]>
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)) { | |
698 | /* only single bit in buddy2 may be 1 */ | |
699 | if (!mb_test_bit(i << 1, buddy2)) { | |
700 | MB_CHECK_ASSERT( | |
701 | mb_test_bit((i<<1)+1, buddy2)); | |
702 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
703 | MB_CHECK_ASSERT( | |
704 | mb_test_bit(i << 1, buddy2)); | |
705 | } | |
706 | continue; | |
707 | } | |
708 | ||
0a10da73 | 709 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
710 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
711 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
712 | ||
713 | for (j = 0; j < (1 << order); j++) { | |
714 | k = (i * (1 << order)) + j; | |
715 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 716 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
717 | } |
718 | count++; | |
719 | } | |
720 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
721 | order--; | |
722 | } | |
723 | ||
724 | fstart = -1; | |
725 | buddy = mb_find_buddy(e4b, 0, &max); | |
726 | for (i = 0; i < max; i++) { | |
727 | if (!mb_test_bit(i, buddy)) { | |
728 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
729 | if (fstart == -1) { | |
730 | fragments++; | |
731 | fstart = i; | |
732 | } | |
733 | continue; | |
734 | } | |
735 | fstart = -1; | |
736 | /* check used bits only */ | |
737 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
738 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
739 | k = i >> j; | |
740 | MB_CHECK_ASSERT(k < max2); | |
741 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
742 | } | |
743 | } | |
744 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
745 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
746 | ||
747 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
748 | list_for_each(cur, &grp->bb_prealloc_list) { |
749 | ext4_group_t groupnr; | |
750 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
751 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
752 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 753 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 754 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
755 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
756 | } | |
757 | return 0; | |
758 | } | |
759 | #undef MB_CHECK_ASSERT | |
760 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 761 | __FILE__, __func__, __LINE__) |
c9de560d AT |
762 | #else |
763 | #define mb_check_buddy(e4b) | |
764 | #endif | |
765 | ||
7c786059 CL |
766 | /* |
767 | * Divide blocks started from @first with length @len into | |
768 | * smaller chunks with power of 2 blocks. | |
769 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
770 | * then increase bb_counters[] for corresponded chunk size. | |
771 | */ | |
c9de560d | 772 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 773 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
774 | struct ext4_group_info *grp) |
775 | { | |
776 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
777 | ext4_grpblk_t min; |
778 | ext4_grpblk_t max; | |
779 | ext4_grpblk_t chunk; | |
69e43e8c | 780 | unsigned int border; |
c9de560d | 781 | |
7137d7a4 | 782 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
783 | |
784 | border = 2 << sb->s_blocksize_bits; | |
785 | ||
786 | while (len > 0) { | |
787 | /* find how many blocks can be covered since this position */ | |
788 | max = ffs(first | border) - 1; | |
789 | ||
790 | /* find how many blocks of power 2 we need to mark */ | |
791 | min = fls(len) - 1; | |
792 | ||
793 | if (max < min) | |
794 | min = max; | |
795 | chunk = 1 << min; | |
796 | ||
797 | /* mark multiblock chunks only */ | |
798 | grp->bb_counters[min]++; | |
799 | if (min > 0) | |
800 | mb_clear_bit(first >> min, | |
801 | buddy + sbi->s_mb_offsets[min]); | |
802 | ||
803 | len -= chunk; | |
804 | first += chunk; | |
805 | } | |
806 | } | |
807 | ||
196e402a HS |
808 | static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *new, |
809 | int (*cmp)(struct rb_node *, struct rb_node *)) | |
810 | { | |
811 | struct rb_node **iter = &root->rb_node, *parent = NULL; | |
812 | ||
813 | while (*iter) { | |
814 | parent = *iter; | |
815 | if (cmp(new, *iter) > 0) | |
816 | iter = &((*iter)->rb_left); | |
817 | else | |
818 | iter = &((*iter)->rb_right); | |
819 | } | |
820 | ||
821 | rb_link_node(new, parent, iter); | |
822 | rb_insert_color(new, root); | |
823 | } | |
824 | ||
825 | static int | |
826 | ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2) | |
827 | { | |
828 | struct ext4_group_info *grp1 = rb_entry(rb1, | |
829 | struct ext4_group_info, | |
830 | bb_avg_fragment_size_rb); | |
831 | struct ext4_group_info *grp2 = rb_entry(rb2, | |
832 | struct ext4_group_info, | |
833 | bb_avg_fragment_size_rb); | |
834 | int num_frags_1, num_frags_2; | |
835 | ||
836 | num_frags_1 = grp1->bb_fragments ? | |
837 | grp1->bb_free / grp1->bb_fragments : 0; | |
838 | num_frags_2 = grp2->bb_fragments ? | |
839 | grp2->bb_free / grp2->bb_fragments : 0; | |
840 | ||
841 | return (num_frags_2 - num_frags_1); | |
842 | } | |
843 | ||
844 | /* | |
845 | * Reinsert grpinfo into the avg_fragment_size tree with new average | |
846 | * fragment size. | |
847 | */ | |
848 | static void | |
849 | mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) | |
850 | { | |
851 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
852 | ||
853 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) | |
854 | return; | |
855 | ||
856 | write_lock(&sbi->s_mb_rb_lock); | |
857 | if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) { | |
858 | rb_erase(&grp->bb_avg_fragment_size_rb, | |
859 | &sbi->s_mb_avg_fragment_size_root); | |
860 | RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb); | |
861 | } | |
862 | ||
863 | ext4_mb_rb_insert(&sbi->s_mb_avg_fragment_size_root, | |
864 | &grp->bb_avg_fragment_size_rb, | |
865 | ext4_mb_avg_fragment_size_cmp); | |
866 | write_unlock(&sbi->s_mb_rb_lock); | |
867 | } | |
868 | ||
869 | /* | |
870 | * Choose next group by traversing largest_free_order lists. Updates *new_cr if | |
871 | * cr level needs an update. | |
872 | */ | |
873 | static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, | |
874 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
875 | { | |
876 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
877 | struct ext4_group_info *iter, *grp; | |
878 | int i; | |
879 | ||
880 | if (ac->ac_status == AC_STATUS_FOUND) | |
881 | return; | |
882 | ||
883 | if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED)) | |
884 | atomic_inc(&sbi->s_bal_cr0_bad_suggestions); | |
885 | ||
886 | grp = NULL; | |
887 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
888 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) | |
889 | continue; | |
890 | read_lock(&sbi->s_mb_largest_free_orders_locks[i]); | |
891 | if (list_empty(&sbi->s_mb_largest_free_orders[i])) { | |
892 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
893 | continue; | |
894 | } | |
895 | grp = NULL; | |
896 | list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], | |
897 | bb_largest_free_order_node) { | |
898 | if (sbi->s_mb_stats) | |
899 | atomic64_inc(&sbi->s_bal_cX_groups_considered[0]); | |
900 | if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) { | |
901 | grp = iter; | |
902 | break; | |
903 | } | |
904 | } | |
905 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
906 | if (grp) | |
907 | break; | |
908 | } | |
909 | ||
910 | if (!grp) { | |
911 | /* Increment cr and search again */ | |
912 | *new_cr = 1; | |
913 | } else { | |
914 | *group = grp->bb_group; | |
915 | ac->ac_last_optimal_group = *group; | |
916 | ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED; | |
917 | } | |
918 | } | |
919 | ||
920 | /* | |
921 | * Choose next group by traversing average fragment size tree. Updates *new_cr | |
922 | * if cr lvel needs an update. Sets EXT4_MB_SEARCH_NEXT_LINEAR to indicate that | |
923 | * the linear search should continue for one iteration since there's lock | |
924 | * contention on the rb tree lock. | |
925 | */ | |
926 | static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, | |
927 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
928 | { | |
929 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
930 | int avg_fragment_size, best_so_far; | |
931 | struct rb_node *node, *found; | |
932 | struct ext4_group_info *grp; | |
933 | ||
934 | /* | |
935 | * If there is contention on the lock, instead of waiting for the lock | |
936 | * to become available, just continue searching lineraly. We'll resume | |
937 | * our rb tree search later starting at ac->ac_last_optimal_group. | |
938 | */ | |
939 | if (!read_trylock(&sbi->s_mb_rb_lock)) { | |
940 | ac->ac_flags |= EXT4_MB_SEARCH_NEXT_LINEAR; | |
941 | return; | |
942 | } | |
943 | ||
944 | if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) { | |
945 | if (sbi->s_mb_stats) | |
946 | atomic_inc(&sbi->s_bal_cr1_bad_suggestions); | |
947 | /* We have found something at CR 1 in the past */ | |
948 | grp = ext4_get_group_info(ac->ac_sb, ac->ac_last_optimal_group); | |
949 | for (found = rb_next(&grp->bb_avg_fragment_size_rb); found != NULL; | |
950 | found = rb_next(found)) { | |
951 | grp = rb_entry(found, struct ext4_group_info, | |
952 | bb_avg_fragment_size_rb); | |
953 | if (sbi->s_mb_stats) | |
954 | atomic64_inc(&sbi->s_bal_cX_groups_considered[1]); | |
955 | if (likely(ext4_mb_good_group(ac, grp->bb_group, 1))) | |
956 | break; | |
957 | } | |
958 | goto done; | |
959 | } | |
960 | ||
961 | node = sbi->s_mb_avg_fragment_size_root.rb_node; | |
962 | best_so_far = 0; | |
963 | found = NULL; | |
964 | ||
965 | while (node) { | |
966 | grp = rb_entry(node, struct ext4_group_info, | |
967 | bb_avg_fragment_size_rb); | |
968 | avg_fragment_size = 0; | |
969 | if (ext4_mb_good_group(ac, grp->bb_group, 1)) { | |
970 | avg_fragment_size = grp->bb_fragments ? | |
971 | grp->bb_free / grp->bb_fragments : 0; | |
972 | if (!best_so_far || avg_fragment_size < best_so_far) { | |
973 | best_so_far = avg_fragment_size; | |
974 | found = node; | |
975 | } | |
976 | } | |
977 | if (avg_fragment_size > ac->ac_g_ex.fe_len) | |
978 | node = node->rb_right; | |
979 | else | |
980 | node = node->rb_left; | |
981 | } | |
982 | ||
983 | done: | |
984 | if (found) { | |
985 | grp = rb_entry(found, struct ext4_group_info, | |
986 | bb_avg_fragment_size_rb); | |
987 | *group = grp->bb_group; | |
988 | ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED; | |
989 | } else { | |
990 | *new_cr = 2; | |
991 | } | |
992 | ||
993 | read_unlock(&sbi->s_mb_rb_lock); | |
994 | ac->ac_last_optimal_group = *group; | |
995 | } | |
996 | ||
997 | static inline int should_optimize_scan(struct ext4_allocation_context *ac) | |
998 | { | |
999 | if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) | |
1000 | return 0; | |
1001 | if (ac->ac_criteria >= 2) | |
1002 | return 0; | |
1003 | if (ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) | |
1004 | return 0; | |
1005 | return 1; | |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * Return next linear group for allocation. If linear traversal should not be | |
1010 | * performed, this function just returns the same group | |
1011 | */ | |
1012 | static int | |
1013 | next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups) | |
1014 | { | |
1015 | if (!should_optimize_scan(ac)) | |
1016 | goto inc_and_return; | |
1017 | ||
1018 | if (ac->ac_groups_linear_remaining) { | |
1019 | ac->ac_groups_linear_remaining--; | |
1020 | goto inc_and_return; | |
1021 | } | |
1022 | ||
1023 | if (ac->ac_flags & EXT4_MB_SEARCH_NEXT_LINEAR) { | |
1024 | ac->ac_flags &= ~EXT4_MB_SEARCH_NEXT_LINEAR; | |
1025 | goto inc_and_return; | |
1026 | } | |
1027 | ||
1028 | return group; | |
1029 | inc_and_return: | |
1030 | /* | |
1031 | * Artificially restricted ngroups for non-extent | |
1032 | * files makes group > ngroups possible on first loop. | |
1033 | */ | |
1034 | return group + 1 >= ngroups ? 0 : group + 1; | |
1035 | } | |
1036 | ||
1037 | /* | |
1038 | * ext4_mb_choose_next_group: choose next group for allocation. | |
1039 | * | |
1040 | * @ac Allocation Context | |
1041 | * @new_cr This is an output parameter. If the there is no good group | |
1042 | * available at current CR level, this field is updated to indicate | |
1043 | * the new cr level that should be used. | |
1044 | * @group This is an input / output parameter. As an input it indicates the | |
1045 | * next group that the allocator intends to use for allocation. As | |
1046 | * output, this field indicates the next group that should be used as | |
1047 | * determined by the optimization functions. | |
1048 | * @ngroups Total number of groups | |
1049 | */ | |
1050 | static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, | |
1051 | int *new_cr, ext4_group_t *group, ext4_group_t ngroups) | |
1052 | { | |
1053 | *new_cr = ac->ac_criteria; | |
1054 | ||
1055 | if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) | |
1056 | return; | |
1057 | ||
1058 | if (*new_cr == 0) { | |
1059 | ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups); | |
1060 | } else if (*new_cr == 1) { | |
1061 | ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups); | |
1062 | } else { | |
1063 | /* | |
1064 | * TODO: For CR=2, we can arrange groups in an rb tree sorted by | |
1065 | * bb_free. But until that happens, we should never come here. | |
1066 | */ | |
1067 | WARN_ON(1); | |
1068 | } | |
1069 | } | |
1070 | ||
8a57d9d6 CW |
1071 | /* |
1072 | * Cache the order of the largest free extent we have available in this block | |
1073 | * group. | |
1074 | */ | |
1075 | static void | |
1076 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
1077 | { | |
196e402a | 1078 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
8a57d9d6 | 1079 | int i; |
8a57d9d6 | 1080 | |
196e402a HS |
1081 | if (test_opt2(sb, MB_OPTIMIZE_SCAN) && grp->bb_largest_free_order >= 0) { |
1082 | write_lock(&sbi->s_mb_largest_free_orders_locks[ | |
1083 | grp->bb_largest_free_order]); | |
1084 | list_del_init(&grp->bb_largest_free_order_node); | |
1085 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1086 | grp->bb_largest_free_order]); | |
1087 | } | |
8a57d9d6 CW |
1088 | grp->bb_largest_free_order = -1; /* uninit */ |
1089 | ||
196e402a | 1090 | for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) { |
8a57d9d6 CW |
1091 | if (grp->bb_counters[i] > 0) { |
1092 | grp->bb_largest_free_order = i; | |
1093 | break; | |
1094 | } | |
1095 | } | |
196e402a HS |
1096 | if (test_opt2(sb, MB_OPTIMIZE_SCAN) && |
1097 | grp->bb_largest_free_order >= 0 && grp->bb_free) { | |
1098 | write_lock(&sbi->s_mb_largest_free_orders_locks[ | |
1099 | grp->bb_largest_free_order]); | |
1100 | list_add_tail(&grp->bb_largest_free_order_node, | |
1101 | &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); | |
1102 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1103 | grp->bb_largest_free_order]); | |
1104 | } | |
8a57d9d6 CW |
1105 | } |
1106 | ||
089ceecc ES |
1107 | static noinline_for_stack |
1108 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
1109 | void *buddy, void *bitmap, ext4_group_t group) |
1110 | { | |
1111 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 1112 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 1113 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
1114 | ext4_grpblk_t i = 0; |
1115 | ext4_grpblk_t first; | |
1116 | ext4_grpblk_t len; | |
c9de560d AT |
1117 | unsigned free = 0; |
1118 | unsigned fragments = 0; | |
1119 | unsigned long long period = get_cycles(); | |
1120 | ||
1121 | /* initialize buddy from bitmap which is aggregation | |
1122 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 1123 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
1124 | grp->bb_first_free = i; |
1125 | while (i < max) { | |
1126 | fragments++; | |
1127 | first = i; | |
ffad0a44 | 1128 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
1129 | len = i - first; |
1130 | free += len; | |
1131 | if (len > 1) | |
1132 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
1133 | else | |
1134 | grp->bb_counters[0]++; | |
1135 | if (i < max) | |
ffad0a44 | 1136 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
1137 | } |
1138 | grp->bb_fragments = fragments; | |
1139 | ||
1140 | if (free != grp->bb_free) { | |
e29136f8 | 1141 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
1142 | "block bitmap and bg descriptor " |
1143 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 1144 | free, grp->bb_free); |
e56eb659 | 1145 | /* |
163a203d | 1146 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
1147 | * corrupt and update bb_free using bitmap value |
1148 | */ | |
c9de560d | 1149 | grp->bb_free = free; |
db79e6d1 WS |
1150 | ext4_mark_group_bitmap_corrupted(sb, group, |
1151 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 1152 | } |
8a57d9d6 | 1153 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
1154 | |
1155 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
1156 | ||
1157 | period = get_cycles() - period; | |
67d25186 HS |
1158 | atomic_inc(&sbi->s_mb_buddies_generated); |
1159 | atomic64_add(period, &sbi->s_mb_generation_time); | |
196e402a | 1160 | mb_update_avg_fragment_size(sb, grp); |
c9de560d AT |
1161 | } |
1162 | ||
1163 | /* The buddy information is attached the buddy cache inode | |
1164 | * for convenience. The information regarding each group | |
1165 | * is loaded via ext4_mb_load_buddy. The information involve | |
1166 | * block bitmap and buddy information. The information are | |
1167 | * stored in the inode as | |
1168 | * | |
1169 | * { page } | |
c3a326a6 | 1170 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
1171 | * |
1172 | * | |
1173 | * one block each for bitmap and buddy information. | |
1174 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 1175 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
1176 | * So it can have information regarding groups_per_page which |
1177 | * is blocks_per_page/2 | |
8a57d9d6 CW |
1178 | * |
1179 | * Locking note: This routine takes the block group lock of all groups | |
1180 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
1181 | */ |
1182 | ||
adb7ef60 | 1183 | static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) |
c9de560d | 1184 | { |
8df9675f | 1185 | ext4_group_t ngroups; |
c9de560d AT |
1186 | int blocksize; |
1187 | int blocks_per_page; | |
1188 | int groups_per_page; | |
1189 | int err = 0; | |
1190 | int i; | |
813e5727 | 1191 | ext4_group_t first_group, group; |
c9de560d AT |
1192 | int first_block; |
1193 | struct super_block *sb; | |
1194 | struct buffer_head *bhs; | |
fa77dcfa | 1195 | struct buffer_head **bh = NULL; |
c9de560d AT |
1196 | struct inode *inode; |
1197 | char *data; | |
1198 | char *bitmap; | |
9b8b7d35 | 1199 | struct ext4_group_info *grinfo; |
c9de560d | 1200 | |
c9de560d AT |
1201 | inode = page->mapping->host; |
1202 | sb = inode->i_sb; | |
8df9675f | 1203 | ngroups = ext4_get_groups_count(sb); |
93407472 | 1204 | blocksize = i_blocksize(inode); |
09cbfeaf | 1205 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d | 1206 | |
d3df1453 RH |
1207 | mb_debug(sb, "init page %lu\n", page->index); |
1208 | ||
c9de560d AT |
1209 | groups_per_page = blocks_per_page >> 1; |
1210 | if (groups_per_page == 0) | |
1211 | groups_per_page = 1; | |
1212 | ||
1213 | /* allocate buffer_heads to read bitmaps */ | |
1214 | if (groups_per_page > 1) { | |
c9de560d | 1215 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 1216 | bh = kzalloc(i, gfp); |
813e5727 TT |
1217 | if (bh == NULL) { |
1218 | err = -ENOMEM; | |
c9de560d | 1219 | goto out; |
813e5727 | 1220 | } |
c9de560d AT |
1221 | } else |
1222 | bh = &bhs; | |
1223 | ||
1224 | first_group = page->index * blocks_per_page / 2; | |
1225 | ||
1226 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
1227 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
1228 | if (group >= ngroups) | |
c9de560d AT |
1229 | break; |
1230 | ||
813e5727 | 1231 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
1232 | /* |
1233 | * If page is uptodate then we came here after online resize | |
1234 | * which added some new uninitialized group info structs, so | |
1235 | * we must skip all initialized uptodate buddies on the page, | |
1236 | * which may be currently in use by an allocating task. | |
1237 | */ | |
1238 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
1239 | bh[i] = NULL; | |
1240 | continue; | |
1241 | } | |
cfd73237 | 1242 | bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
9008a58e DW |
1243 | if (IS_ERR(bh[i])) { |
1244 | err = PTR_ERR(bh[i]); | |
1245 | bh[i] = NULL; | |
c9de560d | 1246 | goto out; |
2ccb5fb9 | 1247 | } |
d3df1453 | 1248 | mb_debug(sb, "read bitmap for group %u\n", group); |
c9de560d AT |
1249 | } |
1250 | ||
1251 | /* wait for I/O completion */ | |
813e5727 | 1252 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
1253 | int err2; |
1254 | ||
1255 | if (!bh[i]) | |
1256 | continue; | |
1257 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
1258 | if (!err) | |
1259 | err = err2; | |
813e5727 | 1260 | } |
c9de560d AT |
1261 | |
1262 | first_block = page->index * blocks_per_page; | |
1263 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 1264 | group = (first_block + i) >> 1; |
8df9675f | 1265 | if (group >= ngroups) |
c9de560d AT |
1266 | break; |
1267 | ||
9b8b7d35 AG |
1268 | if (!bh[group - first_group]) |
1269 | /* skip initialized uptodate buddy */ | |
1270 | continue; | |
1271 | ||
bbdc322f LC |
1272 | if (!buffer_verified(bh[group - first_group])) |
1273 | /* Skip faulty bitmaps */ | |
1274 | continue; | |
1275 | err = 0; | |
1276 | ||
c9de560d AT |
1277 | /* |
1278 | * data carry information regarding this | |
1279 | * particular group in the format specified | |
1280 | * above | |
1281 | * | |
1282 | */ | |
1283 | data = page_address(page) + (i * blocksize); | |
1284 | bitmap = bh[group - first_group]->b_data; | |
1285 | ||
1286 | /* | |
1287 | * We place the buddy block and bitmap block | |
1288 | * close together | |
1289 | */ | |
1290 | if ((first_block + i) & 1) { | |
1291 | /* this is block of buddy */ | |
1292 | BUG_ON(incore == NULL); | |
d3df1453 | 1293 | mb_debug(sb, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 1294 | group, page->index, i * blocksize); |
f307333e | 1295 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
1296 | grinfo = ext4_get_group_info(sb, group); |
1297 | grinfo->bb_fragments = 0; | |
1298 | memset(grinfo->bb_counters, 0, | |
1927805e | 1299 | sizeof(*grinfo->bb_counters) * |
4b68f6df | 1300 | (MB_NUM_ORDERS(sb))); |
c9de560d AT |
1301 | /* |
1302 | * incore got set to the group block bitmap below | |
1303 | */ | |
7a2fcbf7 | 1304 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
1305 | /* init the buddy */ |
1306 | memset(data, 0xff, blocksize); | |
c9de560d | 1307 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 1308 | ext4_unlock_group(sb, group); |
c9de560d AT |
1309 | incore = NULL; |
1310 | } else { | |
1311 | /* this is block of bitmap */ | |
1312 | BUG_ON(incore != NULL); | |
d3df1453 | 1313 | mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 1314 | group, page->index, i * blocksize); |
f307333e | 1315 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
1316 | |
1317 | /* see comments in ext4_mb_put_pa() */ | |
