ea | 6462 | ext4_grpblk_t minblocks, bool set_trimmed) |
7360d173 | 6463 | { |
78944086 | 6464 | struct ext4_buddy e4b; |
6920b391 | 6465 | int ret; |
7360d173 | 6466 | |
b3d4c2b1 TM |
6467 | trace_ext4_trim_all_free(sb, group, start, max); |
6468 | ||
78944086 LC |
6469 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
6470 | if (ret) { | |
9651e6b2 KK |
6471 | ext4_warning(sb, "Error %d loading buddy information for %u", |
6472 | ret, group); | |
78944086 LC |
6473 | return ret; |
6474 | } | |
28739eea LC |
6475 | |
6476 | ext4_lock_group(sb, group); | |
7360d173 | 6477 | |
6920b391 | 6478 | if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || |
2327fb2e | 6479 | minblocks < EXT4_SB(sb)->s_last_trim_minblks) { |
6920b391 | 6480 | ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); |
d63c00ea | 6481 | if (ret >= 0 && set_trimmed) |
6920b391 WJ |
6482 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
6483 | } else { | |
6484 | ret = 0; | |
7360d173 | 6485 | } |
3d56b8d2 | 6486 | |
7360d173 | 6487 | ext4_unlock_group(sb, group); |
78944086 | 6488 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
6489 | |
6490 | ext4_debug("trimmed %d blocks in the group %d\n", | |
6920b391 | 6491 | ret, group); |
7360d173 | 6492 | |
d71c1ae2 | 6493 | return ret; |
7360d173 LC |
6494 | } |
6495 | ||
6496 | /** | |
6497 | * ext4_trim_fs() -- trim ioctl handle function | |
6498 | * @sb: superblock for filesystem | |
6499 | * @range: fstrim_range structure | |
6500 | * | |
6501 | * start: First Byte to trim | |
6502 | * len: number of Bytes to trim from start | |
6503 | * minlen: minimum extent length in Bytes | |
6504 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
6505 | * start to start+len. For each such a group ext4_trim_all_free function | |
6506 | * is invoked to trim all free space. | |
6507 | */ | |
6508 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
6509 | { | |
7b47ef52 | 6510 | unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); |
78944086 | 6511 | struct ext4_group_info *grp; |
913eed83 | 6512 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 6513 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 6514 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
6515 | ext4_fsblk_t first_data_blk = |
6516 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 6517 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
d63c00ea | 6518 | bool whole_group, eof = false; |
7360d173 LC |
6519 | int ret = 0; |
6520 | ||
6521 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 6522 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
6523 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
6524 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 6525 | |
5de35e8d LC |
6526 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
6527 | start >= max_blks || | |
6528 | range->len < sb->s_blocksize) | |
7360d173 | 6529 | return -EINVAL; |
173b6e38 | 6530 | /* No point to try to trim less than discard granularity */ |
7b47ef52 | 6531 | if (range->minlen < discard_granularity) { |
173b6e38 | 6532 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
7b47ef52 | 6533 | discard_granularity >> sb->s_blocksize_bits); |
173b6e38 JK |
6534 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) |
6535 | goto out; | |
6536 | } | |
d63c00ea | 6537 | if (end >= max_blks - 1) { |
913eed83 | 6538 | end = max_blks - 1; |
d63c00ea DM |
6539 | eof = true; |
6540 | } | |
913eed83 | 6541 | if (end <= first_data_blk) |
22f10457 | 6542 | goto out; |
913eed83 | 6543 | if (start < first_data_blk) |
0f0a25bf | 6544 | start = first_data_blk; |
7360d173 | 6545 | |
913eed83 | 6546 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 6547 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 6548 | &first_group, &first_cluster); |
913eed83 | 6549 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 6550 | &last_group, &last_cluster); |
7360d173 | 6551 | |
913eed83 LC |
6552 | /* end now represents the last cluster to discard in this group */ |
6553 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
d63c00ea | 6554 | whole_group = true; |
7360d173 LC |
6555 | |
6556 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
6557 | grp = ext4_get_group_info(sb, group); |
6558 | /* We only do this if the grp has never been initialized */ | |
6559 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 6560 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
6561 | if (ret) |
6562 | break; | |
7360d173 LC |
6563 | } |
6564 | ||
0ba08517 | 6565 | /* |
913eed83 LC |
6566 | * For all the groups except the last one, last cluster will |
6567 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
6568 | * change it for the last group, note that last_cluster is | |
6569 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 6570 | */ |
d63c00ea | 6571 | if (group == last_group) { |
913eed83 | 6572 | end = last_cluster; |
d63c00ea DM |
6573 | whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1; |
6574 | } | |
78944086 | 6575 | if (grp->bb_free >= minlen) { |
7137d7a4 | 6576 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
d63c00ea | 6577 | end, minlen, whole_group); |
7360d173 LC |
6578 | if (cnt < 0) { |
6579 | ret = cnt; | |
7360d173 LC |
6580 | break; |
6581 | } | |
21e7fd22 | 6582 | trimmed += cnt; |
7360d173 | 6583 | } |
913eed83 LC |
6584 | |
6585 | /* | |
6586 | * For every group except the first one, we are sure | |
6587 | * that the first cluster to discard will be cluster #0. | |
6588 | */ | |
7137d7a4 | 6589 | first_cluster = 0; |
7360d173 | 6590 | } |
7360d173 | 6591 | |
3d56b8d2 | 6592 | if (!ret) |
2327fb2e | 6593 | EXT4_SB(sb)->s_last_trim_minblks = minlen; |
3d56b8d2 | 6594 | |
22f10457 | 6595 | out: |
aaf7d73e | 6596 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
6597 | return ret; |
6598 | } | |
0c9ec4be DW |
6599 | |
6600 | /* Iterate all the free extents in the group. */ | |
6601 | int | |
6602 | ext4_mballoc_query_range( | |
6603 | struct super_block *sb, | |
6604 | ext4_group_t group, | |
6605 | ext4_grpblk_t start, | |
6606 | ext4_grpblk_t end, | |
6607 | ext4_mballoc_query_range_fn formatter, | |
6608 | void *priv) | |
6609 | { | |
6610 | void *bitmap; | |
6611 | ext4_grpblk_t next; | |
6612 | struct ext4_buddy e4b; | |
6613 | int error; | |
6614 | ||
6615 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
6616 | if (error) | |
6617 | return error; | |
6618 | bitmap = e4b.bd_bitmap; | |
6619 | ||
6620 | ext4_lock_group(sb, group); | |
6621 | ||
6622 | start = (e4b.bd_info->bb_first_free > start) ? | |
6623 | e4b.bd_info->bb_first_free : start; | |
6624 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
6625 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
6626 | ||
6627 | while (start <= end) { | |
6628 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
6629 | if (start > end) | |
6630 | break; | |
6631 | next = mb_find_next_bit(bitmap, end + 1, start); | |
6632 | ||
6633 | ext4_unlock_group(sb, group); | |
6634 | error = formatter(sb, group, start, next - start, priv); | |
6635 | if (error) | |
6636 | goto out_unload; | |
6637 | ext4_lock_group(sb, group); | |
6638 | ||
6639 | start = next + 1; | |
6640 | } | |
6641 | ||
6642 | ext4_unlock_group(sb, group); | |
6643 | out_unload: | |
6644 | ext4_mb_unload_buddy(&e4b); | |
6645 | ||
6646 | return error; | |
6647 | } |