1318 | ext4_lock_group(sb, group); | |
1319 | memcpy(data, bitmap, blocksize); | |
1320 | ||
1321 | /* mark all preallocated blks used in in-core bitmap */ | |
1322 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 1323 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
1324 | ext4_unlock_group(sb, group); |
1325 | ||
1326 | /* set incore so that the buddy information can be | |
1327 | * generated using this | |
1328 | */ | |
1329 | incore = data; | |
1330 | } | |
1331 | } | |
1332 | SetPageUptodate(page); | |
1333 | ||
1334 | out: | |
1335 | if (bh) { | |
9b8b7d35 | 1336 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
1337 | brelse(bh[i]); |
1338 | if (bh != &bhs) | |
1339 | kfree(bh); | |
1340 | } | |
1341 | return err; | |
1342 | } | |
1343 | ||
eee4adc7 | 1344 | /* |
2de8807b AG |
1345 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
1346 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1347 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
1348 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 1349 | */ |
2de8807b | 1350 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1351 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1352 | { |
2de8807b AG |
1353 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1354 | int block, pnum, poff; | |
eee4adc7 | 1355 | int blocks_per_page; |
2de8807b AG |
1356 | struct page *page; |
1357 | ||
1358 | e4b->bd_buddy_page = NULL; | |
1359 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 | 1360 | |
09cbfeaf | 1361 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1362 | /* |
1363 | * the buddy cache inode stores the block bitmap | |
1364 | * and buddy information in consecutive blocks. | |
1365 | * So for each group we need two blocks. | |
1366 | */ | |
1367 | block = group * 2; | |
1368 | pnum = block / blocks_per_page; | |
2de8807b | 1369 | poff = block % blocks_per_page; |
adb7ef60 | 1370 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1371 | if (!page) |
c57ab39b | 1372 | return -ENOMEM; |
2de8807b AG |
1373 | BUG_ON(page->mapping != inode->i_mapping); |
1374 | e4b->bd_bitmap_page = page; | |
1375 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1376 | ||
1377 | if (blocks_per_page >= 2) { | |
1378 | /* buddy and bitmap are on the same page */ | |
1379 | return 0; | |
eee4adc7 | 1380 | } |
2de8807b AG |
1381 | |
1382 | block++; | |
1383 | pnum = block / blocks_per_page; | |
adb7ef60 | 1384 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
2de8807b | 1385 | if (!page) |
c57ab39b | 1386 | return -ENOMEM; |
2de8807b AG |
1387 | BUG_ON(page->mapping != inode->i_mapping); |
1388 | e4b->bd_buddy_page = page; | |
1389 | return 0; | |
eee4adc7 ES |
1390 | } |
1391 | ||
2de8807b | 1392 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1393 | { |
2de8807b AG |
1394 | if (e4b->bd_bitmap_page) { |
1395 | unlock_page(e4b->bd_bitmap_page); | |
09cbfeaf | 1396 | put_page(e4b->bd_bitmap_page); |
2de8807b AG |
1397 | } |
1398 | if (e4b->bd_buddy_page) { | |
1399 | unlock_page(e4b->bd_buddy_page); | |
09cbfeaf | 1400 | put_page(e4b->bd_buddy_page); |
eee4adc7 | 1401 | } |
eee4adc7 ES |
1402 | } |
1403 | ||
8a57d9d6 CW |
1404 | /* |
1405 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1406 | * block group lock of all groups for this page; do not hold the BG lock when | |
1407 | * calling this routine! | |
1408 | */ | |
b6a758ec | 1409 | static noinline_for_stack |
adb7ef60 | 1410 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1411 | { |
1412 | ||
b6a758ec | 1413 | struct ext4_group_info *this_grp; |
2de8807b AG |
1414 | struct ext4_buddy e4b; |
1415 | struct page *page; | |
1416 | int ret = 0; | |
b6a758ec | 1417 | |
b10a44c3 | 1418 | might_sleep(); |
d3df1453 | 1419 | mb_debug(sb, "init group %u\n", group); |
b6a758ec AK |
1420 | this_grp = ext4_get_group_info(sb, group); |
1421 | /* | |
08c3a813 AK |
1422 | * This ensures that we don't reinit the buddy cache |
1423 | * page which map to the group from which we are already | |
1424 | * allocating. If we are looking at the buddy cache we would | |
1425 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1426 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1427 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1428 | * page accessed. | |
b6a758ec | 1429 | */ |
adb7ef60 | 1430 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1431 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1432 | /* |
1433 | * somebody initialized the group | |
1434 | * return without doing anything | |
1435 | */ | |
b6a758ec AK |
1436 | goto err; |
1437 | } | |
2de8807b AG |
1438 | |
1439 | page = e4b.bd_bitmap_page; | |
adb7ef60 | 1440 | ret = ext4_mb_init_cache(page, NULL, gfp); |
2de8807b AG |
1441 | if (ret) |
1442 | goto err; | |
1443 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1444 | ret = -EIO; |
1445 | goto err; | |
1446 | } | |
b6a758ec | 1447 | |
2de8807b | 1448 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1449 | /* |
1450 | * If both the bitmap and buddy are in | |
1451 | * the same page we don't need to force | |
1452 | * init the buddy | |
1453 | */ | |
2de8807b AG |
1454 | ret = 0; |
1455 | goto err; | |
b6a758ec | 1456 | } |
2de8807b AG |
1457 | /* init buddy cache */ |
1458 | page = e4b.bd_buddy_page; | |
adb7ef60 | 1459 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); |
2de8807b AG |
1460 | if (ret) |
1461 | goto err; | |
1462 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1463 | ret = -EIO; |
1464 | goto err; | |
1465 | } | |
b6a758ec | 1466 | err: |
2de8807b | 1467 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1468 | return ret; |
1469 | } | |
1470 | ||
8a57d9d6 CW |
1471 | /* |
1472 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1473 | * block group lock of all groups for this page; do not hold the BG lock when | |
1474 | * calling this routine! | |
1475 | */ | |
4ddfef7b | 1476 | static noinline_for_stack int |
adb7ef60 KK |
1477 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1478 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1479 | { |
c9de560d AT |
1480 | int blocks_per_page; |
1481 | int block; | |
1482 | int pnum; | |
1483 | int poff; | |
1484 | struct page *page; | |
fdf6c7a7 | 1485 | int ret; |
920313a7 AK |
1486 | struct ext4_group_info *grp; |
1487 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1488 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1489 | |
b10a44c3 | 1490 | might_sleep(); |
d3df1453 | 1491 | mb_debug(sb, "load group %u\n", group); |
c9de560d | 1492 | |
09cbfeaf | 1493 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1494 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1495 | |
1496 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1497 | e4b->bd_info = grp; |
c9de560d AT |
1498 | e4b->bd_sb = sb; |
1499 | e4b->bd_group = group; | |
1500 | e4b->bd_buddy_page = NULL; | |
1501 | e4b->bd_bitmap_page = NULL; | |
1502 | ||
f41c0750 | 1503 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1504 | /* |
1505 | * we need full data about the group | |
1506 | * to make a good selection | |
1507 | */ | |
adb7ef60 | 1508 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1509 | if (ret) |
1510 | return ret; | |
f41c0750 AK |
1511 | } |
1512 | ||
c9de560d AT |
1513 | /* |
1514 | * the buddy cache inode stores the block bitmap | |
1515 | * and buddy information in consecutive blocks. | |
1516 | * So for each group we need two blocks. | |
1517 | */ | |
1518 | block = group * 2; | |
1519 | pnum = block / blocks_per_page; | |
1520 | poff = block % blocks_per_page; | |
1521 | ||
1522 | /* we could use find_or_create_page(), but it locks page | |
1523 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1524 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1525 | if (page == NULL || !PageUptodate(page)) { |
1526 | if (page) | |
920313a7 AK |
1527 | /* |
1528 | * drop the page reference and try | |
1529 | * to get the page with lock. If we | |
1530 | * are not uptodate that implies | |
1531 | * somebody just created the page but | |
1532 | * is yet to initialize the same. So | |
1533 | * wait for it to initialize. | |
1534 | */ | |
09cbfeaf | 1535 | put_page(page); |
adb7ef60 | 1536 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1537 | if (page) { |
1538 | BUG_ON(page->mapping != inode->i_mapping); | |
1539 | if (!PageUptodate(page)) { | |
adb7ef60 | 1540 | ret = ext4_mb_init_cache(page, NULL, gfp); |
fdf6c7a7 SF |
1541 | if (ret) { |
1542 | unlock_page(page); | |
1543 | goto err; | |
1544 | } | |
c9de560d AT |
1545 | mb_cmp_bitmaps(e4b, page_address(page) + |
1546 | (poff * sb->s_blocksize)); | |
1547 | } | |
1548 | unlock_page(page); | |
1549 | } | |
1550 | } | |
c57ab39b YL |
1551 | if (page == NULL) { |
1552 | ret = -ENOMEM; | |
1553 | goto err; | |
1554 | } | |
1555 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1556 | ret = -EIO; |
c9de560d | 1557 | goto err; |
fdf6c7a7 | 1558 | } |
2457aec6 MG |
1559 | |
1560 | /* Pages marked accessed already */ | |
c9de560d AT |
1561 | e4b->bd_bitmap_page = page; |
1562 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1563 | |
1564 | block++; | |
1565 | pnum = block / blocks_per_page; | |
1566 | poff = block % blocks_per_page; | |
1567 | ||
2457aec6 | 1568 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1569 | if (page == NULL || !PageUptodate(page)) { |
1570 | if (page) | |
09cbfeaf | 1571 | put_page(page); |
adb7ef60 | 1572 | page = find_or_create_page(inode->i_mapping, pnum, gfp); |
c9de560d AT |
1573 | if (page) { |
1574 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 | 1575 | if (!PageUptodate(page)) { |
adb7ef60 KK |
1576 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap, |
1577 | gfp); | |
fdf6c7a7 SF |
1578 | if (ret) { |
1579 | unlock_page(page); | |
1580 | goto err; | |
1581 | } | |
1582 | } | |
c9de560d AT |
1583 | unlock_page(page); |
1584 | } | |
1585 | } | |
c57ab39b YL |
1586 | if (page == NULL) { |
1587 | ret = -ENOMEM; | |
1588 | goto err; | |
1589 | } | |
1590 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1591 | ret = -EIO; |
c9de560d | 1592 | goto err; |
fdf6c7a7 | 1593 | } |
2457aec6 MG |
1594 | |
1595 | /* Pages marked accessed already */ | |
c9de560d AT |
1596 | e4b->bd_buddy_page = page; |
1597 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d | 1598 | |
c9de560d AT |
1599 | return 0; |
1600 | ||
1601 | err: | |
26626f11 | 1602 | if (page) |
09cbfeaf | 1603 | put_page(page); |
c9de560d | 1604 | if (e4b->bd_bitmap_page) |
09cbfeaf | 1605 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1606 | if (e4b->bd_buddy_page) |
09cbfeaf | 1607 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1608 | e4b->bd_buddy = NULL; |
1609 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1610 | return ret; |
c9de560d AT |
1611 | } |
1612 | ||
adb7ef60 KK |
1613 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1614 | struct ext4_buddy *e4b) | |
1615 | { | |
1616 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1617 | } | |
1618 | ||
e39e07fd | 1619 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1620 | { |
1621 | if (e4b->bd_bitmap_page) | |
09cbfeaf | 1622 | put_page(e4b->bd_bitmap_page); |
c9de560d | 1623 | if (e4b->bd_buddy_page) |
09cbfeaf | 1624 | put_page(e4b->bd_buddy_page); |
c9de560d AT |
1625 | } |
1626 | ||
1627 | ||
1628 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1629 | { | |
ce3cca33 | 1630 | int order = 1, max; |
c9de560d AT |
1631 | void *bb; |
1632 | ||
c5e8f3f3 | 1633 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1634 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1635 | ||
c9de560d | 1636 | while (order <= e4b->bd_blkbits + 1) { |
ce3cca33 CX |
1637 | bb = mb_find_buddy(e4b, order, &max); |
1638 | if (!mb_test_bit(block >> order, bb)) { | |
c9de560d AT |
1639 | /* this block is part of buddy of order 'order' */ |
1640 | return order; | |
1641 | } | |
c9de560d AT |
1642 | order++; |
1643 | } | |
1644 | return 0; | |
1645 | } | |
1646 | ||
955ce5f5 | 1647 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1648 | { |
1649 | __u32 *addr; | |
1650 | ||
1651 | len = cur + len; | |
1652 | while (cur < len) { | |
1653 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1654 | /* fast path: clear whole word at once */ | |
1655 | addr = bm + (cur >> 3); | |
1656 | *addr = 0; | |
1657 | cur += 32; | |
1658 | continue; | |
1659 | } | |
955ce5f5 | 1660 | mb_clear_bit(cur, bm); |
c9de560d AT |
1661 | cur++; |
1662 | } | |
1663 | } | |
1664 | ||
eabe0444 AS |
1665 | /* clear bits in given range |
1666 | * will return first found zero bit if any, -1 otherwise | |
1667 | */ | |
1668 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1669 | { | |
1670 | __u32 *addr; | |
1671 | int zero_bit = -1; | |
1672 | ||
1673 | len = cur + len; | |
1674 | while (cur < len) { | |
1675 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1676 | /* fast path: clear whole word at once */ | |
1677 | addr = bm + (cur >> 3); | |
1678 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1679 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1680 | *addr = 0; | |
1681 | cur += 32; | |
1682 | continue; | |
1683 | } | |
1684 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1685 | zero_bit = cur; | |
1686 | cur++; | |
1687 | } | |
1688 | ||
1689 | return zero_bit; | |
1690 | } | |
1691 | ||
c3e94d1d | 1692 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1693 | { |
1694 | __u32 *addr; | |
1695 | ||
1696 | len = cur + len; | |
1697 | while (cur < len) { | |
1698 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1699 | /* fast path: set whole word at once */ | |
1700 | addr = bm + (cur >> 3); | |
1701 | *addr = 0xffffffff; | |
1702 | cur += 32; | |
1703 | continue; | |
1704 | } | |
955ce5f5 | 1705 | mb_set_bit(cur, bm); |
c9de560d AT |
1706 | cur++; |
1707 | } | |
1708 | } | |
1709 | ||
eabe0444 AS |
1710 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
1711 | { | |
1712 | if (mb_test_bit(*bit + side, bitmap)) { | |
1713 | mb_clear_bit(*bit, bitmap); | |
1714 | (*bit) -= side; | |
1715 | return 1; | |
1716 | } | |
1717 | else { | |
1718 | (*bit) += side; | |
1719 | mb_set_bit(*bit, bitmap); | |
1720 | return -1; | |
1721 | } | |
1722 | } | |
1723 | ||
1724 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1725 | { | |
1726 | int max; | |
1727 | int order = 1; | |
1728 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1729 | ||
1730 | while (buddy) { | |
1731 | void *buddy2; | |
1732 | ||
1733 | /* Bits in range [first; last] are known to be set since | |
1734 | * corresponding blocks were allocated. Bits in range | |
1735 | * (first; last) will stay set because they form buddies on | |
1736 | * upper layer. We just deal with borders if they don't | |
1737 | * align with upper layer and then go up. | |
1738 | * Releasing entire group is all about clearing | |
1739 | * single bit of highest order buddy. | |
1740 | */ | |
1741 | ||
1742 | /* Example: | |
1743 | * --------------------------------- | |
1744 | * | 1 | 1 | 1 | 1 | | |
1745 | * --------------------------------- | |
1746 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1747 | * --------------------------------- | |
1748 | * 0 1 2 3 4 5 6 7 | |
1749 | * \_____________________/ | |
1750 | * | |
1751 | * Neither [1] nor [6] is aligned to above layer. | |
1752 | * Left neighbour [0] is free, so mark it busy, | |
1753 | * decrease bb_counters and extend range to | |
1754 | * [0; 6] | |
1755 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1756 | * mark [6] free, increase bb_counters and shrink range to | |
1757 | * [0; 5]. | |
1758 | * Then shift range to [0; 2], go up and do the same. | |
1759 | */ | |
1760 | ||
1761 | ||
1762 | if (first & 1) | |
1763 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1764 | if (!(last & 1)) | |
1765 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1766 | if (first > last) | |
1767 | break; | |
1768 | order++; | |
1769 | ||
1770 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1771 | mb_clear_bits(buddy, first, last - first + 1); | |
1772 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1773 | break; | |
1774 | } | |
1775 | first >>= 1; | |
1776 | last >>= 1; | |
1777 | buddy = buddy2; | |
1778 | } | |
1779 | } | |
1780 | ||
7e5a8cdd | 1781 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1782 | int first, int count) |
c9de560d | 1783 | { |
eabe0444 AS |
1784 | int left_is_free = 0; |
1785 | int right_is_free = 0; | |
1786 | int block; | |
1787 | int last = first + count - 1; | |
c9de560d AT |
1788 | struct super_block *sb = e4b->bd_sb; |
1789 | ||
c99d1e6e TT |
1790 | if (WARN_ON(count == 0)) |
1791 | return; | |
eabe0444 | 1792 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1793 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1794 | /* Don't bother if the block group is corrupt. */ |
1795 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1796 | return; | |
1797 | ||
c9de560d AT |
1798 | mb_check_buddy(e4b); |
1799 | mb_free_blocks_double(inode, e4b, first, count); | |
1800 | ||
07b5b8e1 | 1801 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1802 | e4b->bd_info->bb_free += count; |
1803 | if (first < e4b->bd_info->bb_first_free) | |
1804 | e4b->bd_info->bb_first_free = first; | |
1805 | ||
eabe0444 AS |
1806 | /* access memory sequentially: check left neighbour, |
1807 | * clear range and then check right neighbour | |
1808 | */ | |
c9de560d | 1809 | if (first != 0) |
eabe0444 AS |
1810 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1811 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1812 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1813 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1814 | ||
1815 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1816 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1817 | ext4_fsblk_t blocknr; |
1818 | ||
1819 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
49598e04 | 1820 | blocknr += EXT4_C2B(sbi, block); |
8016e29f HS |
1821 | if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { |
1822 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1823 | inode ? inode->i_ino : 0, | |
1824 | blocknr, | |
1825 | "freeing already freed block (bit %u); block bitmap corrupt.", | |
1826 | block); | |
1827 | ext4_mark_group_bitmap_corrupted( | |
1828 | sb, e4b->bd_group, | |
db79e6d1 | 1829 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
8016e29f | 1830 | } |
eabe0444 AS |
1831 | goto done; |
1832 | } | |
1833 | ||
1834 | /* let's maintain fragments counter */ | |
1835 | if (left_is_free && right_is_free) | |
c9de560d | 1836 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1837 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1838 | e4b->bd_info->bb_fragments++; |
1839 | ||
eabe0444 AS |
1840 | /* buddy[0] == bd_bitmap is a special case, so handle |
1841 | * it right away and let mb_buddy_mark_free stay free of | |
1842 | * zero order checks. | |
1843 | * Check if neighbours are to be coaleasced, | |
1844 | * adjust bitmap bb_counters and borders appropriately. | |
1845 | */ | |
1846 | if (first & 1) { | |
1847 | first += !left_is_free; | |
1848 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1849 | } | |
1850 | if (!(last & 1)) { | |
1851 | last -= !right_is_free; | |
1852 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1853 | } | |
c9de560d | 1854 | |
eabe0444 AS |
1855 | if (first <= last) |
1856 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1857 | |
eabe0444 | 1858 | done: |
8a57d9d6 | 1859 | mb_set_largest_free_order(sb, e4b->bd_info); |
196e402a | 1860 | mb_update_avg_fragment_size(sb, e4b->bd_info); |
c9de560d | 1861 | mb_check_buddy(e4b); |
c9de560d AT |
1862 | } |
1863 | ||
15c006a2 | 1864 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1865 | int needed, struct ext4_free_extent *ex) |
1866 | { | |
1867 | int next = block; | |
15c006a2 | 1868 | int max, order; |
c9de560d AT |
1869 | void *buddy; |
1870 | ||
bc8e6740 | 1871 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1872 | BUG_ON(ex == NULL); |
1873 | ||
15c006a2 | 1874 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1875 | BUG_ON(buddy == NULL); |
1876 | BUG_ON(block >= max); | |
1877 | if (mb_test_bit(block, buddy)) { | |
1878 | ex->fe_len = 0; | |
1879 | ex->fe_start = 0; | |
1880 | ex->fe_group = 0; | |
1881 | return 0; | |
1882 | } | |
1883 | ||
15c006a2 RD |
1884 | /* find actual order */ |
1885 | order = mb_find_order_for_block(e4b, block); | |
1886 | block = block >> order; | |
c9de560d AT |
1887 | |
1888 | ex->fe_len = 1 << order; | |
1889 | ex->fe_start = block << order; | |
1890 | ex->fe_group = e4b->bd_group; | |
1891 | ||
1892 | /* calc difference from given start */ | |
1893 | next = next - ex->fe_start; | |
1894 | ex->fe_len -= next; | |
1895 | ex->fe_start += next; | |
1896 | ||
1897 | while (needed > ex->fe_len && | |
d8ec0c39 | 1898 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1899 | |
1900 | if (block + 1 >= max) | |
1901 | break; | |
1902 | ||
1903 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1904 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1905 | break; |
1906 | ||
b051d8dc | 1907 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1908 | |
c9de560d AT |
1909 | block = next >> order; |
1910 | ex->fe_len += 1 << order; | |
1911 | } | |
1912 | ||
31562b95 | 1913 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
1914 | /* Should never happen! (but apparently sometimes does?!?) */ |
1915 | WARN_ON(1); | |
cd84bbba SB |
1916 | ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, |
1917 | "corruption or bug in mb_find_extent " | |
1918 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
1919 | block, order, needed, ex->fe_group, ex->fe_start, | |
1920 | ex->fe_len, ex->fe_logical); | |
43c73221 TT |
1921 | ex->fe_len = 0; |
1922 | ex->fe_start = 0; | |
1923 | ex->fe_group = 0; | |
1924 | } | |
c9de560d AT |
1925 | return ex->fe_len; |
1926 | } | |
1927 | ||
1928 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1929 | { | |
1930 | int ord; | |
1931 | int mlen = 0; | |
1932 | int max = 0; | |
1933 | int cur; | |
1934 | int start = ex->fe_start; | |
1935 | int len = ex->fe_len; | |
1936 | unsigned ret = 0; | |
1937 | int len0 = len; | |
1938 | void *buddy; | |
1939 | ||
1940 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1941 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1942 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1943 | mb_check_buddy(e4b); |
1944 | mb_mark_used_double(e4b, start, len); | |
1945 | ||
07b5b8e1 | 1946 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
1947 | e4b->bd_info->bb_free -= len; |
1948 | if (e4b->bd_info->bb_first_free == start) | |
1949 | e4b->bd_info->bb_first_free += len; | |
1950 | ||
1951 | /* let's maintain fragments counter */ | |
1952 | if (start != 0) | |
c5e8f3f3 | 1953 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1954 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1955 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1956 | if (mlen && max) |
1957 | e4b->bd_info->bb_fragments++; | |
1958 | else if (!mlen && !max) | |
1959 | e4b->bd_info->bb_fragments--; | |
1960 | ||
1961 | /* let's maintain buddy itself */ | |
1962 | while (len) { | |
1963 | ord = mb_find_order_for_block(e4b, start); | |
1964 | ||
1965 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1966 | /* the whole chunk may be allocated at once! */ | |
1967 | mlen = 1 << ord; | |
1968 | buddy = mb_find_buddy(e4b, ord, &max); | |
1969 | BUG_ON((start >> ord) >= max); | |
1970 | mb_set_bit(start >> ord, buddy); | |
1971 | e4b->bd_info->bb_counters[ord]--; | |
1972 | start += mlen; | |
1973 | len -= mlen; | |
1974 | BUG_ON(len < 0); | |
1975 | continue; | |
1976 | } | |
1977 | ||
1978 | /* store for history */ | |
1979 | if (ret == 0) | |
1980 | ret = len | (ord << 16); | |
1981 | ||
1982 | /* we have to split large buddy */ | |
1983 | BUG_ON(ord <= 0); | |
1984 | buddy = mb_find_buddy(e4b, ord, &max); | |
1985 | mb_set_bit(start >> ord, buddy); | |
1986 | e4b->bd_info->bb_counters[ord]--; | |
1987 | ||
1988 | ord--; | |
1989 | cur = (start >> ord) & ~1U; | |
1990 | buddy = mb_find_buddy(e4b, ord, &max); | |
1991 | mb_clear_bit(cur, buddy); | |
1992 | mb_clear_bit(cur + 1, buddy); | |
1993 | e4b->bd_info->bb_counters[ord]++; | |
1994 | e4b->bd_info->bb_counters[ord]++; | |
1995 | } | |
8a57d9d6 | 1996 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1997 | |
196e402a | 1998 | mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); |
c5e8f3f3 | 1999 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
2000 | mb_check_buddy(e4b); |
2001 | ||
2002 | return ret; | |
2003 | } | |
2004 | ||
2005 | /* | |
2006 | * Must be called under group lock! | |
2007 | */ | |
2008 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
2009 | struct ext4_buddy *e4b) | |
2010 | { | |
2011 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2012 | int ret; | |
2013 | ||
2014 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
2015 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
2016 | ||
2017 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
2018 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
2019 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
2020 | ||
2021 | /* preallocation can change ac_b_ex, thus we store actually | |
2022 | * allocated blocks for history */ | |
2023 | ac->ac_f_ex = ac->ac_b_ex; | |
2024 | ||
2025 | ac->ac_status = AC_STATUS_FOUND; | |
2026 | ac->ac_tail = ret & 0xffff; | |
2027 | ac->ac_buddy = ret >> 16; | |
2028 | ||
c3a326a6 AK |
2029 | /* |
2030 | * take the page reference. We want the page to be pinned | |
2031 | * so that we don't get a ext4_mb_init_cache_call for this | |
2032 | * group until we update the bitmap. That would mean we | |
2033 | * double allocate blocks. The reference is dropped | |
2034 | * in ext4_mb_release_context | |
2035 | */ | |
c9de560d AT |
2036 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
2037 | get_page(ac->ac_bitmap_page); | |
2038 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
2039 | get_page(ac->ac_buddy_page); | |
c9de560d | 2040 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 2041 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
2042 | spin_lock(&sbi->s_md_lock); |
2043 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
2044 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
2045 | spin_unlock(&sbi->s_md_lock); | |
2046 | } | |
53f86b17 RH |
2047 | /* |
2048 | * As we've just preallocated more space than | |
2049 | * user requested originally, we store allocated | |
2050 | * space in a special descriptor. | |
2051 | */ | |
2052 | if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
2053 | ext4_mb_new_preallocation(ac); | |
2054 | ||
c9de560d AT |
2055 | } |
2056 | ||
c9de560d AT |
2057 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
2058 | struct ext4_buddy *e4b, | |
2059 | int finish_group) | |
2060 | { | |
2061 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2062 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2063 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2064 | struct ext4_free_extent ex; | |
2065 | int max; | |
2066 | ||
032115fc AK |
2067 | if (ac->ac_status == AC_STATUS_FOUND) |
2068 | return; | |
c9de560d AT |
2069 | /* |
2070 | * We don't want to scan for a whole year | |
2071 | */ | |
2072 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
2073 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2074 | ac->ac_status = AC_STATUS_BREAK; | |
2075 | return; | |
2076 | } | |
2077 | ||
2078 | /* | |
2079 | * Haven't found good chunk so far, let's continue | |
2080 | */ | |
2081 | if (bex->fe_len < gex->fe_len) | |
2082 | return; | |
2083 | ||
2084 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
2085 | && bex->fe_group == e4b->bd_group) { | |
2086 | /* recheck chunk's availability - we don't know | |
2087 | * when it was found (within this lock-unlock | |
2088 | * period or not) */ | |
15c006a2 | 2089 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
2090 | if (max >= gex->fe_len) { |
2091 | ext4_mb_use_best_found(ac, e4b); | |
2092 | return; | |
2093 | } | |
2094 | } | |
2095 | } | |
2096 | ||
2097 | /* | |
2098 | * The routine checks whether found extent is good enough. If it is, | |
2099 | * then the extent gets marked used and flag is set to the context | |
2100 | * to stop scanning. Otherwise, the extent is compared with the | |
2101 | * previous found extent and if new one is better, then it's stored | |
2102 | * in the context. Later, the best found extent will be used, if | |
2103 | * mballoc can't find good enough extent. | |
2104 | * | |
2105 | * FIXME: real allocation policy is to be designed yet! | |
2106 | */ | |
2107 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
2108 | struct ext4_free_extent *ex, | |
2109 | struct ext4_buddy *e4b) | |
2110 | { | |
2111 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2112 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2113 | ||
2114 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
2115 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
2116 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
2117 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
2118 | ||
2119 | ac->ac_found++; | |
2120 | ||
2121 | /* | |
2122 | * The special case - take what you catch first | |
2123 | */ | |
2124 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2125 | *bex = *ex; | |
2126 | ext4_mb_use_best_found(ac, e4b); | |
2127 | return; | |
2128 | } | |
2129 | ||
2130 | /* | |
2131 | * Let's check whether the chuck is good enough | |
2132 | */ | |
2133 | if (ex->fe_len == gex->fe_len) { | |
2134 | *bex = *ex; | |
2135 | ext4_mb_use_best_found(ac, e4b); | |
2136 | return; | |
2137 | } | |
2138 | ||
2139 | /* | |
2140 | * If this is first found extent, just store it in the context | |
2141 | */ | |
2142 | if (bex->fe_len == 0) { | |
2143 | *bex = *ex; | |
2144 | return; | |
2145 | } | |
2146 | ||
2147 | /* | |
2148 | * If new found extent is better, store it in the context | |
2149 | */ | |
2150 | if (bex->fe_len < gex->fe_len) { | |
2151 | /* if the request isn't satisfied, any found extent | |
2152 | * larger than previous best one is better */ | |
2153 | if (ex->fe_len > bex->fe_len) | |
2154 | *bex = *ex; | |
2155 | } else if (ex->fe_len > gex->fe_len) { | |
2156 | /* if the request is satisfied, then we try to find | |
2157 | * an extent that still satisfy the request, but is | |
2158 | * smaller than previous one */ | |
2159 | if (ex->fe_len < bex->fe_len) | |
2160 | *bex = *ex; | |
2161 | } | |
2162 | ||
2163 | ext4_mb_check_limits(ac, e4b, 0); | |
2164 | } | |
2165 | ||
089ceecc ES |
2166 | static noinline_for_stack |
2167 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
2168 | struct ext4_buddy *e4b) |
2169 | { | |
2170 | struct ext4_free_extent ex = ac->ac_b_ex; | |
2171 | ext4_group_t group = ex.fe_group; | |
2172 | int max; | |
2173 | int err; | |
2174 | ||
2175 | BUG_ON(ex.fe_len <= 0); | |
2176 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2177 | if (err) | |
2178 | return err; | |
2179 | ||
2180 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 2181 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
2182 | |
2183 | if (max > 0) { | |
2184 | ac->ac_b_ex = ex; | |
2185 | ext4_mb_use_best_found(ac, e4b); | |
2186 | } | |
2187 | ||
2188 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2189 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2190 | |
2191 | return 0; | |
2192 | } | |
2193 | ||
089ceecc ES |
2194 | static noinline_for_stack |
2195 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
2196 | struct ext4_buddy *e4b) |
2197 | { | |
2198 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
2199 | int max; | |
2200 | int err; | |
2201 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 2202 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
2203 | struct ext4_free_extent ex; |
2204 | ||
2205 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
2206 | return 0; | |
838cd0cf YY |
2207 | if (grp->bb_free == 0) |
2208 | return 0; | |
c9de560d AT |
2209 | |
2210 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2211 | if (err) | |
2212 | return err; | |
2213 | ||
163a203d DW |
2214 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
2215 | ext4_mb_unload_buddy(e4b); | |
2216 | return 0; | |
2217 | } | |
2218 | ||
c9de560d | 2219 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 2220 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 2221 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 2222 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
2223 | |
2224 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
2225 | ext4_fsblk_t start; | |
2226 | ||
5661bd68 AM |
2227 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
2228 | ex.fe_start; | |
c9de560d AT |
2229 | /* use do_div to get remainder (would be 64-bit modulo) */ |
2230 | if (do_div(start, sbi->s_stripe) == 0) { | |
2231 | ac->ac_found++; | |
2232 | ac->ac_b_ex = ex; | |
2233 | ext4_mb_use_best_found(ac, e4b); | |
2234 | } | |
2235 | } else if (max >= ac->ac_g_ex.fe_len) { | |
2236 | BUG_ON(ex.fe_len <= 0); | |
2237 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2238 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2239 | ac->ac_found++; | |
2240 | ac->ac_b_ex = ex; | |
2241 | ext4_mb_use_best_found(ac, e4b); | |
2242 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
2243 | /* Sometimes, caller may want to merge even small | |
2244 | * number of blocks to an existing extent */ | |
2245 | BUG_ON(ex.fe_len <= 0); | |
2246 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2247 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2248 | ac->ac_found++; | |
2249 | ac->ac_b_ex = ex; | |
2250 | ext4_mb_use_best_found(ac, e4b); | |
2251 | } | |
2252 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 2253 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2254 | |
2255 | return 0; | |
2256 | } | |
2257 | ||
2258 | /* | |
2259 | * The routine scans buddy structures (not bitmap!) from given order | |
2260 | * to max order and tries to find big enough chunk to satisfy the req | |
2261 | */ | |
089ceecc ES |
2262 | static noinline_for_stack |
2263 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2264 | struct ext4_buddy *e4b) |
2265 | { | |
2266 | struct super_block *sb = ac->ac_sb; | |
2267 | struct ext4_group_info *grp = e4b->bd_info; | |
2268 | void *buddy; | |
2269 | int i; | |
2270 | int k; | |
2271 | int max; | |
2272 | ||
2273 | BUG_ON(ac->ac_2order <= 0); | |
4b68f6df | 2274 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { |
c9de560d AT |
2275 | if (grp->bb_counters[i] == 0) |
2276 | continue; | |
2277 | ||
2278 | buddy = mb_find_buddy(e4b, i, &max); | |
2279 | BUG_ON(buddy == NULL); | |
2280 | ||
ffad0a44 | 2281 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 DM |
2282 | if (k >= max) { |
2283 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, | |
2284 | "%d free clusters of order %d. But found 0", | |
2285 | grp->bb_counters[i], i); | |
2286 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, | |
2287 | e4b->bd_group, | |
2288 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
2289 | break; | |
2290 | } | |
c9de560d AT |
2291 | ac->ac_found++; |
2292 | ||
2293 | ac->ac_b_ex.fe_len = 1 << i; | |
2294 | ac->ac_b_ex.fe_start = k << i; | |
2295 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
2296 | ||
2297 | ext4_mb_use_best_found(ac, e4b); | |
2298 | ||
53f86b17 | 2299 | BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
c9de560d AT |
2300 | |
2301 | if (EXT4_SB(sb)->s_mb_stats) | |
2302 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
2303 | ||
2304 | break; | |
2305 | } | |
2306 | } | |
2307 | ||
2308 | /* | |
2309 | * The routine scans the group and measures all found extents. | |
2310 | * In order to optimize scanning, caller must pass number of | |
2311 | * free blocks in the group, so the routine can know upper limit. | |
2312 | */ | |
089ceecc ES |
2313 | static noinline_for_stack |
2314 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2315 | struct ext4_buddy *e4b) |
2316 | { | |
2317 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 2318 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2319 | struct ext4_free_extent ex; |
2320 | int i; | |
2321 | int free; | |
2322 | ||
2323 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
2324 | if (WARN_ON(free <= 0)) |
2325 | return; | |
c9de560d AT |
2326 | |
2327 | i = e4b->bd_info->bb_first_free; | |
2328 | ||
2329 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 2330 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
2331 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
2332 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 2333 | /* |
e56eb659 | 2334 | * IF we have corrupt bitmap, we won't find any |
26346ff6 | 2335 | * free blocks even though group info says we |
b483bb77 | 2336 | * have free blocks |
26346ff6 | 2337 | */ |
e29136f8 | 2338 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2339 | "%d free clusters as per " |
fde4d95a | 2340 | "group info. But bitmap says 0", |
26346ff6 | 2341 | free); |
736dedbb WS |
2342 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2343 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d AT |
2344 | break; |
2345 | } | |
2346 | ||
15c006a2 | 2347 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2348 | if (WARN_ON(ex.fe_len <= 0)) |
2349 | break; | |
26346ff6 | 2350 | if (free < ex.fe_len) { |
e29136f8 | 2351 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2352 | "%d free clusters as per " |
fde4d95a | 2353 | "group info. But got %d blocks", |
26346ff6 | 2354 | free, ex.fe_len); |
736dedbb WS |
2355 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2356 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e56eb659 AK |
2357 | /* |
2358 | * The number of free blocks differs. This mostly | |
2359 | * indicate that the bitmap is corrupt. So exit | |
2360 | * without claiming the space. | |
2361 | */ | |
2362 | break; | |
26346ff6 | 2363 | } |
ab0c00fc | 2364 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2365 | ext4_mb_measure_extent(ac, &ex, e4b); |
2366 | ||
2367 | i += ex.fe_len; | |
2368 | free -= ex.fe_len; | |
2369 | } | |
2370 | ||
2371 | ext4_mb_check_limits(ac, e4b, 1); | |
2372 | } | |
2373 | ||
2374 | /* | |
2375 | * This is a special case for storages like raid5 | |
506bf2d8 | 2376 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2377 | */ |
089ceecc ES |
2378 | static noinline_for_stack |
2379 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2380 | struct ext4_buddy *e4b) |
2381 | { | |
2382 | struct super_block *sb = ac->ac_sb; | |
2383 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2384 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2385 | struct ext4_free_extent ex; |
2386 | ext4_fsblk_t first_group_block; | |
2387 | ext4_fsblk_t a; | |
2388 | ext4_grpblk_t i; | |
2389 | int max; | |
2390 | ||
2391 | BUG_ON(sbi->s_stripe == 0); | |
2392 | ||
2393 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2394 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2395 | ||
c9de560d AT |
2396 | a = first_group_block + sbi->s_stripe - 1; |
2397 | do_div(a, sbi->s_stripe); | |
2398 | i = (a * sbi->s_stripe) - first_group_block; | |
2399 | ||
7137d7a4 | 2400 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2401 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 2402 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
2403 | if (max >= sbi->s_stripe) { |
2404 | ac->ac_found++; | |
ab0c00fc | 2405 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2406 | ac->ac_b_ex = ex; |
2407 | ext4_mb_use_best_found(ac, e4b); | |
2408 | break; | |
2409 | } | |
2410 | } | |
2411 | i += sbi->s_stripe; | |
2412 | } | |
2413 | } | |
2414 | ||
42ac1848 | 2415 | /* |
8ef123fe | 2416 | * This is also called BEFORE we load the buddy bitmap. |
42ac1848 | 2417 | * Returns either 1 or 0 indicating that the group is either suitable |
8ef123fe | 2418 | * for the allocation or not. |
42ac1848 | 2419 | */ |
8ef123fe | 2420 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
c9de560d AT |
2421 | ext4_group_t group, int cr) |
2422 | { | |
8ef123fe | 2423 | ext4_grpblk_t free, fragments; |
a4912123 | 2424 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2425 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2426 | ||
2427 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2428 | |
dddcd2f9 | 2429 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
8ef123fe | 2430 | return false; |
01fc48e8 | 2431 | |
dddcd2f9 | 2432 | free = grp->bb_free; |
2433 | if (free == 0) | |
8ef123fe | 2434 | return false; |
c9de560d | 2435 | |
c9de560d | 2436 | fragments = grp->bb_fragments; |
c9de560d | 2437 | if (fragments == 0) |
8ef123fe | 2438 | return false; |
c9de560d AT |
2439 | |
2440 | switch (cr) { | |
2441 | case 0: | |
2442 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2443 | |
a4912123 TT |
2444 | /* Avoid using the first bg of a flexgroup for data files */ |
2445 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2446 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2447 | ((group % flex_size) == 0)) | |
8ef123fe | 2448 | return false; |
a4912123 | 2449 | |
dddcd2f9 | 2450 | if (free < ac->ac_g_ex.fe_len) |
2451 | return false; | |
2452 | ||
4b68f6df | 2453 | if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) |
8ef123fe | 2454 | return true; |
40ae3487 TT |
2455 | |
2456 | if (grp->bb_largest_free_order < ac->ac_2order) | |
8ef123fe | 2457 | return false; |
40ae3487 | 2458 | |
8ef123fe | 2459 | return true; |
c9de560d AT |
2460 | case 1: |
2461 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2462 | return true; |
c9de560d AT |
2463 | break; |
2464 | case 2: | |
2465 | if (free >= ac->ac_g_ex.fe_len) | |
8ef123fe | 2466 | return true; |
c9de560d AT |
2467 | break; |
2468 | case 3: | |
8ef123fe | 2469 | return true; |
c9de560d AT |
2470 | default: |
2471 | BUG(); | |
2472 | } | |
2473 | ||
8ef123fe RH |
2474 | return false; |
2475 | } | |
2476 | ||
2477 | /* | |
2478 | * This could return negative error code if something goes wrong | |
2479 | * during ext4_mb_init_group(). This should not be called with | |
2480 | * ext4_lock_group() held. | |
a5fda113 TT |
2481 | * |
2482 | * Note: because we are conditionally operating with the group lock in | |
2483 | * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this | |
2484 | * function using __acquire and __release. This means we need to be | |
2485 | * super careful before messing with the error path handling via "goto | |
2486 | * out"! | |
8ef123fe RH |
2487 | */ |
2488 | static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, | |
2489 | ext4_group_t group, int cr) | |
2490 | { | |
2491 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
99377830 | 2492 | struct super_block *sb = ac->ac_sb; |
c1d2c7d4 | 2493 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
99377830 | 2494 | bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
8ef123fe RH |
2495 | ext4_grpblk_t free; |
2496 | int ret = 0; | |
2497 | ||
a6c75eaf HS |
2498 | if (sbi->s_mb_stats) |
2499 | atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); | |
a5fda113 | 2500 | if (should_lock) { |
99377830 | 2501 | ext4_lock_group(sb, group); |
a5fda113 TT |
2502 | __release(ext4_group_lock_ptr(sb, group)); |
2503 | } | |
8ef123fe RH |
2504 | free = grp->bb_free; |
2505 | if (free == 0) | |
2506 | goto out; | |
2507 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2508 | goto out; | |
2509 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
2510 | goto out; | |
a5fda113 TT |
2511 | if (should_lock) { |
2512 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2513 | ext4_unlock_group(sb, group); |
a5fda113 | 2514 | } |
8ef123fe RH |
2515 | |
2516 | /* We only do this if the grp has never been initialized */ | |
2517 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
c1d2c7d4 AZ |
2518 | struct ext4_group_desc *gdp = |
2519 | ext4_get_group_desc(sb, group, NULL); | |
2520 | int ret; | |
2521 | ||
2522 | /* cr=0/1 is a very optimistic search to find large | |
2523 | * good chunks almost for free. If buddy data is not | |
2524 | * ready, then this optimization makes no sense. But | |
2525 | * we never skip the first block group in a flex_bg, | |
2526 | * since this gets used for metadata block allocation, | |
2527 | * and we want to make sure we locate metadata blocks | |
2528 | * in the first block group in the flex_bg if possible. | |
2529 | */ | |
2530 | if (cr < 2 && | |
2531 | (!sbi->s_log_groups_per_flex || | |
2532 | ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && | |
2533 | !(ext4_has_group_desc_csum(sb) && | |
2534 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) | |
2535 | return 0; | |
2536 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); | |
8ef123fe RH |
2537 | if (ret) |
2538 | return ret; | |
2539 | } | |
2540 | ||
a5fda113 | 2541 | if (should_lock) { |
99377830 | 2542 | ext4_lock_group(sb, group); |
a5fda113 TT |
2543 | __release(ext4_group_lock_ptr(sb, group)); |
2544 | } | |
8ef123fe RH |
2545 | ret = ext4_mb_good_group(ac, group, cr); |
2546 | out: | |
a5fda113 TT |
2547 | if (should_lock) { |
2548 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2549 | ext4_unlock_group(sb, group); |
a5fda113 | 2550 | } |
8ef123fe | 2551 | return ret; |
c9de560d AT |
2552 | } |
2553 | ||
cfd73237 AZ |
2554 | /* |
2555 | * Start prefetching @nr block bitmaps starting at @group. | |
2556 | * Return the next group which needs to be prefetched. | |
2557 | */ | |
3d392b26 TT |
2558 | ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
2559 | unsigned int nr, int *cnt) | |
cfd73237 AZ |
2560 | { |
2561 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2562 | struct buffer_head *bh; | |
2563 | struct blk_plug plug; | |
2564 | ||
2565 | blk_start_plug(&plug); | |
2566 | while (nr-- > 0) { | |
2567 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2568 | NULL); | |
2569 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2570 | ||
2571 | /* | |
2572 | * Prefetch block groups with free blocks; but don't | |
2573 | * bother if it is marked uninitialized on disk, since | |
2574 | * it won't require I/O to read. Also only try to | |
2575 | * prefetch once, so we avoid getblk() call, which can | |
2576 | * be expensive. | |
2577 | */ | |
2578 | if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) && | |
2579 | EXT4_MB_GRP_NEED_INIT(grp) && | |
2580 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2581 | !(ext4_has_group_desc_csum(sb) && | |
2582 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2583 | bh = ext4_read_block_bitmap_nowait(sb, group, true); | |
2584 | if (bh && !IS_ERR(bh)) { | |
2585 | if (!buffer_uptodate(bh) && cnt) | |
2586 | (*cnt)++; | |
2587 | brelse(bh); | |
2588 | } | |
2589 | } | |
2590 | if (++group >= ngroups) | |
2591 | group = 0; | |
2592 | } | |
2593 | blk_finish_plug(&plug); | |
2594 | return group; | |
2595 | } | |
2596 | ||
2597 | /* | |
2598 | * Prefetching reads the block bitmap into the buffer cache; but we | |
2599 | * need to make sure that the buddy bitmap in the page cache has been | |
2600 | * initialized. Note that ext4_mb_init_group() will block if the I/O | |
2601 | * is not yet completed, or indeed if it was not initiated by | |
2602 | * ext4_mb_prefetch did not start the I/O. | |
2603 | * | |
2604 | * TODO: We should actually kick off the buddy bitmap setup in a work | |
2605 | * queue when the buffer I/O is completed, so that we don't block | |
2606 | * waiting for the block allocation bitmap read to finish when | |
2607 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). | |
2608 | */ | |
3d392b26 TT |
2609 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
2610 | unsigned int nr) | |
cfd73237 AZ |
2611 | { |
2612 | while (nr-- > 0) { | |
2613 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2614 | NULL); | |
2615 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2616 | ||
2617 | if (!group) | |
2618 | group = ext4_get_groups_count(sb); | |
2619 | group--; | |
2620 | grp = ext4_get_group_info(sb, group); | |
2621 | ||
2622 | if (EXT4_MB_GRP_NEED_INIT(grp) && | |
2623 | ext4_free_group_clusters(sb, gdp) > 0 && | |
2624 | !(ext4_has_group_desc_csum(sb) && | |
2625 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) { | |
2626 | if (ext4_mb_init_group(sb, group, GFP_NOFS)) | |
2627 | break; | |
2628 | } | |
2629 | } | |
2630 | } | |
2631 | ||
4ddfef7b ES |
2632 | static noinline_for_stack int |
2633 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2634 | { |
cfd73237 | 2635 | ext4_group_t prefetch_grp = 0, ngroups, group, i; |
bbc4ec77 | 2636 | int cr = -1; |
42ac1848 | 2637 | int err = 0, first_err = 0; |
cfd73237 | 2638 | unsigned int nr = 0, prefetch_ios = 0; |
c9de560d AT |
2639 | struct ext4_sb_info *sbi; |
2640 | struct super_block *sb; | |
2641 | struct ext4_buddy e4b; | |
66d5e027 | 2642 | int lost; |
c9de560d AT |
2643 | |
2644 | sb = ac->ac_sb; | |
2645 | sbi = EXT4_SB(sb); | |
8df9675f | 2646 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2647 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2648 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2649 | ngroups = sbi->s_blockfile_groups; |
2650 | ||
c9de560d AT |
2651 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2652 | ||
2653 | /* first, try the goal */ | |
2654 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2655 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2656 | goto out; | |
2657 | ||
2658 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2659 | goto out; | |
2660 | ||
2661 | /* | |
e9a3cd48 | 2662 | * ac->ac_2order is set only if the fe_len is a power of 2 |
2663 | * if ac->ac_2order is set we also set criteria to 0 so that we | |
c9de560d AT |
2664 | * try exact allocation using buddy. |
2665 | */ | |
2666 | i = fls(ac->ac_g_ex.fe_len); | |
2667 | ac->ac_2order = 0; | |
2668 | /* | |
2669 | * We search using buddy data only if the order of the request | |
2670 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2671 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2672 | * We also support searching for power-of-two requests only for |
2673 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2674 | */ |
4b68f6df | 2675 | if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { |
c9de560d AT |
2676 | /* |
2677 | * This should tell if fe_len is exactly power of 2 | |
2678 | */ | |
2679 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
1a5d5e5d | 2680 | ac->ac_2order = array_index_nospec(i - 1, |
4b68f6df | 2681 | MB_NUM_ORDERS(sb)); |
c9de560d AT |
2682 | } |
2683 | ||
4ba74d00 TT |
2684 | /* if stream allocation is enabled, use global goal */ |
2685 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2686 | /* TBD: may be hot point */ |
2687 | spin_lock(&sbi->s_md_lock); | |
2688 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2689 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2690 | spin_unlock(&sbi->s_md_lock); | |
2691 | } | |
4ba74d00 | 2692 | |
c9de560d AT |
2693 | /* Let's just scan groups to find more-less suitable blocks */ |
2694 | cr = ac->ac_2order ? 0 : 1; | |
2695 | /* | |
2696 | * cr == 0 try to get exact allocation, | |
2697 | * cr == 3 try to get anything | |
2698 | */ | |
2699 | repeat: | |
2700 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2701 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2702 | /* |
2703 | * searching for the right group start | |
2704 | * from the goal value specified | |
2705 | */ | |
2706 | group = ac->ac_g_ex.fe_group; | |
196e402a HS |
2707 | ac->ac_last_optimal_group = group; |
2708 | ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; | |
cfd73237 | 2709 | prefetch_grp = group; |
ed8f9c75 | 2710 | |
196e402a HS |
2711 | for (i = 0; i < ngroups; group = next_linear_group(ac, group, ngroups), |
2712 | i++) { | |
2713 | int ret = 0, new_cr; | |
2714 | ||
2ed5724d | 2715 | cond_resched(); |
196e402a HS |
2716 | |
2717 | ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups); | |
2718 | if (new_cr != cr) { | |
2719 | cr = new_cr; | |
2720 | goto repeat; | |
2721 | } | |
c9de560d | 2722 | |
cfd73237 AZ |
2723 | /* |
2724 | * Batch reads of the block allocation bitmaps | |
2725 | * to get multiple READs in flight; limit | |
2726 | * prefetching at cr=0/1, otherwise mballoc can | |
2727 | * spend a lot of time loading imperfect groups | |
2728 | */ | |
2729 | if ((prefetch_grp == group) && | |
2730 | (cr > 1 || | |
2731 | prefetch_ios < sbi->s_mb_prefetch_limit)) { | |
2732 | unsigned int curr_ios = prefetch_ios; | |
2733 | ||
2734 | nr = sbi->s_mb_prefetch; | |
2735 | if (ext4_has_feature_flex_bg(sb)) { | |
82ef1370 CX |
2736 | nr = 1 << sbi->s_log_groups_per_flex; |
2737 | nr -= group & (nr - 1); | |
2738 | nr = min(nr, sbi->s_mb_prefetch); | |
cfd73237 AZ |
2739 | } |
2740 | prefetch_grp = ext4_mb_prefetch(sb, group, | |
2741 | nr, &prefetch_ios); | |
2742 | if (prefetch_ios == curr_ios) | |
2743 | nr = 0; | |
2744 | } | |
2745 | ||
8a57d9d6 | 2746 | /* This now checks without needing the buddy page */ |
8ef123fe | 2747 | ret = ext4_mb_good_group_nolock(ac, group, cr); |
42ac1848 LC |
2748 | if (ret <= 0) { |
2749 | if (!first_err) | |
2750 | first_err = ret; | |
c9de560d | 2751 | continue; |
42ac1848 | 2752 | } |
c9de560d | 2753 | |
c9de560d AT |
2754 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2755 | if (err) | |
2756 | goto out; | |
2757 | ||
2758 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2759 | |
2760 | /* | |
2761 | * We need to check again after locking the | |
2762 | * block group | |
2763 | */ | |
42ac1848 | 2764 | ret = ext4_mb_good_group(ac, group, cr); |
8ef123fe | 2765 | if (ret == 0) { |
c9de560d | 2766 | ext4_unlock_group(sb, group); |
e39e07fd | 2767 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2768 | continue; |
2769 | } | |
2770 | ||
2771 | ac->ac_groups_scanned++; | |
d9b22cf9 | 2772 | if (cr == 0) |
c9de560d | 2773 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2774 | else if (cr == 1 && sbi->s_stripe && |
2775 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2776 | ext4_mb_scan_aligned(ac, &e4b); |
2777 | else | |
2778 | ext4_mb_complex_scan_group(ac, &e4b); | |
2779 | ||
2780 | ext4_unlock_group(sb, group); | |
e39e07fd | 2781 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2782 | |
2783 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2784 | break; | |
2785 | } | |
a6c75eaf HS |
2786 | /* Processed all groups and haven't found blocks */ |
2787 | if (sbi->s_mb_stats && i == ngroups) | |
2788 | atomic64_inc(&sbi->s_bal_cX_failed[cr]); | |
c9de560d AT |
2789 | } |
2790 | ||
2791 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2792 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2793 | /* | |
2794 | * We've been searching too long. Let's try to allocate | |
2795 | * the best chunk we've found so far | |
2796 | */ | |
c9de560d AT |
2797 | ext4_mb_try_best_found(ac, &e4b); |
2798 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2799 | /* | |
2800 | * Someone more lucky has already allocated it. | |
2801 | * The only thing we can do is just take first | |
2802 | * found block(s) | |
c9de560d | 2803 | */ |
66d5e027 | 2804 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
2805 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", | |
c55ee7d2 | 2806 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
2807 | ac->ac_b_ex.fe_len, lost); | |
2808 | ||
c9de560d AT |
2809 | ac->ac_b_ex.fe_group = 0; |
2810 | ac->ac_b_ex.fe_start = 0; | |
2811 | ac->ac_b_ex.fe_len = 0; | |
2812 | ac->ac_status = AC_STATUS_CONTINUE; | |
2813 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2814 | cr = 3; | |
c9de560d AT |
2815 | goto repeat; |
2816 | } | |
2817 | } | |
a6c75eaf HS |
2818 | |
2819 | if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) | |
2820 | atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); | |
c9de560d | 2821 | out: |
42ac1848 LC |
2822 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2823 | err = first_err; | |
bbc4ec77 | 2824 | |
d3df1453 | 2825 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
bbc4ec77 RH |
2826 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
2827 | ac->ac_flags, cr, err); | |
cfd73237 AZ |
2828 | |
2829 | if (nr) | |
2830 | ext4_mb_prefetch_fini(sb, prefetch_grp, nr); | |
2831 | ||
c9de560d AT |
2832 | return err; |
2833 | } | |
2834 | ||
c9de560d AT |
2835 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2836 | { | |
247dbed8 | 2837 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2838 | ext4_group_t group; |
2839 | ||
8df9675f | 2840 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2841 | return NULL; |
c9de560d | 2842 | group = *pos + 1; |
a9df9a49 | 2843 | return (void *) ((unsigned long) group); |
c9de560d AT |
2844 | } |
2845 | ||
2846 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2847 | { | |
247dbed8 | 2848 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
c9de560d AT |
2849 | ext4_group_t group; |
2850 | ||
2851 | ++*pos; | |
8df9675f | 2852 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2853 | return NULL; |
2854 | group = *pos + 1; | |
a9df9a49 | 2855 | return (void *) ((unsigned long) group); |
c9de560d AT |
2856 | } |
2857 | ||
2858 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2859 | { | |
247dbed8 | 2860 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); |
a9df9a49 | 2861 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2862 | int i; |
1c8457ca | 2863 | int err, buddy_loaded = 0; |
c9de560d | 2864 | struct ext4_buddy e4b; |
1c8457ca | 2865 | struct ext4_group_info *grinfo; |
2df2c340 AB |
2866 | unsigned char blocksize_bits = min_t(unsigned char, |
2867 | sb->s_blocksize_bits, | |
2868 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
2869 | struct sg { |
2870 | struct ext4_group_info info; | |
b80b32b6 | 2871 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
2872 | } sg; |
2873 | ||
2874 | group--; | |
2875 | if (group == 0) | |
97b4af2f RV |
2876 | seq_puts(seq, "#group: free frags first [" |
2877 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 2878 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 2879 | |
b80b32b6 TT |
2880 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
2881 | sizeof(struct ext4_group_info); | |
2882 | ||
1c8457ca AK |
2883 | grinfo = ext4_get_group_info(sb, group); |
2884 | /* Load the group info in memory only if not already loaded. */ | |
2885 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2886 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2887 | if (err) { | |
2888 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2889 | return 0; | |
2890 | } | |
2891 | buddy_loaded = 1; | |
c9de560d | 2892 | } |
1c8457ca | 2893 | |
b80b32b6 | 2894 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2895 | |
2896 | if (buddy_loaded) | |
2897 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2898 | |
a9df9a49 | 2899 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2900 | sg.info.bb_fragments, sg.info.bb_first_free); |
2901 | for (i = 0; i <= 13; i++) | |
2df2c340 | 2902 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d | 2903 | sg.info.bb_counters[i] : 0); |
e0d438c7 | 2904 | seq_puts(seq, " ]\n"); |
c9de560d AT |
2905 | |
2906 | return 0; | |
2907 | } | |
2908 | ||
2909 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2910 | { | |
2911 | } | |
2912 | ||
247dbed8 | 2913 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2914 | .start = ext4_mb_seq_groups_start, |
2915 | .next = ext4_mb_seq_groups_next, | |
2916 | .stop = ext4_mb_seq_groups_stop, | |
2917 | .show = ext4_mb_seq_groups_show, | |
2918 | }; | |
2919 | ||
a6c75eaf HS |
2920 | int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) |
2921 | { | |
2922 | struct super_block *sb = (struct super_block *)seq->private; | |
2923 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2924 | ||
2925 | seq_puts(seq, "mballoc:\n"); | |
2926 | if (!sbi->s_mb_stats) { | |
2927 | seq_puts(seq, "\tmb stats collection turned off.\n"); | |
2928 | seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); | |
2929 | return 0; | |
2930 | } | |
2931 | seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); | |
2932 | seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); | |
2933 | ||
2934 | seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned)); | |
2935 | ||
2936 | seq_puts(seq, "\tcr0_stats:\n"); | |
2937 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0])); | |
2938 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2939 | atomic64_read(&sbi->s_bal_cX_groups_considered[0])); | |
2940 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2941 | atomic64_read(&sbi->s_bal_cX_failed[0])); | |
196e402a HS |
2942 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2943 | atomic_read(&sbi->s_bal_cr0_bad_suggestions)); | |
a6c75eaf HS |
2944 | |
2945 | seq_puts(seq, "\tcr1_stats:\n"); | |
2946 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1])); | |
2947 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2948 | atomic64_read(&sbi->s_bal_cX_groups_considered[1])); | |
2949 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2950 | atomic64_read(&sbi->s_bal_cX_failed[1])); | |
196e402a HS |
2951 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
2952 | atomic_read(&sbi->s_bal_cr1_bad_suggestions)); | |
a6c75eaf HS |
2953 | |
2954 | seq_puts(seq, "\tcr2_stats:\n"); | |
2955 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2])); | |
2956 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2957 | atomic64_read(&sbi->s_bal_cX_groups_considered[2])); | |
2958 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2959 | atomic64_read(&sbi->s_bal_cX_failed[2])); | |
2960 | ||
2961 | seq_puts(seq, "\tcr3_stats:\n"); | |
2962 | seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3])); | |
2963 | seq_printf(seq, "\t\tgroups_considered: %llu\n", | |
2964 | atomic64_read(&sbi->s_bal_cX_groups_considered[3])); | |
2965 | seq_printf(seq, "\t\tuseless_loops: %llu\n", | |
2966 | atomic64_read(&sbi->s_bal_cX_failed[3])); | |
2967 | seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned)); | |
2968 | seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); | |
2969 | seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); | |
2970 | seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); | |
2971 | seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); | |
2972 | ||
2973 | seq_printf(seq, "\tbuddies_generated: %u/%u\n", | |
2974 | atomic_read(&sbi->s_mb_buddies_generated), | |
2975 | ext4_get_groups_count(sb)); | |
2976 | seq_printf(seq, "\tbuddies_time_used: %llu\n", | |
2977 | atomic64_read(&sbi->s_mb_generation_time)); | |
2978 | seq_printf(seq, "\tpreallocated: %u\n", | |
2979 | atomic_read(&sbi->s_mb_preallocated)); | |
2980 | seq_printf(seq, "\tdiscarded: %u\n", | |
2981 | atomic_read(&sbi->s_mb_discarded)); | |
2982 | return 0; | |
2983 | } | |
2984 | ||
f68f4063 | 2985 | static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) |
a5fda113 | 2986 | __acquires(&EXT4_SB(sb)->s_mb_rb_lock) |
f68f4063 HS |
2987 | { |
2988 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); | |
2989 | unsigned long position; | |
2990 | ||
2991 | read_lock(&EXT4_SB(sb)->s_mb_rb_lock); | |
2992 | ||
2993 | if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1) | |
2994 | return NULL; | |
2995 | position = *pos + 1; | |
2996 | return (void *) ((unsigned long) position); | |
2997 | } | |
2998 | ||
2999 | static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) | |
3000 | { | |
3001 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); | |
3002 | unsigned long position; | |
3003 | ||
3004 | ++*pos; | |
3005 | if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1) | |
3006 | return NULL; | |
3007 | position = *pos + 1; | |
3008 | return (void *) ((unsigned long) position); | |
3009 | } | |
3010 | ||
3011 | static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) | |
3012 | { | |
3013 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); | |
3014 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3015 | unsigned long position = ((unsigned long) v); | |
3016 | struct ext4_group_info *grp; | |
3017 | struct rb_node *n; | |
3018 | unsigned int count, min, max; | |
3019 | ||
3020 | position--; | |
3021 | if (position >= MB_NUM_ORDERS(sb)) { | |
3022 | seq_puts(seq, "fragment_size_tree:\n"); | |
3023 | n = rb_first(&sbi->s_mb_avg_fragment_size_root); | |
3024 | if (!n) { | |
3025 | seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n"); | |
3026 | return 0; | |
3027 | } | |
3028 | grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb); | |
3029 | min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0; | |
3030 | count = 1; | |
3031 | while (rb_next(n)) { | |
3032 | count++; | |
3033 | n = rb_next(n); | |
3034 | } | |
3035 | grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb); | |
3036 | max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0; | |
3037 | ||
3038 | seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n", | |
3039 | min, max, count); | |
3040 | return 0; | |
3041 | } | |
3042 | ||
3043 | if (position == 0) { | |
3044 | seq_printf(seq, "optimize_scan: %d\n", | |
3045 | test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); | |
3046 | seq_puts(seq, "max_free_order_lists:\n"); | |
3047 | } | |
3048 | count = 0; | |
3049 | list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], | |
3050 | bb_largest_free_order_node) | |
3051 | count++; | |
3052 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", | |
3053 | (unsigned int)position, count); | |
3054 | ||
3055 | return 0; | |
3056 | } | |
3057 | ||
3058 | static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) | |
a5fda113 | 3059 | __releases(&EXT4_SB(sb)->s_mb_rb_lock) |
f68f4063 HS |
3060 | { |
3061 | struct super_block *sb = PDE_DATA(file_inode(seq->file)); | |
3062 | ||
3063 | read_unlock(&EXT4_SB(sb)->s_mb_rb_lock); | |
3064 | } | |
3065 | ||
3066 | const struct seq_operations ext4_mb_seq_structs_summary_ops = { | |
3067 | .start = ext4_mb_seq_structs_summary_start, | |
3068 | .next = ext4_mb_seq_structs_summary_next, | |
3069 | .stop = ext4_mb_seq_structs_summary_stop, | |
3070 | .show = ext4_mb_seq_structs_summary_show, | |
3071 | }; | |
3072 | ||
fb1813f4 CW |
3073 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
3074 | { | |
3075 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3076 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
3077 | ||
3078 | BUG_ON(!cachep); | |
3079 | return cachep; | |
3080 | } | |
5f21b0e6 | 3081 | |
28623c2f TT |
3082 | /* |
3083 | * Allocate the top-level s_group_info array for the specified number | |
3084 | * of groups | |
3085 | */ | |
3086 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
3087 | { | |
3088 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3089 | unsigned size; | |
df3da4ea | 3090 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
3091 | |
3092 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
3093 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
3094 | if (size <= sbi->s_group_info_size) | |
3095 | return 0; | |
3096 | ||
3097 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 3098 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
3099 | if (!new_groupinfo) { |
3100 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
3101 | return -ENOMEM; | |
3102 | } | |
df3da4ea SJS |
3103 | rcu_read_lock(); |
3104 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
3105 | if (old_groupinfo) | |
3106 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 3107 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
3108 | rcu_read_unlock(); |
3109 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 3110 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
3111 | if (old_groupinfo) |
3112 | ext4_kvfree_array_rcu(old_groupinfo); | |
666245d9 | 3113 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
28623c2f TT |
3114 | sbi->s_group_info_size); |
3115 | return 0; | |
3116 | } | |
3117 | ||
5f21b0e6 | 3118 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 3119 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
3120 | struct ext4_group_desc *desc) |
3121 | { | |
fb1813f4 | 3122 | int i; |
5f21b0e6 | 3123 | int metalen = 0; |
df3da4ea | 3124 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
3125 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3126 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 3127 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
3128 | |
3129 | /* | |
3130 | * First check if this group is the first of a reserved block. | |
3131 | * If it's true, we have to allocate a new table of pointers | |
3132 | * to ext4_group_info structures | |
3133 | */ | |
3134 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
3135 | metalen = sizeof(*meta_group_info) << | |
3136 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 3137 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 3138 | if (meta_group_info == NULL) { |
7f6a11e7 | 3139 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 3140 | "for a buddy group"); |
5f21b0e6 FB |
3141 | goto exit_meta_group_info; |
3142 | } | |
df3da4ea SJS |
3143 | rcu_read_lock(); |
3144 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
3145 | rcu_read_unlock(); | |
5f21b0e6 FB |
3146 | } |
3147 | ||
df3da4ea | 3148 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
3149 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
3150 | ||
4fdb5543 | 3151 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 3152 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 3153 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
3154 | goto exit_group_info; |
3155 | } | |
3156 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
3157 | &(meta_group_info[i]->bb_state)); | |
3158 | ||
3159 | /* | |
3160 | * initialize bb_free to be able to skip | |
3161 | * empty groups without initialization | |
3162 | */ | |
8844618d TT |
3163 | if (ext4_has_group_desc_csum(sb) && |
3164 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 3165 | meta_group_info[i]->bb_free = |
cff1dfd7 | 3166 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
3167 | } else { |
3168 | meta_group_info[i]->bb_free = | |
021b65bb | 3169 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
3170 | } |
3171 | ||
3172 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 3173 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 3174 | meta_group_info[i]->bb_free_root = RB_ROOT; |
196e402a HS |
3175 | INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); |
3176 | RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb); | |
8a57d9d6 | 3177 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
196e402a | 3178 | meta_group_info[i]->bb_group = group; |
5f21b0e6 | 3179 | |
a3450215 | 3180 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
3181 | return 0; |
3182 | ||
3183 | exit_group_info: | |
3184 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 3185 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
3186 | struct ext4_group_info ***group_info; |
3187 | ||
3188 | rcu_read_lock(); | |
3189 | group_info = rcu_dereference(sbi->s_group_info); | |
3190 | kfree(group_info[idx]); | |
3191 | group_info[idx] = NULL; | |
3192 | rcu_read_unlock(); | |
caaf7a29 | 3193 | } |
5f21b0e6 FB |
3194 | exit_meta_group_info: |
3195 | return -ENOMEM; | |
3196 | } /* ext4_mb_add_groupinfo */ | |
3197 | ||
c9de560d AT |
3198 | static int ext4_mb_init_backend(struct super_block *sb) |
3199 | { | |
8df9675f | 3200 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 3201 | ext4_group_t i; |
c9de560d | 3202 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 3203 | int err; |
5f21b0e6 | 3204 | struct ext4_group_desc *desc; |
df3da4ea | 3205 | struct ext4_group_info ***group_info; |
fb1813f4 | 3206 | struct kmem_cache *cachep; |
5f21b0e6 | 3207 | |
28623c2f TT |
3208 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
3209 | if (err) | |
3210 | return err; | |
c9de560d | 3211 | |
c9de560d AT |
3212 | sbi->s_buddy_cache = new_inode(sb); |
3213 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 3214 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
3215 | goto err_freesgi; |
3216 | } | |
48e6061b YJ |
3217 | /* To avoid potentially colliding with an valid on-disk inode number, |
3218 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
3219 | * not in the inode hash, so it should never be found by iget(), but | |
3220 | * this will avoid confusion if it ever shows up during debugging. */ | |
3221 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 3222 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 3223 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3224 | cond_resched(); |
c9de560d AT |
3225 | desc = ext4_get_group_desc(sb, i, NULL); |
3226 | if (desc == NULL) { | |
9d8b9ec4 | 3227 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
3228 | goto err_freebuddy; |
3229 | } | |
5f21b0e6 FB |
3230 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
3231 | goto err_freebuddy; | |
c9de560d AT |
3232 | } |
3233 | ||
cfd73237 | 3234 | if (ext4_has_feature_flex_bg(sb)) { |
f91436d5 ST |
3235 | /* a single flex group is supposed to be read by a single IO. |
3236 | * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is | |
3237 | * unsigned integer, so the maximum shift is 32. | |
3238 | */ | |
3239 | if (sbi->s_es->s_log_groups_per_flex >= 32) { | |
3240 | ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); | |
a8867f4e | 3241 | goto err_freebuddy; |
f91436d5 ST |
3242 | } |
3243 | sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, | |
82ef1370 | 3244 | BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); |
cfd73237 AZ |
3245 | sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ |
3246 | } else { | |
3247 | sbi->s_mb_prefetch = 32; | |
3248 | } | |
3249 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) | |
3250 | sbi->s_mb_prefetch = ext4_get_groups_count(sb); | |
3251 | /* now many real IOs to prefetch within a single allocation at cr=0 | |
3252 | * given cr=0 is an CPU-related optimization we shouldn't try to | |
3253 | * load too many groups, at some point we should start to use what | |
3254 | * we've got in memory. | |
3255 | * with an average random access time 5ms, it'd take a second to get | |
3256 | * 200 groups (* N with flex_bg), so let's make this limit 4 | |
3257 | */ | |
3258 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; | |
3259 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) | |
3260 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); | |
3261 | ||
c9de560d AT |
3262 | return 0; |
3263 | ||
3264 | err_freebuddy: | |
fb1813f4 | 3265 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 3266 | while (i-- > 0) |
fb1813f4 | 3267 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 3268 | i = sbi->s_group_info_size; |
df3da4ea SJS |
3269 | rcu_read_lock(); |
3270 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 3271 | while (i-- > 0) |
df3da4ea SJS |
3272 | kfree(group_info[i]); |
3273 | rcu_read_unlock(); | |
c9de560d AT |
3274 | iput(sbi->s_buddy_cache); |
3275 | err_freesgi: | |
df3da4ea SJS |
3276 | rcu_read_lock(); |
3277 | kvfree(rcu_dereference(sbi->s_group_info)); | |
3278 | rcu_read_unlock(); | |
c9de560d AT |
3279 | return -ENOMEM; |
3280 | } | |
3281 | ||
2892c15d ES |
3282 | static void ext4_groupinfo_destroy_slabs(void) |
3283 | { | |
3284 | int i; | |
3285 | ||
3286 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 3287 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
3288 | ext4_groupinfo_caches[i] = NULL; |
3289 | } | |
3290 | } | |
3291 | ||
3292 | static int ext4_groupinfo_create_slab(size_t size) | |
3293 | { | |
3294 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
3295 | int slab_size; | |
3296 | int blocksize_bits = order_base_2(size); | |
3297 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3298 | struct kmem_cache *cachep; | |
3299 | ||
3300 | if (cache_index >= NR_GRPINFO_CACHES) | |
3301 | return -EINVAL; | |
3302 | ||
3303 | if (unlikely(cache_index < 0)) | |
3304 | cache_index = 0; | |
3305 | ||
3306 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
3307 | if (ext4_groupinfo_caches[cache_index]) { | |
3308 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
3309 | return 0; /* Already created */ | |
3310 | } | |
3311 | ||
3312 | slab_size = offsetof(struct ext4_group_info, | |
3313 | bb_counters[blocksize_bits + 2]); | |
3314 | ||
3315 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
3316 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
3317 | NULL); | |
3318 | ||
823ba01f TM |
3319 | ext4_groupinfo_caches[cache_index] = cachep; |
3320 | ||
2892c15d ES |
3321 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
3322 | if (!cachep) { | |
9d8b9ec4 TT |
3323 | printk(KERN_EMERG |
3324 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
3325 | return -ENOMEM; |
3326 | } | |
3327 | ||
2892c15d ES |
3328 | return 0; |
3329 | } | |
3330 | ||
55cdd0af WJ |
3331 | static void ext4_discard_work(struct work_struct *work) |
3332 | { | |
3333 | struct ext4_sb_info *sbi = container_of(work, | |
3334 | struct ext4_sb_info, s_discard_work); | |
3335 | struct super_block *sb = sbi->s_sb; | |
3336 | struct ext4_free_data *fd, *nfd; | |
3337 | struct ext4_buddy e4b; | |
3338 | struct list_head discard_list; | |
3339 | ext4_group_t grp, load_grp; | |
3340 | int err = 0; | |
3341 | ||
3342 | INIT_LIST_HEAD(&discard_list); | |
3343 | spin_lock(&sbi->s_md_lock); | |
3344 | list_splice_init(&sbi->s_discard_list, &discard_list); | |
3345 | spin_unlock(&sbi->s_md_lock); | |
3346 | ||
3347 | load_grp = UINT_MAX; | |
3348 | list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { | |
3349 | /* | |
5036ab8d WJ |
3350 | * If filesystem is umounting or no memory or suffering |
3351 | * from no space, give up the discard | |
55cdd0af | 3352 | */ |
5036ab8d WJ |
3353 | if ((sb->s_flags & SB_ACTIVE) && !err && |
3354 | !atomic_read(&sbi->s_retry_alloc_pending)) { | |
55cdd0af WJ |
3355 | grp = fd->efd_group; |
3356 | if (grp != load_grp) { | |
3357 | if (load_grp != UINT_MAX) | |
3358 | ext4_mb_unload_buddy(&e4b); | |
3359 | ||
3360 | err = ext4_mb_load_buddy(sb, grp, &e4b); | |
3361 | if (err) { | |
3362 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3363 | load_grp = UINT_MAX; | |
3364 | continue; | |
3365 | } else { | |
3366 | load_grp = grp; | |
3367 | } | |
3368 | } | |
3369 | ||
3370 | ext4_lock_group(sb, grp); | |
3371 | ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, | |
3372 | fd->efd_start_cluster + fd->efd_count - 1, 1); | |
3373 | ext4_unlock_group(sb, grp); | |
3374 | } | |
3375 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3376 | } | |
3377 | ||
3378 | if (load_grp != UINT_MAX) | |
3379 | ext4_mb_unload_buddy(&e4b); | |
3380 | } | |
3381 | ||
9d99012f | 3382 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
3383 | { |
3384 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 3385 | unsigned i, j; |
935244cd | 3386 | unsigned offset, offset_incr; |
c9de560d | 3387 | unsigned max; |
74767c5a | 3388 | int ret; |
c9de560d | 3389 | |
4b68f6df | 3390 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
3391 | |
3392 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
3393 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
3394 | ret = -ENOMEM; |
3395 | goto out; | |
c9de560d | 3396 | } |
ff7ef329 | 3397 | |
4b68f6df | 3398 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
3399 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
3400 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
3401 | ret = -ENOMEM; |
3402 | goto out; | |
3403 | } | |
3404 | ||
2892c15d ES |
3405 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
3406 | if (ret < 0) | |
3407 | goto out; | |
c9de560d AT |
3408 | |
3409 | /* order 0 is regular bitmap */ | |
3410 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
3411 | sbi->s_mb_offsets[0] = 0; | |
3412 | ||
3413 | i = 1; | |
3414 | offset = 0; | |
935244cd | 3415 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
3416 | max = sb->s_blocksize << 2; |
3417 | do { | |
3418 | sbi->s_mb_offsets[i] = offset; | |
3419 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
3420 | offset += offset_incr; |
3421 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
3422 | max = max >> 1; |
3423 | i++; | |
4b68f6df HS |
3424 | } while (i < MB_NUM_ORDERS(sb)); |
3425 | ||
196e402a HS |
3426 | sbi->s_mb_avg_fragment_size_root = RB_ROOT; |
3427 | sbi->s_mb_largest_free_orders = | |
3428 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3429 | GFP_KERNEL); | |
3430 | if (!sbi->s_mb_largest_free_orders) { | |
3431 | ret = -ENOMEM; | |
3432 | goto out; | |
3433 | } | |
3434 | sbi->s_mb_largest_free_orders_locks = | |
3435 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3436 | GFP_KERNEL); | |
3437 | if (!sbi->s_mb_largest_free_orders_locks) { | |
3438 | ret = -ENOMEM; | |
3439 | goto out; | |
3440 | } | |
3441 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3442 | INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); | |
3443 | rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); | |
3444 | } | |
3445 | rwlock_init(&sbi->s_mb_rb_lock); | |
c9de560d | 3446 | |
c9de560d | 3447 | spin_lock_init(&sbi->s_md_lock); |
d08854f5 | 3448 | sbi->s_mb_free_pending = 0; |
a0154344 | 3449 | INIT_LIST_HEAD(&sbi->s_freed_data_list); |
55cdd0af WJ |
3450 | INIT_LIST_HEAD(&sbi->s_discard_list); |
3451 | INIT_WORK(&sbi->s_discard_work, ext4_discard_work); | |
5036ab8d | 3452 | atomic_set(&sbi->s_retry_alloc_pending, 0); |
c9de560d AT |
3453 | |
3454 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
3455 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
3456 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
3457 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
3458 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27bc446e | 3459 | sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC; |
27baebb8 TT |
3460 | /* |
3461 | * The default group preallocation is 512, which for 4k block | |
3462 | * sizes translates to 2 megabytes. However for bigalloc file | |
3463 | * systems, this is probably too big (i.e, if the cluster size | |
3464 | * is 1 megabyte, then group preallocation size becomes half a | |
3465 | * gigabyte!). As a default, we will keep a two megabyte | |
3466 | * group pralloc size for cluster sizes up to 64k, and after | |
3467 | * that, we will force a minimum group preallocation size of | |
3468 | * 32 clusters. This translates to 8 megs when the cluster | |
3469 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
3470 | * which seems reasonable as a default. | |
3471 | */ | |
3472 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
3473 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
3474 | /* |
3475 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
3476 | * to the lowest multiple of s_stripe which is bigger than | |
3477 | * the s_mb_group_prealloc as determined above. We want | |
3478 | * the preallocation size to be an exact multiple of the | |
3479 | * RAID stripe size so that preallocations don't fragment | |
3480 | * the stripes. | |
3481 | */ | |
3482 | if (sbi->s_stripe > 1) { | |
3483 | sbi->s_mb_group_prealloc = roundup( | |
3484 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
3485 | } | |
c9de560d | 3486 | |
730c213c | 3487 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 3488 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 3489 | ret = -ENOMEM; |
029b10c5 | 3490 | goto out; |
c9de560d | 3491 | } |
730c213c | 3492 | for_each_possible_cpu(i) { |
c9de560d | 3493 | struct ext4_locality_group *lg; |
730c213c | 3494 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 3495 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
3496 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
3497 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
3498 | spin_lock_init(&lg->lg_prealloc_lock); |
3499 | } | |
3500 | ||
196e402a HS |
3501 | if (blk_queue_nonrot(bdev_get_queue(sb->s_bdev))) |
3502 | sbi->s_mb_max_linear_groups = 0; | |
3503 | else | |
3504 | sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; | |
79a77c5a YJ |
3505 | /* init file for buddy data */ |
3506 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
3507 | if (ret != 0) |
3508 | goto out_free_locality_groups; | |
79a77c5a | 3509 | |
7aa0baea TM |
3510 | return 0; |
3511 | ||
3512 | out_free_locality_groups: | |
3513 | free_percpu(sbi->s_locality_groups); | |
3514 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 3515 | out: |
196e402a HS |
3516 | kfree(sbi->s_mb_largest_free_orders); |
3517 | kfree(sbi->s_mb_largest_free_orders_locks); | |
7aa0baea TM |
3518 | kfree(sbi->s_mb_offsets); |
3519 | sbi->s_mb_offsets = NULL; | |
3520 | kfree(sbi->s_mb_maxs); | |
3521 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 3522 | return ret; |
c9de560d AT |
3523 | } |
3524 | ||
955ce5f5 | 3525 | /* need to called with the ext4 group lock held */ |
d3df1453 | 3526 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
c9de560d AT |
3527 | { |
3528 | struct ext4_prealloc_space *pa; | |
3529 | struct list_head *cur, *tmp; | |
3530 | int count = 0; | |
3531 | ||
3532 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
3533 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3534 | list_del(&pa->pa_group_list); | |
3535 | count++; | |
688f05a0 | 3536 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d | 3537 | } |
d3df1453 | 3538 | return count; |
c9de560d AT |
3539 | } |
3540 | ||
3541 | int ext4_mb_release(struct super_block *sb) | |
3542 | { | |
8df9675f | 3543 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
3544 | ext4_group_t i; |
3545 | int num_meta_group_infos; | |
df3da4ea | 3546 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 3547 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 3548 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
d3df1453 | 3549 | int count; |
c9de560d | 3550 | |
55cdd0af WJ |
3551 | if (test_opt(sb, DISCARD)) { |
3552 | /* | |
3553 | * wait the discard work to drain all of ext4_free_data | |
3554 | */ | |
3555 | flush_work(&sbi->s_discard_work); | |
3556 | WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); | |
3557 | } | |
3558 | ||
c9de560d | 3559 | if (sbi->s_group_info) { |
8df9675f | 3560 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3561 | cond_resched(); |
c9de560d | 3562 | grinfo = ext4_get_group_info(sb, i); |
a3450215 | 3563 | mb_group_bb_bitmap_free(grinfo); |
c9de560d | 3564 | ext4_lock_group(sb, i); |
d3df1453 RH |
3565 | count = ext4_mb_cleanup_pa(grinfo); |
3566 | if (count) | |
3567 | mb_debug(sb, "mballoc: %d PAs left\n", | |
3568 | count); | |
c9de560d | 3569 | ext4_unlock_group(sb, i); |
fb1813f4 | 3570 | kmem_cache_free(cachep, grinfo); |
c9de560d | 3571 | } |
8df9675f | 3572 | num_meta_group_infos = (ngroups + |
c9de560d AT |
3573 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
3574 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
3575 | rcu_read_lock(); |
3576 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 3577 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
3578 | kfree(group_info[i]); |
3579 | kvfree(group_info); | |
3580 | rcu_read_unlock(); | |
c9de560d | 3581 | } |
196e402a HS |
3582 | kfree(sbi->s_mb_largest_free_orders); |
3583 | kfree(sbi->s_mb_largest_free_orders_locks); | |
c9de560d AT |
3584 | kfree(sbi->s_mb_offsets); |
3585 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 3586 | iput(sbi->s_buddy_cache); |
c9de560d | 3587 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
3588 | ext4_msg(sb, KERN_INFO, |
3589 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
3590 | atomic_read(&sbi->s_bal_allocated), |
3591 | atomic_read(&sbi->s_bal_reqs), | |
3592 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 | 3593 | ext4_msg(sb, KERN_INFO, |
a6c75eaf | 3594 | "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " |
9d8b9ec4 | 3595 | "%u 2^N hits, %u breaks, %u lost", |
c9de560d | 3596 | atomic_read(&sbi->s_bal_ex_scanned), |
a6c75eaf | 3597 | atomic_read(&sbi->s_bal_groups_scanned), |
c9de560d AT |
3598 | atomic_read(&sbi->s_bal_goals), |
3599 | atomic_read(&sbi->s_bal_2orders), | |
3600 | atomic_read(&sbi->s_bal_breaks), | |
3601 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 | 3602 | ext4_msg(sb, KERN_INFO, |
67d25186 HS |
3603 | "mballoc: %u generated and it took %llu", |
3604 | atomic_read(&sbi->s_mb_buddies_generated), | |
3605 | atomic64_read(&sbi->s_mb_generation_time)); | |
9d8b9ec4 TT |
3606 | ext4_msg(sb, KERN_INFO, |
3607 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
3608 | atomic_read(&sbi->s_mb_preallocated), |
3609 | atomic_read(&sbi->s_mb_discarded)); | |
3610 | } | |
3611 | ||
730c213c | 3612 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
3613 | |
3614 | return 0; | |
3615 | } | |
3616 | ||
77ca6cdf | 3617 | static inline int ext4_issue_discard(struct super_block *sb, |
a0154344 DJ |
3618 | ext4_group_t block_group, ext4_grpblk_t cluster, int count, |
3619 | struct bio **biop) | |
5c521830 | 3620 | { |
5c521830 JZ |
3621 | ext4_fsblk_t discard_block; |
3622 | ||
84130193 TT |
3623 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
3624 | ext4_group_first_block_no(sb, block_group)); | |
3625 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
3626 | trace_ext4_discard_blocks(sb, |
3627 | (unsigned long long) discard_block, count); | |
a0154344 DJ |
3628 | if (biop) { |
3629 | return __blkdev_issue_discard(sb->s_bdev, | |
3630 | (sector_t)discard_block << (sb->s_blocksize_bits - 9), | |
3631 | (sector_t)count << (sb->s_blocksize_bits - 9), | |
3632 | GFP_NOFS, 0, biop); | |
3633 | } else | |
3634 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
3635 | } |
3636 | ||
a0154344 DJ |
3637 | static void ext4_free_data_in_buddy(struct super_block *sb, |
3638 | struct ext4_free_data *entry) | |
c9de560d | 3639 | { |
c9de560d | 3640 | struct ext4_buddy e4b; |
c894058d | 3641 | struct ext4_group_info *db; |
d9f34504 | 3642 | int err, count = 0, count2 = 0; |
c9de560d | 3643 | |
d3df1453 | 3644 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
18aadd47 | 3645 | entry->efd_count, entry->efd_group, entry); |
c9de560d | 3646 | |
18aadd47 BJ |
3647 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
3648 | /* we expect to find existing buddy because it's pinned */ | |
3649 | BUG_ON(err != 0); | |
b90f6870 | 3650 | |
d08854f5 TT |
3651 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
3652 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
3653 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 3654 | |
18aadd47 BJ |
3655 | db = e4b.bd_info; |
3656 | /* there are blocks to put in buddy to make them really free */ | |
3657 | count += entry->efd_count; | |
3658 | count2++; | |
3659 | ext4_lock_group(sb, entry->efd_group); | |
3660 | /* Take it out of per group rb tree */ | |
3661 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
3662 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 3663 | |
18aadd47 BJ |
3664 | /* |
3665 | * Clear the trimmed flag for the group so that the next | |
3666 | * ext4_trim_fs can trim it. | |
3667 | * If the volume is mounted with -o discard, online discard | |
3668 | * is supported and the free blocks will be trimmed online. | |
3669 | */ | |
3670 | if (!test_opt(sb, DISCARD)) | |
3671 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 3672 | |
18aadd47 BJ |
3673 | if (!db->bb_free_root.rb_node) { |
3674 | /* No more items in the per group rb tree | |
3675 | * balance refcounts from ext4_mb_free_metadata() | |
3676 | */ | |
09cbfeaf KS |
3677 | put_page(e4b.bd_buddy_page); |
3678 | put_page(e4b.bd_bitmap_page); | |
3e624fc7 | 3679 | } |
18aadd47 | 3680 | ext4_unlock_group(sb, entry->efd_group); |
18aadd47 | 3681 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 3682 | |
d3df1453 RH |
3683 | mb_debug(sb, "freed %d blocks in %d structures\n", count, |
3684 | count2); | |
c9de560d AT |
3685 | } |
3686 | ||
a0154344 DJ |
3687 | /* |
3688 | * This function is called by the jbd2 layer once the commit has finished, | |
3689 | * so we know we can free the blocks that were released with that commit. | |
3690 | */ | |
3691 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
3692 | { | |
3693 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3694 | struct ext4_free_data *entry, *tmp; | |
a0154344 DJ |
3695 | struct list_head freed_data_list; |
3696 | struct list_head *cut_pos = NULL; | |
55cdd0af | 3697 | bool wake; |
a0154344 DJ |
3698 | |
3699 | INIT_LIST_HEAD(&freed_data_list); | |
3700 | ||
3701 | spin_lock(&sbi->s_md_lock); | |
3702 | list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { | |
3703 | if (entry->efd_tid != commit_tid) | |
3704 | break; | |
3705 | cut_pos = &entry->efd_list; | |
3706 | } | |
3707 | if (cut_pos) | |
3708 | list_cut_position(&freed_data_list, &sbi->s_freed_data_list, | |
3709 | cut_pos); | |
3710 | spin_unlock(&sbi->s_md_lock); | |
3711 | ||
55cdd0af WJ |
3712 | list_for_each_entry(entry, &freed_data_list, efd_list) |
3713 | ext4_free_data_in_buddy(sb, entry); | |
a0154344 | 3714 | |
55cdd0af WJ |
3715 | if (test_opt(sb, DISCARD)) { |
3716 | spin_lock(&sbi->s_md_lock); | |
3717 | wake = list_empty(&sbi->s_discard_list); | |
3718 | list_splice_tail(&freed_data_list, &sbi->s_discard_list); | |
3719 | spin_unlock(&sbi->s_md_lock); | |
3720 | if (wake) | |
3721 | queue_work(system_unbound_wq, &sbi->s_discard_work); | |
3722 | } else { | |
3723 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
3724 | kmem_cache_free(ext4_free_data_cachep, entry); | |
a0154344 | 3725 | } |
a0154344 DJ |
3726 | } |
3727 | ||
5dabfc78 | 3728 | int __init ext4_init_mballoc(void) |
c9de560d | 3729 | { |
16828088 TT |
3730 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
3731 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 3732 | if (ext4_pspace_cachep == NULL) |
f283529a | 3733 | goto out; |
c9de560d | 3734 | |
16828088 TT |
3735 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
3736 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3737 | if (ext4_ac_cachep == NULL) |
3738 | goto out_pa_free; | |
c894058d | 3739 | |
18aadd47 BJ |
3740 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
3741 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3742 | if (ext4_free_data_cachep == NULL) |
3743 | goto out_ac_free; | |
3744 | ||
c9de560d | 3745 | return 0; |
f283529a RH |
3746 | |
3747 | out_ac_free: | |
3748 | kmem_cache_destroy(ext4_ac_cachep); | |
3749 | out_pa_free: | |
3750 | kmem_cache_destroy(ext4_pspace_cachep); | |
3751 | out: | |
3752 | return -ENOMEM; | |
c9de560d AT |
3753 | } |
3754 | ||
5dabfc78 | 3755 | void ext4_exit_mballoc(void) |
c9de560d | 3756 | { |
60e6679e | 3757 | /* |
3e03f9ca JDB |
3758 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
3759 | * before destroying the slab cache. | |
3760 | */ | |
3761 | rcu_barrier(); | |
c9de560d | 3762 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 3763 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 3764 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 3765 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
3766 | } |
3767 | ||
3768 | ||
3769 | /* | |
73b2c716 | 3770 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
3771 | * Returns 0 if success or error code |
3772 | */ | |
4ddfef7b ES |
3773 | static noinline_for_stack int |
3774 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 3775 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
3776 | { |
3777 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
3778 | struct ext4_group_desc *gdp; |
3779 | struct buffer_head *gdp_bh; | |
3780 | struct ext4_sb_info *sbi; | |
3781 | struct super_block *sb; | |
3782 | ext4_fsblk_t block; | |
519deca0 | 3783 | int err, len; |
c9de560d AT |
3784 | |
3785 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3786 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
3787 | ||
3788 | sb = ac->ac_sb; | |
3789 | sbi = EXT4_SB(sb); | |
c9de560d | 3790 | |
574ca174 | 3791 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
9008a58e DW |
3792 | if (IS_ERR(bitmap_bh)) { |
3793 | err = PTR_ERR(bitmap_bh); | |
3794 | bitmap_bh = NULL; | |
c9de560d | 3795 | goto out_err; |
9008a58e | 3796 | } |
c9de560d | 3797 | |
5d601255 | 3798 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
3799 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
3800 | EXT4_JTR_NONE); | |
c9de560d AT |
3801 | if (err) |
3802 | goto out_err; | |
3803 | ||
3804 | err = -EIO; | |
3805 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
3806 | if (!gdp) | |
3807 | goto out_err; | |
3808 | ||
a9df9a49 | 3809 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 3810 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 3811 | |
5d601255 | 3812 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
188c299e | 3813 | err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE); |
c9de560d AT |
3814 | if (err) |
3815 | goto out_err; | |
3816 | ||
bda00de7 | 3817 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 3818 | |
53accfa9 | 3819 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
ce9f24cc | 3820 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
12062ddd | 3821 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 3822 | "fs metadata", block, block+len); |
519deca0 | 3823 | /* File system mounted not to panic on error |
554a5ccc | 3824 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
3825 | * We leak some of the blocks here. |
3826 | */ | |
955ce5f5 | 3827 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
3828 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3829 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 3830 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 3831 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 | 3832 | if (!err) |
554a5ccc | 3833 | err = -EFSCORRUPTED; |
519deca0 | 3834 | goto out_err; |
c9de560d | 3835 | } |
955ce5f5 AK |
3836 | |
3837 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
3838 | #ifdef AGGRESSIVE_CHECK |
3839 | { | |
3840 | int i; | |
3841 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
3842 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
3843 | bitmap_bh->b_data)); | |
3844 | } | |
3845 | } | |
3846 | #endif | |
c3e94d1d YY |
3847 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
3848 | ac->ac_b_ex.fe_len); | |
8844618d TT |
3849 | if (ext4_has_group_desc_csum(sb) && |
3850 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
c9de560d | 3851 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
021b65bb | 3852 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 3853 | ext4_free_clusters_after_init(sb, |
021b65bb | 3854 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 3855 | } |
021b65bb TT |
3856 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
3857 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 3858 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 3859 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
3860 | |
3861 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 3862 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 3863 | /* |
6bc6e63f | 3864 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 3865 | */ |
6bc6e63f AK |
3866 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
3867 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
3868 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
3869 | reserv_clstrs); | |
c9de560d | 3870 | |
772cb7c8 JS |
3871 | if (sbi->s_log_groups_per_flex) { |
3872 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
3873 | ac->ac_b_ex.fe_group); | |
90ba983f | 3874 | atomic64_sub(ac->ac_b_ex.fe_len, |
7c990728 SJS |
3875 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
3876 | flex_group)->free_clusters); | |
772cb7c8 JS |
3877 | } |
3878 | ||
0390131b | 3879 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
3880 | if (err) |
3881 | goto out_err; | |
0390131b | 3882 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
3883 | |
3884 | out_err: | |
42a10add | 3885 | brelse(bitmap_bh); |
c9de560d AT |
3886 | return err; |
3887 | } | |
3888 | ||
8016e29f HS |
3889 | /* |
3890 | * Idempotent helper for Ext4 fast commit replay path to set the state of | |
3891 | * blocks in bitmaps and update counters. | |
3892 | */ | |
3893 | void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, | |
3894 | int len, int state) | |
3895 | { | |
3896 | struct buffer_head *bitmap_bh = NULL; | |
3897 | struct ext4_group_desc *gdp; | |
3898 | struct buffer_head *gdp_bh; | |
3899 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3900 | ext4_group_t group; | |
3901 | ext4_grpblk_t blkoff; | |
3902 | int i, clen, err; | |
3903 | int already; | |
3904 | ||
3905 | clen = EXT4_B2C(sbi, len); | |
3906 | ||
3907 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
3908 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
3909 | if (IS_ERR(bitmap_bh)) { | |
3910 | err = PTR_ERR(bitmap_bh); | |
3911 | bitmap_bh = NULL; | |
3912 | goto out_err; | |
3913 | } | |
3914 | ||
3915 | err = -EIO; | |
3916 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
3917 | if (!gdp) | |
3918 | goto out_err; | |
3919 | ||
3920 | ext4_lock_group(sb, group); | |
3921 | already = 0; | |
3922 | for (i = 0; i < clen; i++) | |
3923 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == !state) | |
3924 | already++; | |
3925 | ||
3926 | if (state) | |
3927 | ext4_set_bits(bitmap_bh->b_data, blkoff, clen); | |
3928 | else | |
3929 | mb_test_and_clear_bits(bitmap_bh->b_data, blkoff, clen); | |
3930 | if (ext4_has_group_desc_csum(sb) && | |
3931 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
3932 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
3933 | ext4_free_group_clusters_set(sb, gdp, | |
3934 | ext4_free_clusters_after_init(sb, | |
3935 | group, gdp)); | |
3936 | } | |
3937 | if (state) | |
3938 | clen = ext4_free_group_clusters(sb, gdp) - clen + already; | |
3939 | else | |
3940 | clen = ext4_free_group_clusters(sb, gdp) + clen - already; | |
3941 | ||
3942 | ext4_free_group_clusters_set(sb, gdp, clen); | |
3943 | ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh); | |
3944 | ext4_group_desc_csum_set(sb, group, gdp); | |
3945 | ||
3946 | ext4_unlock_group(sb, group); | |
3947 | ||
3948 | if (sbi->s_log_groups_per_flex) { | |
3949 | ext4_group_t flex_group = ext4_flex_group(sbi, group); | |
3950 | ||
3951 | atomic64_sub(len, | |
3952 | &sbi_array_rcu_deref(sbi, s_flex_groups, | |
3953 | flex_group)->free_clusters); | |
3954 | } | |
3955 | ||
3956 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
3957 | if (err) | |
3958 | goto out_err; | |
3959 | sync_dirty_buffer(bitmap_bh); | |
3960 | err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
3961 | sync_dirty_buffer(gdp_bh); | |
3962 | ||
3963 | out_err: | |
3964 | brelse(bitmap_bh); | |
3965 | } | |
3966 | ||
c9de560d AT |
3967 | /* |
3968 | * here we normalize request for locality group | |
d7a1fee1 DE |
3969 | * Group request are normalized to s_mb_group_prealloc, which goes to |
3970 | * s_strip if we set the same via mount option. | |
3971 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 3972 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
3973 | * |
3974 | * XXX: should we try to preallocate more than the group has now? | |
3975 | */ | |
3976 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
3977 | { | |
3978 | struct super_block *sb = ac->ac_sb; | |
3979 | struct ext4_locality_group *lg = ac->ac_lg; | |
3980 | ||
3981 | BUG_ON(lg == NULL); | |
d7a1fee1 | 3982 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
d3df1453 | 3983 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
c9de560d AT |
3984 | } |
3985 | ||
3986 | /* | |
3987 | * Normalization means making request better in terms of | |
3988 | * size and alignment | |
3989 | */ | |
4ddfef7b ES |
3990 | static noinline_for_stack void |
3991 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3992 | struct ext4_allocation_request *ar) |
3993 | { | |
53accfa9 | 3994 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3995 | int bsbits, max; |
3996 | ext4_lblk_t end; | |
1592d2c5 CW |
3997 | loff_t size, start_off; |
3998 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3999 | ext4_lblk_t start; |
c9de560d | 4000 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 4001 | struct ext4_prealloc_space *pa; |
c9de560d AT |
4002 | |
4003 | /* do normalize only data requests, metadata requests | |
4004 | do not need preallocation */ | |
4005 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4006 | return; | |
4007 | ||
4008 | /* sometime caller may want exact blocks */ | |
4009 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4010 | return; | |
4011 | ||
4012 | /* caller may indicate that preallocation isn't | |
4013 | * required (it's a tail, for example) */ | |
4014 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
4015 | return; | |
4016 | ||
4017 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
4018 | ext4_mb_normalize_group_request(ac); | |
4019 | return ; | |
4020 | } | |
4021 | ||
4022 | bsbits = ac->ac_sb->s_blocksize_bits; | |
4023 | ||
4024 | /* first, let's learn actual file size | |
4025 | * given current request is allocated */ | |
53accfa9 | 4026 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
4027 | size = size << bsbits; |
4028 | if (size < i_size_read(ac->ac_inode)) | |
4029 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 4030 | orig_size = size; |
c9de560d | 4031 | |
1930479c VC |
4032 | /* max size of free chunks */ |
4033 | max = 2 << bsbits; | |
c9de560d | 4034 | |
1930479c VC |
4035 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
4036 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
4037 | |
4038 | /* first, try to predict filesize */ | |
4039 | /* XXX: should this table be tunable? */ | |
4040 | start_off = 0; | |
4041 | if (size <= 16 * 1024) { | |
4042 | size = 16 * 1024; | |
4043 | } else if (size <= 32 * 1024) { | |
4044 | size = 32 * 1024; | |
4045 | } else if (size <= 64 * 1024) { | |
4046 | size = 64 * 1024; | |
4047 | } else if (size <= 128 * 1024) { | |
4048 | size = 128 * 1024; | |
4049 | } else if (size <= 256 * 1024) { | |
4050 | size = 256 * 1024; | |
4051 | } else if (size <= 512 * 1024) { | |
4052 | size = 512 * 1024; | |
4053 | } else if (size <= 1024 * 1024) { | |
4054 | size = 1024 * 1024; | |
1930479c | 4055 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 4056 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
4057 | (21 - bsbits)) << 21; |
4058 | size = 2 * 1024 * 1024; | |
4059 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
4060 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4061 | (22 - bsbits)) << 22; | |
4062 | size = 4 * 1024 * 1024; | |
4063 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 4064 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
4065 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4066 | (23 - bsbits)) << 23; | |
4067 | size = 8 * 1024 * 1024; | |
4068 | } else { | |
b27b1535 XW |
4069 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
4070 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
4071 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 4072 | } |
5a0790c2 AK |
4073 | size = size >> bsbits; |
4074 | start = start_off >> bsbits; | |
c9de560d AT |
4075 | |
4076 | /* don't cover already allocated blocks in selected range */ | |
4077 | if (ar->pleft && start <= ar->lleft) { | |
4078 | size -= ar->lleft + 1 - start; | |
4079 | start = ar->lleft + 1; | |
4080 | } | |
4081 | if (ar->pright && start + size - 1 >= ar->lright) | |
4082 | size -= start + size - ar->lright; | |
4083 | ||
cd648b8a JK |
4084 | /* |
4085 | * Trim allocation request for filesystems with artificially small | |
4086 | * groups. | |
4087 | */ | |
4088 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
4089 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
4090 | ||
c9de560d AT |
4091 | end = start + size; |
4092 | ||
4093 | /* check we don't cross already preallocated blocks */ | |
4094 | rcu_read_lock(); | |
9a0762c5 | 4095 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 4096 | ext4_lblk_t pa_end; |
c9de560d | 4097 | |
c9de560d AT |
4098 | if (pa->pa_deleted) |
4099 | continue; | |
4100 | spin_lock(&pa->pa_lock); | |
4101 | if (pa->pa_deleted) { | |
4102 | spin_unlock(&pa->pa_lock); | |
4103 | continue; | |
4104 | } | |
4105 | ||
53accfa9 TT |
4106 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
4107 | pa->pa_len); | |
c9de560d AT |
4108 | |
4109 | /* PA must not overlap original request */ | |
4110 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
4111 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
4112 | ||
38877f4e ES |
4113 | /* skip PAs this normalized request doesn't overlap with */ |
4114 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
4115 | spin_unlock(&pa->pa_lock); |
4116 | continue; | |
4117 | } | |
4118 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
4119 | ||
38877f4e | 4120 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
4121 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
4122 | BUG_ON(pa_end < start); | |
4123 | start = pa_end; | |
38877f4e | 4124 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
4125 | BUG_ON(pa->pa_lstart > end); |
4126 | end = pa->pa_lstart; | |
4127 | } | |
4128 | spin_unlock(&pa->pa_lock); | |
4129 | } | |
4130 | rcu_read_unlock(); | |
4131 | size = end - start; | |
4132 | ||
4133 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
4134 | rcu_read_lock(); | |
9a0762c5 | 4135 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 4136 | ext4_lblk_t pa_end; |
53accfa9 | 4137 | |
c9de560d AT |
4138 | spin_lock(&pa->pa_lock); |
4139 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
4140 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
4141 | pa->pa_len); | |
c9de560d AT |
4142 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
4143 | } | |
4144 | spin_unlock(&pa->pa_lock); | |
4145 | } | |
4146 | rcu_read_unlock(); | |
4147 | ||
4148 | if (start + size <= ac->ac_o_ex.fe_logical && | |
4149 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
4150 | ext4_msg(ac->ac_sb, KERN_ERR, |
4151 | "start %lu, size %lu, fe_logical %lu", | |
4152 | (unsigned long) start, (unsigned long) size, | |
4153 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 4154 | BUG(); |
c9de560d | 4155 | } |
b5b60778 | 4156 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
4157 | |
4158 | /* now prepare goal request */ | |
4159 | ||
4160 | /* XXX: is it better to align blocks WRT to logical | |
4161 | * placement or satisfy big request as is */ | |
4162 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 4163 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
4164 | |
4165 | /* define goal start in order to merge */ | |
4166 | if (ar->pright && (ar->lright == (start + size))) { | |
4167 | /* merge to the right */ | |
4168 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
4169 | &ac->ac_f_ex.fe_group, | |
4170 | &ac->ac_f_ex.fe_start); | |
4171 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
4172 | } | |
4173 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
4174 | /* merge to the left */ | |
4175 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
4176 | &ac->ac_f_ex.fe_group, | |
4177 | &ac->ac_f_ex.fe_start); | |
4178 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
4179 | } | |
4180 | ||
d3df1453 RH |
4181 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
4182 | orig_size, start); | |
c9de560d AT |
4183 | } |
4184 | ||
4185 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
4186 | { | |
4187 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4188 | ||
a6c75eaf | 4189 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { |
c9de560d AT |
4190 | atomic_inc(&sbi->s_bal_reqs); |
4191 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 4192 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
4193 | atomic_inc(&sbi->s_bal_success); |
4194 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
a6c75eaf | 4195 | atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); |
c9de560d AT |
4196 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && |
4197 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
4198 | atomic_inc(&sbi->s_bal_goals); | |
4199 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
4200 | atomic_inc(&sbi->s_bal_breaks); | |
4201 | } | |
4202 | ||
296c355c TT |
4203 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
4204 | trace_ext4_mballoc_alloc(ac); | |
4205 | else | |
4206 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
4207 | } |
4208 | ||
b844167e CW |
4209 | /* |
4210 | * Called on failure; free up any blocks from the inode PA for this | |
4211 | * context. We don't need this for MB_GROUP_PA because we only change | |
4212 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
4213 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
4214 | */ | |
4215 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
4216 | { | |
4217 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
4218 | struct ext4_buddy e4b; |
4219 | int err; | |
b844167e | 4220 | |
86f0afd4 | 4221 | if (pa == NULL) { |
c99d1e6e TT |
4222 | if (ac->ac_f_ex.fe_len == 0) |
4223 | return; | |
86f0afd4 TT |
4224 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
4225 | if (err) { | |
4226 | /* | |
4227 | * This should never happen since we pin the | |
4228 | * pages in the ext4_allocation_context so | |
4229 | * ext4_mb_load_buddy() should never fail. | |
4230 | */ | |
4231 | WARN(1, "mb_load_buddy failed (%d)", err); | |
4232 | return; | |
4233 | } | |
4234 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
4235 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
4236 | ac->ac_f_ex.fe_len); | |
4237 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 4238 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
4239 | return; |
4240 | } | |
4241 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 4242 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
4243 | } |
4244 | ||
c9de560d AT |
4245 | /* |
4246 | * use blocks preallocated to inode | |
4247 | */ | |
4248 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
4249 | struct ext4_prealloc_space *pa) | |
4250 | { | |
53accfa9 | 4251 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
4252 | ext4_fsblk_t start; |
4253 | ext4_fsblk_t end; | |
4254 | int len; | |
4255 | ||
4256 | /* found preallocated blocks, use them */ | |
4257 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
4258 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
4259 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
4260 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
4261 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
4262 | &ac->ac_b_ex.fe_start); | |
4263 | ac->ac_b_ex.fe_len = len; | |
4264 | ac->ac_status = AC_STATUS_FOUND; | |
4265 | ac->ac_pa = pa; | |
4266 | ||
4267 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 4268 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
4269 | BUG_ON(pa->pa_free < len); |
4270 | pa->pa_free -= len; | |
4271 | ||
d3df1453 | 4272 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
4273 | } |
4274 | ||
4275 | /* | |
4276 | * use blocks preallocated to locality group | |
4277 | */ | |
4278 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
4279 | struct ext4_prealloc_space *pa) | |
4280 | { | |
03cddb80 | 4281 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 4282 | |
c9de560d AT |
4283 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
4284 | &ac->ac_b_ex.fe_group, | |
4285 | &ac->ac_b_ex.fe_start); | |
4286 | ac->ac_b_ex.fe_len = len; | |
4287 | ac->ac_status = AC_STATUS_FOUND; | |
4288 | ac->ac_pa = pa; | |
4289 | ||
4290 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 4291 | * possible race when the group is being loaded concurrently |
c9de560d | 4292 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
4293 | * in on-disk bitmap -- see ext4_mb_release_context() |
4294 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 4295 | */ |
d3df1453 RH |
4296 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
4297 | pa->pa_lstart-len, len, pa); | |
c9de560d AT |
4298 | } |
4299 | ||
5e745b04 AK |
4300 | /* |
4301 | * Return the prealloc space that have minimal distance | |
4302 | * from the goal block. @cpa is the prealloc | |
4303 | * space that is having currently known minimal distance | |
4304 | * from the goal block. | |
4305 | */ | |
4306 | static struct ext4_prealloc_space * | |
4307 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
4308 | struct ext4_prealloc_space *pa, | |
4309 | struct ext4_prealloc_space *cpa) | |
4310 | { | |
4311 | ext4_fsblk_t cur_distance, new_distance; | |
4312 | ||
4313 | if (cpa == NULL) { | |
4314 | atomic_inc(&pa->pa_count); | |
4315 | return pa; | |
4316 | } | |
79211c8e AM |
4317 | cur_distance = abs(goal_block - cpa->pa_pstart); |
4318 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 4319 | |
5a54b2f1 | 4320 | if (cur_distance <= new_distance) |
5e745b04 AK |
4321 | return cpa; |
4322 | ||
4323 | /* drop the previous reference */ | |
4324 | atomic_dec(&cpa->pa_count); | |
4325 | atomic_inc(&pa->pa_count); | |
4326 | return pa; | |
4327 | } | |
4328 | ||
c9de560d AT |
4329 | /* |
4330 | * search goal blocks in preallocated space | |
4331 | */ | |
4fca8f07 | 4332 | static noinline_for_stack bool |
4ddfef7b | 4333 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 4334 | { |
53accfa9 | 4335 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 4336 | int order, i; |
c9de560d AT |
4337 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
4338 | struct ext4_locality_group *lg; | |
5e745b04 AK |
4339 | struct ext4_prealloc_space *pa, *cpa = NULL; |
4340 | ext4_fsblk_t goal_block; | |
c9de560d AT |
4341 | |
4342 | /* only data can be preallocated */ | |
4343 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 4344 | return false; |
c9de560d AT |
4345 | |
4346 | /* first, try per-file preallocation */ | |
4347 | rcu_read_lock(); | |
9a0762c5 | 4348 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
4349 | |
4350 | /* all fields in this condition don't change, | |
4351 | * so we can skip locking for them */ | |
4352 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
4353 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
4354 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
4355 | continue; |
4356 | ||
fb0a387d | 4357 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 4358 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
4359 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
4360 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
4361 | continue; |
4362 | ||
c9de560d AT |
4363 | /* found preallocated blocks, use them */ |
4364 | spin_lock(&pa->pa_lock); | |
4365 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
4366 | atomic_inc(&pa->pa_count); | |
4367 | ext4_mb_use_inode_pa(ac, pa); | |
4368 | spin_unlock(&pa->pa_lock); | |
4369 | ac->ac_criteria = 10; | |
4370 | rcu_read_unlock(); | |
4fca8f07 | 4371 | return true; |
c9de560d AT |
4372 | } |
4373 | spin_unlock(&pa->pa_lock); | |
4374 | } | |
4375 | rcu_read_unlock(); | |
4376 | ||
4377 | /* can we use group allocation? */ | |
4378 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 4379 | return false; |
c9de560d AT |
4380 | |
4381 | /* inode may have no locality group for some reason */ | |
4382 | lg = ac->ac_lg; | |
4383 | if (lg == NULL) | |
4fca8f07 | 4384 | return false; |
6be2ded1 AK |
4385 | order = fls(ac->ac_o_ex.fe_len) - 1; |
4386 | if (order > PREALLOC_TB_SIZE - 1) | |
4387 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4388 | order = PREALLOC_TB_SIZE - 1; | |
4389 | ||
bda00de7 | 4390 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
4391 | /* |
4392 | * search for the prealloc space that is having | |
4393 | * minimal distance from the goal block. | |
4394 | */ | |
6be2ded1 AK |
4395 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
4396 | rcu_read_lock(); | |
4397 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
4398 | pa_inode_list) { | |
4399 | spin_lock(&pa->pa_lock); | |
4400 | if (pa->pa_deleted == 0 && | |
4401 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
4402 | |
4403 | cpa = ext4_mb_check_group_pa(goal_block, | |
4404 | pa, cpa); | |
6be2ded1 | 4405 | } |
c9de560d | 4406 | spin_unlock(&pa->pa_lock); |
c9de560d | 4407 | } |
6be2ded1 | 4408 | rcu_read_unlock(); |
c9de560d | 4409 | } |
5e745b04 AK |
4410 | if (cpa) { |
4411 | ext4_mb_use_group_pa(ac, cpa); | |
4412 | ac->ac_criteria = 20; | |
4fca8f07 | 4413 | return true; |
5e745b04 | 4414 | } |
4fca8f07 | 4415 | return false; |
c9de560d AT |
4416 | } |
4417 | ||
7a2fcbf7 AK |
4418 | /* |
4419 | * the function goes through all block freed in the group | |
4420 | * but not yet committed and marks them used in in-core bitmap. | |
4421 | * buddy must be generated from this bitmap | |
955ce5f5 | 4422 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
4423 | */ |
4424 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
4425 | ext4_group_t group) | |
4426 | { | |
4427 | struct rb_node *n; | |
4428 | struct ext4_group_info *grp; | |
4429 | struct ext4_free_data *entry; | |
4430 | ||
4431 | grp = ext4_get_group_info(sb, group); | |
4432 | n = rb_first(&(grp->bb_free_root)); | |
4433 | ||
4434 | while (n) { | |
18aadd47 BJ |
4435 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
4436 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
4437 | n = rb_next(n); |
4438 | } | |
4439 | return; | |
4440 | } | |
4441 | ||
c9de560d AT |
4442 | /* |
4443 | * the function goes through all preallocation in this group and marks them | |
4444 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 4445 | * Need to be called with ext4 group lock held |
c9de560d | 4446 | */ |
089ceecc ES |
4447 | static noinline_for_stack |
4448 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
4449 | ext4_group_t group) |
4450 | { | |
4451 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4452 | struct ext4_prealloc_space *pa; | |
4453 | struct list_head *cur; | |
4454 | ext4_group_t groupnr; | |
4455 | ext4_grpblk_t start; | |
4456 | int preallocated = 0; | |
c9de560d AT |
4457 | int len; |
4458 | ||
4459 | /* all form of preallocation discards first load group, | |
4460 | * so the only competing code is preallocation use. | |
4461 | * we don't need any locking here | |
4462 | * notice we do NOT ignore preallocations with pa_deleted | |
4463 | * otherwise we could leave used blocks available for | |
4464 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
4465 | * is dropping preallocation | |
4466 | */ | |
4467 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4468 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
4469 | spin_lock(&pa->pa_lock); | |
4470 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4471 | &groupnr, &start); | |
4472 | len = pa->pa_len; | |
4473 | spin_unlock(&pa->pa_lock); | |
4474 | if (unlikely(len == 0)) | |
4475 | continue; | |
4476 | BUG_ON(groupnr != group); | |
c3e94d1d | 4477 | ext4_set_bits(bitmap, start, len); |
c9de560d | 4478 | preallocated += len; |
c9de560d | 4479 | } |
d3df1453 | 4480 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
4481 | } |
4482 | ||
27bc446e | 4483 | static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
4484 | struct ext4_prealloc_space *pa) | |
4485 | { | |
4486 | struct ext4_inode_info *ei; | |
4487 | ||
4488 | if (pa->pa_deleted) { | |
4489 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", | |
4490 | pa->pa_type, pa->pa_pstart, pa->pa_lstart, | |
4491 | pa->pa_len); | |
4492 | return; | |
4493 | } | |
4494 | ||
4495 | pa->pa_deleted = 1; | |
4496 | ||
4497 | if (pa->pa_type == MB_INODE_PA) { | |
4498 | ei = EXT4_I(pa->pa_inode); | |
4499 | atomic_dec(&ei->i_prealloc_active); | |
4500 | } | |
4501 | } | |
4502 | ||
c9de560d AT |
4503 | static void ext4_mb_pa_callback(struct rcu_head *head) |
4504 | { | |
4505 | struct ext4_prealloc_space *pa; | |
4506 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
4507 | |
4508 | BUG_ON(atomic_read(&pa->pa_count)); | |
4509 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
4510 | kmem_cache_free(ext4_pspace_cachep, pa); |
4511 | } | |
4512 | ||
4513 | /* | |
4514 | * drops a reference to preallocated space descriptor | |
4515 | * if this was the last reference and the space is consumed | |
4516 | */ | |
4517 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
4518 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
4519 | { | |
a9df9a49 | 4520 | ext4_group_t grp; |
d33a1976 | 4521 | ext4_fsblk_t grp_blk; |
c9de560d | 4522 | |
c9de560d AT |
4523 | /* in this short window concurrent discard can set pa_deleted */ |
4524 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
4525 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
4526 | spin_unlock(&pa->pa_lock); | |
4527 | return; | |
4528 | } | |
4529 | ||
c9de560d AT |
4530 | if (pa->pa_deleted == 1) { |
4531 | spin_unlock(&pa->pa_lock); | |
4532 | return; | |
4533 | } | |
4534 | ||
27bc446e | 4535 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4536 | spin_unlock(&pa->pa_lock); |
4537 | ||
d33a1976 | 4538 | grp_blk = pa->pa_pstart; |
60e6679e | 4539 | /* |
cc0fb9ad AK |
4540 | * If doing group-based preallocation, pa_pstart may be in the |
4541 | * next group when pa is used up | |
4542 | */ | |
4543 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
4544 | grp_blk--; |
4545 | ||
bd86298e | 4546 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
4547 | |
4548 | /* | |
4549 | * possible race: | |
4550 | * | |
4551 | * P1 (buddy init) P2 (regular allocation) | |
4552 | * find block B in PA | |
4553 | * copy on-disk bitmap to buddy | |
4554 | * mark B in on-disk bitmap | |
4555 | * drop PA from group | |
4556 | * mark all PAs in buddy | |
4557 | * | |
4558 | * thus, P1 initializes buddy with B available. to prevent this | |
4559 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
4560 | * against that pair | |
4561 | */ | |
4562 | ext4_lock_group(sb, grp); | |
4563 | list_del(&pa->pa_group_list); | |
4564 | ext4_unlock_group(sb, grp); | |
4565 | ||
4566 | spin_lock(pa->pa_obj_lock); | |
4567 | list_del_rcu(&pa->pa_inode_list); | |
4568 | spin_unlock(pa->pa_obj_lock); | |
4569 | ||
4570 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4571 | } | |
4572 | ||
4573 | /* | |
4574 | * creates new preallocated space for given inode | |
4575 | */ | |
53f86b17 | 4576 | static noinline_for_stack void |
4ddfef7b | 4577 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4578 | { |
4579 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 4580 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
4581 | struct ext4_prealloc_space *pa; |
4582 | struct ext4_group_info *grp; | |
4583 | struct ext4_inode_info *ei; | |
4584 | ||
4585 | /* preallocate only when found space is larger then requested */ | |
4586 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4587 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4588 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4589 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4590 | |
53f86b17 | 4591 | pa = ac->ac_pa; |
c9de560d AT |
4592 | |
4593 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
4594 | int winl; | |
4595 | int wins; | |
4596 | int win; | |
4597 | int offs; | |
4598 | ||
4599 | /* we can't allocate as much as normalizer wants. | |
4600 | * so, found space must get proper lstart | |
4601 | * to cover original request */ | |
4602 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
4603 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
4604 | ||
4605 | /* we're limited by original request in that | |
4606 | * logical block must be covered any way | |
4607 | * winl is window we can move our chunk within */ | |
4608 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
4609 | ||
4610 | /* also, we should cover whole original request */ | |
53accfa9 | 4611 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
4612 | |
4613 | /* the smallest one defines real window */ | |
4614 | win = min(winl, wins); | |
4615 | ||
53accfa9 TT |
4616 | offs = ac->ac_o_ex.fe_logical % |
4617 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
4618 | if (offs && offs < win) |
4619 | win = offs; | |
4620 | ||
53accfa9 | 4621 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 4622 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
4623 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
4624 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
4625 | } | |
4626 | ||
4627 | /* preallocation can change ac_b_ex, thus we store actually | |
4628 | * allocated blocks for history */ | |
4629 | ac->ac_f_ex = ac->ac_b_ex; | |
4630 | ||
4631 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
4632 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4633 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4634 | pa->pa_free = pa->pa_len; | |
c9de560d | 4635 | spin_lock_init(&pa->pa_lock); |
d794bf8e AK |
4636 | INIT_LIST_HEAD(&pa->pa_inode_list); |
4637 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 4638 | pa->pa_deleted = 0; |
cc0fb9ad | 4639 | pa->pa_type = MB_INODE_PA; |
c9de560d | 4640 | |
d3df1453 RH |
4641 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4642 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4643 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
4644 | |
4645 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 4646 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
4647 | |
4648 | ei = EXT4_I(ac->ac_inode); | |
4649 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4650 | ||
4651 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
4652 | pa->pa_inode = ac->ac_inode; | |
4653 | ||
c9de560d | 4654 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d AT |
4655 | |
4656 | spin_lock(pa->pa_obj_lock); | |
4657 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
4658 | spin_unlock(pa->pa_obj_lock); | |
27bc446e | 4659 | atomic_inc(&ei->i_prealloc_active); |
c9de560d AT |
4660 | } |
4661 | ||
4662 | /* | |
4663 | * creates new preallocated space for locality group inodes belongs to | |
4664 | */ | |
53f86b17 | 4665 | static noinline_for_stack void |
4ddfef7b | 4666 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
4667 | { |
4668 | struct super_block *sb = ac->ac_sb; | |
4669 | struct ext4_locality_group *lg; | |
4670 | struct ext4_prealloc_space *pa; | |
4671 | struct ext4_group_info *grp; | |
4672 | ||
4673 | /* preallocate only when found space is larger then requested */ | |
4674 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
4675 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4676 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 4677 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 4678 | |
53f86b17 | 4679 | pa = ac->ac_pa; |
c9de560d AT |
4680 | |
4681 | /* preallocation can change ac_b_ex, thus we store actually | |
4682 | * allocated blocks for history */ | |
4683 | ac->ac_f_ex = ac->ac_b_ex; | |
4684 | ||
4685 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
4686 | pa->pa_lstart = pa->pa_pstart; | |
4687 | pa->pa_len = ac->ac_b_ex.fe_len; | |
4688 | pa->pa_free = pa->pa_len; | |
c9de560d | 4689 | spin_lock_init(&pa->pa_lock); |
6be2ded1 | 4690 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 4691 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 4692 | pa->pa_deleted = 0; |
cc0fb9ad | 4693 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 4694 | |
d3df1453 RH |
4695 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
4696 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 4697 | trace_ext4_mb_new_group_pa(ac, pa); |
c9de560d AT |
4698 | |
4699 | ext4_mb_use_group_pa(ac, pa); | |
4700 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
4701 | ||
4702 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
4703 | lg = ac->ac_lg; | |
4704 | BUG_ON(lg == NULL); | |
4705 | ||
4706 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
4707 | pa->pa_inode = NULL; | |
4708 | ||
c9de560d | 4709 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 4710 | |
6be2ded1 AK |
4711 | /* |
4712 | * We will later add the new pa to the right bucket | |
4713 | * after updating the pa_free in ext4_mb_release_context | |
4714 | */ | |
c9de560d AT |
4715 | } |
4716 | ||
53f86b17 | 4717 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
c9de560d | 4718 | { |
c9de560d | 4719 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
53f86b17 | 4720 | ext4_mb_new_group_pa(ac); |
c9de560d | 4721 | else |
53f86b17 | 4722 | ext4_mb_new_inode_pa(ac); |
c9de560d AT |
4723 | } |
4724 | ||
4725 | /* | |
4726 | * finds all unused blocks in on-disk bitmap, frees them in | |
4727 | * in-core bitmap and buddy. | |
4728 | * @pa must be unlinked from inode and group lists, so that | |
4729 | * nobody else can find/use it. | |
4730 | * the caller MUST hold group/inode locks. | |
4731 | * TODO: optimize the case when there are no in-core structures yet | |
4732 | */ | |
4ddfef7b ES |
4733 | static noinline_for_stack int |
4734 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 4735 | struct ext4_prealloc_space *pa) |
c9de560d | 4736 | { |
c9de560d AT |
4737 | struct super_block *sb = e4b->bd_sb; |
4738 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
4739 | unsigned int end; |
4740 | unsigned int next; | |
c9de560d AT |
4741 | ext4_group_t group; |
4742 | ext4_grpblk_t bit; | |
ba80b101 | 4743 | unsigned long long grp_blk_start; |
c9de560d AT |
4744 | int free = 0; |
4745 | ||
4746 | BUG_ON(pa->pa_deleted == 0); | |
4747 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 4748 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
4749 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
4750 | end = bit + pa->pa_len; | |
4751 | ||
c9de560d | 4752 | while (bit < end) { |
ffad0a44 | 4753 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
4754 | if (bit >= end) |
4755 | break; | |
ffad0a44 | 4756 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
d3df1453 | 4757 | mb_debug(sb, "free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
4758 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
4759 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
4760 | free += next - bit; |
4761 | ||
3e1e5f50 | 4762 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
4763 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
4764 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 4765 | next - bit); |
c9de560d AT |
4766 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
4767 | bit = next + 1; | |
4768 | } | |
4769 | if (free != pa->pa_free) { | |
9d8b9ec4 | 4770 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 4771 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
4772 | pa, (unsigned long) pa->pa_lstart, |
4773 | (unsigned long) pa->pa_pstart, | |
36bad423 | 4774 | pa->pa_len); |
e29136f8 | 4775 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 4776 | free, pa->pa_free); |
e56eb659 AK |
4777 | /* |
4778 | * pa is already deleted so we use the value obtained | |
4779 | * from the bitmap and continue. | |
4780 | */ | |
c9de560d | 4781 | } |
c9de560d AT |
4782 | atomic_add(free, &sbi->s_mb_discarded); |
4783 | ||
863c37fc | 4784 | return 0; |
c9de560d AT |
4785 | } |
4786 | ||
4ddfef7b ES |
4787 | static noinline_for_stack int |
4788 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 4789 | struct ext4_prealloc_space *pa) |
c9de560d | 4790 | { |
c9de560d AT |
4791 | struct super_block *sb = e4b->bd_sb; |
4792 | ext4_group_t group; | |
4793 | ext4_grpblk_t bit; | |
4794 | ||
60e07cf5 | 4795 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
4796 | BUG_ON(pa->pa_deleted == 0); |
4797 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
4798 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
4799 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
4800 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 4801 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
4802 | |
4803 | return 0; | |
4804 | } | |
4805 | ||
4806 | /* | |
4807 | * releases all preallocations in given group | |
4808 | * | |
4809 | * first, we need to decide discard policy: | |
4810 | * - when do we discard | |
4811 | * 1) ENOSPC | |
4812 | * - how many do we discard | |
4813 | * 1) how many requested | |
4814 | */ | |
4ddfef7b ES |
4815 | static noinline_for_stack int |
4816 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
8c80fb31 | 4817 | ext4_group_t group, int *busy) |
c9de560d AT |
4818 | { |
4819 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4820 | struct buffer_head *bitmap_bh = NULL; | |
4821 | struct ext4_prealloc_space *pa, *tmp; | |
4822 | struct list_head list; | |
4823 | struct ext4_buddy e4b; | |
4824 | int err; | |
8c80fb31 | 4825 | int free = 0; |
c9de560d | 4826 | |
d3df1453 | 4827 | mb_debug(sb, "discard preallocation for group %u\n", group); |
c9de560d | 4828 | if (list_empty(&grp->bb_prealloc_list)) |
bbc4ec77 | 4829 | goto out_dbg; |
c9de560d | 4830 | |
574ca174 | 4831 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
4832 | if (IS_ERR(bitmap_bh)) { |
4833 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
4834 | ext4_error_err(sb, -err, |
4835 | "Error %d reading block bitmap for %u", | |
4836 | err, group); | |
bbc4ec77 | 4837 | goto out_dbg; |
c9de560d AT |
4838 | } |
4839 | ||
4840 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 4841 | if (err) { |
9651e6b2 KK |
4842 | ext4_warning(sb, "Error %d loading buddy information for %u", |
4843 | err, group); | |
ce89f46c | 4844 | put_bh(bitmap_bh); |
bbc4ec77 | 4845 | goto out_dbg; |
ce89f46c | 4846 | } |
c9de560d | 4847 | |
c9de560d | 4848 | INIT_LIST_HEAD(&list); |
c9de560d AT |
4849 | ext4_lock_group(sb, group); |
4850 | list_for_each_entry_safe(pa, tmp, | |
4851 | &grp->bb_prealloc_list, pa_group_list) { | |
4852 | spin_lock(&pa->pa_lock); | |
4853 | if (atomic_read(&pa->pa_count)) { | |
4854 | spin_unlock(&pa->pa_lock); | |
8c80fb31 | 4855 | *busy = 1; |
c9de560d AT |
4856 | continue; |
4857 | } | |
4858 | if (pa->pa_deleted) { | |
4859 | spin_unlock(&pa->pa_lock); | |
4860 | continue; | |
4861 | } | |
4862 | ||
4863 | /* seems this one can be freed ... */ | |
27bc446e | 4864 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 4865 | |
70022da8 YB |
4866 | if (!free) |
4867 | this_cpu_inc(discard_pa_seq); | |
4868 | ||
c9de560d AT |
4869 | /* we can trust pa_free ... */ |
4870 | free += pa->pa_free; | |
4871 | ||
4872 | spin_unlock(&pa->pa_lock); | |
4873 | ||
4874 | list_del(&pa->pa_group_list); | |
4875 | list_add(&pa->u.pa_tmp_list, &list); | |
4876 | } | |
4877 | ||
c9de560d AT |
4878 | /* now free all selected PAs */ |
4879 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
4880 | ||
4881 | /* remove from object (inode or locality group) */ | |
4882 | spin_lock(pa->pa_obj_lock); | |
4883 | list_del_rcu(&pa->pa_inode_list); | |
4884 | spin_unlock(pa->pa_obj_lock); | |
4885 | ||
cc0fb9ad | 4886 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 4887 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 4888 | else |
3e1e5f50 | 4889 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4890 | |
4891 | list_del(&pa->u.pa_tmp_list); | |
4892 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4893 | } | |
4894 | ||
c9de560d | 4895 | ext4_unlock_group(sb, group); |
e39e07fd | 4896 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 4897 | put_bh(bitmap_bh); |
bbc4ec77 | 4898 | out_dbg: |
d3df1453 | 4899 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
8c80fb31 CX |
4900 | free, group, grp->bb_free); |
4901 | return free; | |
c9de560d AT |
4902 | } |
4903 | ||
4904 | /* | |
4905 | * releases all non-used preallocated blocks for given inode | |
4906 | * | |
4907 | * It's important to discard preallocations under i_data_sem | |
4908 | * We don't want another block to be served from the prealloc | |
4909 | * space when we are discarding the inode prealloc space. | |
4910 | * | |
4911 | * FIXME!! Make sure it is valid at all the call sites | |
4912 | */ | |
27bc446e | 4913 | void ext4_discard_preallocations(struct inode *inode, unsigned int needed) |
c9de560d AT |
4914 | { |
4915 | struct ext4_inode_info *ei = EXT4_I(inode); | |
4916 | struct super_block *sb = inode->i_sb; | |
4917 | struct buffer_head *bitmap_bh = NULL; | |
4918 | struct ext4_prealloc_space *pa, *tmp; | |
4919 | ext4_group_t group = 0; | |
4920 | struct list_head list; | |
4921 | struct ext4_buddy e4b; | |
4922 | int err; | |
4923 | ||
c2ea3fde | 4924 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
4925 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
4926 | return; | |
4927 | } | |
4928 | ||
8016e29f HS |
4929 | if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) |
4930 | return; | |
4931 | ||
d3df1453 RH |
4932 | mb_debug(sb, "discard preallocation for inode %lu\n", |
4933 | inode->i_ino); | |
27bc446e | 4934 | trace_ext4_discard_preallocations(inode, |
4935 | atomic_read(&ei->i_prealloc_active), needed); | |
c9de560d AT |
4936 | |
4937 | INIT_LIST_HEAD(&list); | |
4938 | ||
27bc446e | 4939 | if (needed == 0) |
4940 | needed = UINT_MAX; | |
4941 | ||
c9de560d AT |
4942 | repeat: |
4943 | /* first, collect all pa's in the inode */ | |
4944 | spin_lock(&ei->i_prealloc_lock); | |
27bc446e | 4945 | while (!list_empty(&ei->i_prealloc_list) && needed) { |
4946 | pa = list_entry(ei->i_prealloc_list.prev, | |
c9de560d AT |
4947 | struct ext4_prealloc_space, pa_inode_list); |
4948 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
4949 | spin_lock(&pa->pa_lock); | |
4950 | if (atomic_read(&pa->pa_count)) { | |
4951 | /* this shouldn't happen often - nobody should | |
4952 | * use preallocation while we're discarding it */ | |
4953 | spin_unlock(&pa->pa_lock); | |
4954 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
4955 | ext4_msg(sb, KERN_ERR, |
4956 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
4957 | WARN_ON(1); |
4958 | schedule_timeout_uninterruptible(HZ); | |
4959 | goto repeat; | |
4960 | ||
4961 | } | |
4962 | if (pa->pa_deleted == 0) { | |
27bc446e | 4963 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
4964 | spin_unlock(&pa->pa_lock); |
4965 | list_del_rcu(&pa->pa_inode_list); | |
4966 | list_add(&pa->u.pa_tmp_list, &list); | |
27bc446e | 4967 | needed--; |
c9de560d AT |
4968 | continue; |
4969 | } | |
4970 | ||
4971 | /* someone is deleting pa right now */ | |
4972 | spin_unlock(&pa->pa_lock); | |
4973 | spin_unlock(&ei->i_prealloc_lock); | |
4974 | ||
4975 | /* we have to wait here because pa_deleted | |
4976 | * doesn't mean pa is already unlinked from | |
4977 | * the list. as we might be called from | |
4978 | * ->clear_inode() the inode will get freed | |
4979 | * and concurrent thread which is unlinking | |
4980 | * pa from inode's list may access already | |
4981 | * freed memory, bad-bad-bad */ | |
4982 | ||
4983 | /* XXX: if this happens too often, we can | |
4984 | * add a flag to force wait only in case | |
4985 | * of ->clear_inode(), but not in case of | |
4986 | * regular truncate */ | |
4987 | schedule_timeout_uninterruptible(HZ); | |
4988 | goto repeat; | |
4989 | } | |
4990 | spin_unlock(&ei->i_prealloc_lock); | |
4991 | ||
4992 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 4993 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 4994 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 4995 | |
9651e6b2 KK |
4996 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
4997 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 4998 | if (err) { |
54d3adbc TT |
4999 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5000 | err, group); | |
ce89f46c AK |
5001 | continue; |
5002 | } | |
c9de560d | 5003 | |
574ca174 | 5004 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5005 | if (IS_ERR(bitmap_bh)) { |
5006 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5007 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
5008 | err, group); | |
e39e07fd | 5009 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 5010 | continue; |
c9de560d AT |
5011 | } |
5012 | ||
5013 | ext4_lock_group(sb, group); | |
5014 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5015 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
5016 | ext4_unlock_group(sb, group); |
5017 | ||
e39e07fd | 5018 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
5019 | put_bh(bitmap_bh); |
5020 | ||
5021 | list_del(&pa->u.pa_tmp_list); | |
5022 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5023 | } | |
5024 | } | |
5025 | ||
53f86b17 RH |
5026 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
5027 | { | |
5028 | struct ext4_prealloc_space *pa; | |
5029 | ||
5030 | BUG_ON(ext4_pspace_cachep == NULL); | |
5031 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); | |
5032 | if (!pa) | |
5033 | return -ENOMEM; | |
5034 | atomic_set(&pa->pa_count, 1); | |
5035 | ac->ac_pa = pa; | |
5036 | return 0; | |
5037 | } | |
5038 | ||
5039 | static void ext4_mb_pa_free(struct ext4_allocation_context *ac) | |
5040 | { | |
5041 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
5042 | ||
5043 | BUG_ON(!pa); | |
5044 | ac->ac_pa = NULL; | |
5045 | WARN_ON(!atomic_dec_and_test(&pa->pa_count)); | |
5046 | kmem_cache_free(ext4_pspace_cachep, pa); | |
5047 | } | |
5048 | ||
6ba495e9 | 5049 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 5050 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 5051 | { |
e68cf40c | 5052 | ext4_group_t i, ngroups; |
c9de560d | 5053 | |
9b5f6c9b | 5054 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e3570639 ES |
5055 | return; |
5056 | ||
8df9675f | 5057 | ngroups = ext4_get_groups_count(sb); |
d3df1453 | 5058 | mb_debug(sb, "groups: "); |
8df9675f | 5059 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
5060 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
5061 | struct ext4_prealloc_space *pa; | |
5062 | ext4_grpblk_t start; | |
5063 | struct list_head *cur; | |
5064 | ext4_lock_group(sb, i); | |
5065 | list_for_each(cur, &grp->bb_prealloc_list) { | |
5066 | pa = list_entry(cur, struct ext4_prealloc_space, | |
5067 | pa_group_list); | |
5068 | spin_lock(&pa->pa_lock); | |
5069 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
5070 | NULL, &start); | |
5071 | spin_unlock(&pa->pa_lock); | |
d3df1453 RH |
5072 | mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
5073 | pa->pa_len); | |
c9de560d | 5074 | } |
60bd63d1 | 5075 | ext4_unlock_group(sb, i); |
d3df1453 RH |
5076 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
5077 | grp->bb_fragments); | |
c9de560d | 5078 | } |
c9de560d | 5079 | } |
e68cf40c RH |
5080 | |
5081 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
5082 | { | |
5083 | struct super_block *sb = ac->ac_sb; | |
5084 | ||
9b5f6c9b | 5085 | if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED)) |
e68cf40c RH |
5086 | return; |
5087 | ||
d3df1453 | 5088 | mb_debug(sb, "Can't allocate:" |
e68cf40c | 5089 | " Allocation context details:"); |
d3df1453 | 5090 | mb_debug(sb, "status %u flags 0x%x", |
e68cf40c | 5091 | ac->ac_status, ac->ac_flags); |
d3df1453 | 5092 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
e68cf40c RH |
5093 | "goal %lu/%lu/%lu@%lu, " |
5094 | "best %lu/%lu/%lu@%lu cr %d", | |
5095 | (unsigned long)ac->ac_o_ex.fe_group, | |
5096 | (unsigned long)ac->ac_o_ex.fe_start, | |
5097 | (unsigned long)ac->ac_o_ex.fe_len, | |
5098 | (unsigned long)ac->ac_o_ex.fe_logical, | |
5099 | (unsigned long)ac->ac_g_ex.fe_group, | |
5100 | (unsigned long)ac->ac_g_ex.fe_start, | |
5101 | (unsigned long)ac->ac_g_ex.fe_len, | |
5102 | (unsigned long)ac->ac_g_ex.fe_logical, | |
5103 | (unsigned long)ac->ac_b_ex.fe_group, | |
5104 | (unsigned long)ac->ac_b_ex.fe_start, | |
5105 | (unsigned long)ac->ac_b_ex.fe_len, | |
5106 | (unsigned long)ac->ac_b_ex.fe_logical, | |
5107 | (int)ac->ac_criteria); | |
d3df1453 | 5108 | mb_debug(sb, "%u found", ac->ac_found); |
e68cf40c RH |
5109 | ext4_mb_show_pa(sb); |
5110 | } | |
c9de560d | 5111 | #else |
e68cf40c RH |
5112 | static inline void ext4_mb_show_pa(struct super_block *sb) |
5113 | { | |
5114 | return; | |
5115 | } | |
c9de560d AT |
5116 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
5117 | { | |
e68cf40c | 5118 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
5119 | return; |
5120 | } | |
5121 | #endif | |
5122 | ||
5123 | /* | |
5124 | * We use locality group preallocation for small size file. The size of the | |
5125 | * file is determined by the current size or the resulting size after | |
5126 | * allocation which ever is larger | |
5127 | * | |
b713a5ec | 5128 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
5129 | */ |
5130 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
5131 | { | |
5132 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
5133 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
5134 | loff_t size, isize; | |
5135 | ||
5136 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
5137 | return; | |
5138 | ||
4ba74d00 TT |
5139 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
5140 | return; | |
5141 | ||
53accfa9 | 5142 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
5143 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
5144 | >> bsbits; | |
c9de560d | 5145 | |
82dd124c NB |
5146 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
5147 | !inode_is_open_for_write(ac->ac_inode)) { | |
50797481 TT |
5148 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; |
5149 | return; | |
5150 | } | |
5151 | ||
ebbe0277 RD |
5152 | if (sbi->s_mb_group_prealloc <= 0) { |
5153 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
5154 | return; | |
5155 | } | |
5156 | ||
c9de560d | 5157 | /* don't use group allocation for large files */ |
71780577 | 5158 | size = max(size, isize); |
cc483f10 | 5159 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 5160 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 5161 | return; |
4ba74d00 | 5162 | } |
c9de560d AT |
5163 | |
5164 | BUG_ON(ac->ac_lg != NULL); | |
5165 | /* | |
5166 | * locality group prealloc space are per cpu. The reason for having | |
5167 | * per cpu locality group is to reduce the contention between block | |
5168 | * request from multiple CPUs. | |
5169 | */ | |
a0b6bc63 | 5170 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
5171 | |
5172 | /* we're going to use group allocation */ | |
5173 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
5174 | ||
5175 | /* serialize all allocations in the group */ | |
5176 | mutex_lock(&ac->ac_lg->lg_mutex); | |
5177 | } | |
5178 | ||
4ddfef7b ES |
5179 | static noinline_for_stack int |
5180 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
5181 | struct ext4_allocation_request *ar) |
5182 | { | |
5183 | struct super_block *sb = ar->inode->i_sb; | |
5184 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
5185 | struct ext4_super_block *es = sbi->s_es; | |
5186 | ext4_group_t group; | |
498e5f24 TT |
5187 | unsigned int len; |
5188 | ext4_fsblk_t goal; | |
c9de560d AT |
5189 | ext4_grpblk_t block; |
5190 | ||
5191 | /* we can't allocate > group size */ | |
5192 | len = ar->len; | |
5193 | ||
5194 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
5195 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
5196 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
5197 | |
5198 | /* start searching from the goal */ | |
5199 | goal = ar->goal; | |
5200 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5201 | goal >= ext4_blocks_count(es)) | |
5202 | goal = le32_to_cpu(es->s_first_data_block); | |
5203 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
5204 | ||
5205 | /* set up allocation goals */ | |
f5a44db5 | 5206 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 5207 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
5208 | ac->ac_sb = sb; |
5209 | ac->ac_inode = ar->inode; | |
53accfa9 | 5210 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
5211 | ac->ac_o_ex.fe_group = group; |
5212 | ac->ac_o_ex.fe_start = block; | |
5213 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 5214 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 5215 | ac->ac_flags = ar->flags; |
c9de560d | 5216 | |
3cb77bd2 | 5217 | /* we have to define context: we'll work with a file or |
c9de560d AT |
5218 | * locality group. this is a policy, actually */ |
5219 | ext4_mb_group_or_file(ac); | |
5220 | ||
d3df1453 | 5221 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
c9de560d AT |
5222 | "left: %u/%u, right %u/%u to %swritable\n", |
5223 | (unsigned) ar->len, (unsigned) ar->logical, | |
5224 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
5225 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
5226 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 5227 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
5228 | return 0; |
5229 | ||
5230 | } | |
5231 | ||
6be2ded1 AK |
5232 | static noinline_for_stack void |
5233 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
5234 | struct ext4_locality_group *lg, | |
5235 | int order, int total_entries) | |
5236 | { | |
5237 | ext4_group_t group = 0; | |
5238 | struct ext4_buddy e4b; | |
5239 | struct list_head discard_list; | |
5240 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 5241 | |
d3df1453 | 5242 | mb_debug(sb, "discard locality group preallocation\n"); |
6be2ded1 AK |
5243 | |
5244 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
5245 | |
5246 | spin_lock(&lg->lg_prealloc_lock); | |
5247 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
92e9c58c MB |
5248 | pa_inode_list, |
5249 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
5250 | spin_lock(&pa->pa_lock); |
5251 | if (atomic_read(&pa->pa_count)) { | |
5252 | /* | |
5253 | * This is the pa that we just used | |
5254 | * for block allocation. So don't | |
5255 | * free that | |
5256 | */ | |
5257 | spin_unlock(&pa->pa_lock); | |
5258 | continue; | |
5259 | } | |
5260 | if (pa->pa_deleted) { | |
5261 | spin_unlock(&pa->pa_lock); | |
5262 | continue; | |
5263 | } | |
5264 | /* only lg prealloc space */ | |
cc0fb9ad | 5265 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
5266 | |
5267 | /* seems this one can be freed ... */ | |
27bc446e | 5268 | ext4_mb_mark_pa_deleted(sb, pa); |
6be2ded1 AK |
5269 | spin_unlock(&pa->pa_lock); |
5270 | ||
5271 | list_del_rcu(&pa->pa_inode_list); | |
5272 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
5273 | ||
5274 | total_entries--; | |
5275 | if (total_entries <= 5) { | |
5276 | /* | |
5277 | * we want to keep only 5 entries | |
5278 | * allowing it to grow to 8. This | |
5279 | * mak sure we don't call discard | |
5280 | * soon for this list. | |
5281 | */ | |
5282 | break; | |
5283 | } | |
5284 | } | |
5285 | spin_unlock(&lg->lg_prealloc_lock); | |
5286 | ||
5287 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 5288 | int err; |
6be2ded1 | 5289 | |
bd86298e | 5290 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
5291 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5292 | GFP_NOFS|__GFP_NOFAIL); | |
5293 | if (err) { | |
54d3adbc TT |
5294 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5295 | err, group); | |
6be2ded1 AK |
5296 | continue; |
5297 | } | |
5298 | ext4_lock_group(sb, group); | |
5299 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5300 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
5301 | ext4_unlock_group(sb, group); |
5302 | ||
e39e07fd | 5303 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
5304 | list_del(&pa->u.pa_tmp_list); |
5305 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5306 | } | |
6be2ded1 AK |
5307 | } |
5308 | ||
5309 | /* | |
5310 | * We have incremented pa_count. So it cannot be freed at this | |
5311 | * point. Also we hold lg_mutex. So no parallel allocation is | |
5312 | * possible from this lg. That means pa_free cannot be updated. | |
5313 | * | |
5314 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
5315 | * which can cause the lg_prealloc_list to be updated. | |
5316 | */ | |
5317 | ||
5318 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
5319 | { | |
5320 | int order, added = 0, lg_prealloc_count = 1; | |
5321 | struct super_block *sb = ac->ac_sb; | |
5322 | struct ext4_locality_group *lg = ac->ac_lg; | |
5323 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
5324 | ||
5325 | order = fls(pa->pa_free) - 1; | |
5326 | if (order > PREALLOC_TB_SIZE - 1) | |
5327 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
5328 | order = PREALLOC_TB_SIZE - 1; | |
5329 | /* Add the prealloc space to lg */ | |
f1167009 | 5330 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 5331 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
92e9c58c MB |
5332 | pa_inode_list, |
5333 | lockdep_is_held(&lg->lg_prealloc_lock)) { | |
6be2ded1 AK |
5334 | spin_lock(&tmp_pa->pa_lock); |
5335 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 5336 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
5337 | continue; |
5338 | } | |
5339 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
5340 | /* Add to the tail of the previous entry */ | |
5341 | list_add_tail_rcu(&pa->pa_inode_list, | |
5342 | &tmp_pa->pa_inode_list); | |
5343 | added = 1; | |
5344 | /* | |
5345 | * we want to count the total | |
5346 | * number of entries in the list | |
5347 | */ | |
5348 | } | |
5349 | spin_unlock(&tmp_pa->pa_lock); | |
5350 | lg_prealloc_count++; | |
5351 | } | |
5352 | if (!added) | |
5353 | list_add_tail_rcu(&pa->pa_inode_list, | |
5354 | &lg->lg_prealloc_list[order]); | |
f1167009 | 5355 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
5356 | |
5357 | /* Now trim the list to be not more than 8 elements */ | |
5358 | if (lg_prealloc_count > 8) { | |
5359 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 5360 | order, lg_prealloc_count); |
6be2ded1 AK |
5361 | return; |
5362 | } | |
5363 | return ; | |
5364 | } | |
5365 | ||
27bc446e | 5366 | /* |
5367 | * if per-inode prealloc list is too long, trim some PA | |
5368 | */ | |
5369 | static void ext4_mb_trim_inode_pa(struct inode *inode) | |
5370 | { | |
5371 | struct ext4_inode_info *ei = EXT4_I(inode); | |
5372 | struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); | |
5373 | int count, delta; | |
5374 | ||
5375 | count = atomic_read(&ei->i_prealloc_active); | |
5376 | delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1; | |
5377 | if (count > sbi->s_mb_max_inode_prealloc + delta) { | |
5378 | count -= sbi->s_mb_max_inode_prealloc; | |
5379 | ext4_discard_preallocations(inode, count); | |
5380 | } | |
5381 | } | |
5382 | ||
c9de560d AT |
5383 | /* |
5384 | * release all resource we used in allocation | |
5385 | */ | |
5386 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
5387 | { | |
27bc446e | 5388 | struct inode *inode = ac->ac_inode; |
5389 | struct ext4_inode_info *ei = EXT4_I(inode); | |
53accfa9 | 5390 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
5391 | struct ext4_prealloc_space *pa = ac->ac_pa; |
5392 | if (pa) { | |
cc0fb9ad | 5393 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 5394 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 5395 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
5396 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
5397 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
5398 | pa->pa_free -= ac->ac_b_ex.fe_len; |
5399 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
5400 | spin_unlock(&pa->pa_lock); | |
66d5e027 | 5401 | |
5402 | /* | |
5403 | * We want to add the pa to the right bucket. | |
5404 | * Remove it from the list and while adding | |
5405 | * make sure the list to which we are adding | |
5406 | * doesn't grow big. | |
5407 | */ | |
5408 | if (likely(pa->pa_free)) { | |
5409 | spin_lock(pa->pa_obj_lock); | |
5410 | list_del_rcu(&pa->pa_inode_list); | |
5411 | spin_unlock(pa->pa_obj_lock); | |
5412 | ext4_mb_add_n_trim(ac); | |
5413 | } | |
ba443916 | 5414 | } |
27bc446e | 5415 | |
5416 | if (pa->pa_type == MB_INODE_PA) { | |
5417 | /* | |
5418 | * treat per-inode prealloc list as a lru list, then try | |
5419 | * to trim the least recently used PA. | |
5420 | */ | |
5421 | spin_lock(pa->pa_obj_lock); | |
5422 | list_move(&pa->pa_inode_list, &ei->i_prealloc_list); | |
5423 | spin_unlock(pa->pa_obj_lock); | |
5424 | } | |
5425 | ||
ba443916 AK |
5426 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
5427 | } | |
c9de560d | 5428 | if (ac->ac_bitmap_page) |
09cbfeaf | 5429 | put_page(ac->ac_bitmap_page); |
c9de560d | 5430 | if (ac->ac_buddy_page) |
09cbfeaf | 5431 | put_page(ac->ac_buddy_page); |
c9de560d AT |
5432 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
5433 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
5434 | ext4_mb_collect_stats(ac); | |
27bc446e | 5435 | ext4_mb_trim_inode_pa(inode); |
c9de560d AT |
5436 | return 0; |
5437 | } | |
5438 | ||
5439 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
5440 | { | |
8df9675f | 5441 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d | 5442 | int ret; |
8c80fb31 CX |
5443 | int freed = 0, busy = 0; |
5444 | int retry = 0; | |
c9de560d | 5445 | |
9bffad1e | 5446 | trace_ext4_mb_discard_preallocations(sb, needed); |
8c80fb31 CX |
5447 | |
5448 | if (needed == 0) | |
5449 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; | |
5450 | repeat: | |
8df9675f | 5451 | for (i = 0; i < ngroups && needed > 0; i++) { |
8c80fb31 | 5452 | ret = ext4_mb_discard_group_preallocations(sb, i, &busy); |
c9de560d AT |
5453 | freed += ret; |
5454 | needed -= ret; | |
8c80fb31 CX |
5455 | cond_resched(); |
5456 | } | |
5457 | ||
5458 | if (needed > 0 && busy && ++retry < 3) { | |
5459 | busy = 0; | |
5460 | goto repeat; | |
c9de560d AT |
5461 | } |
5462 | ||
5463 | return freed; | |
5464 | } | |
5465 | ||
cf5e2ca6 | 5466 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
07b5b8e1 | 5467 | struct ext4_allocation_context *ac, u64 *seq) |
cf5e2ca6 RH |
5468 | { |
5469 | int freed; | |
07b5b8e1 RH |
5470 | u64 seq_retry = 0; |
5471 | bool ret = false; | |
cf5e2ca6 RH |
5472 | |
5473 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
07b5b8e1 RH |
5474 | if (freed) { |
5475 | ret = true; | |
5476 | goto out_dbg; | |
5477 | } | |
5478 | seq_retry = ext4_get_discard_pa_seq_sum(); | |
99377830 RH |
5479 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
5480 | ac->ac_flags |= EXT4_MB_STRICT_CHECK; | |
07b5b8e1 RH |
5481 | *seq = seq_retry; |
5482 | ret = true; | |
5483 | } | |
5484 | ||
5485 | out_dbg: | |
5486 | mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); | |
5487 | return ret; | |
cf5e2ca6 RH |
5488 | } |
5489 | ||
8016e29f HS |
5490 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, |
5491 | struct ext4_allocation_request *ar, int *errp); | |
5492 | ||
c9de560d AT |
5493 | /* |
5494 | * Main entry point into mballoc to allocate blocks | |
5495 | * it tries to use preallocation first, then falls back | |
5496 | * to usual allocation | |
5497 | */ | |
5498 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 5499 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 5500 | { |
256bdb49 | 5501 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
5502 | struct ext4_sb_info *sbi; |
5503 | struct super_block *sb; | |
5504 | ext4_fsblk_t block = 0; | |
60e58e0f | 5505 | unsigned int inquota = 0; |
53accfa9 | 5506 | unsigned int reserv_clstrs = 0; |
07b5b8e1 | 5507 | u64 seq; |
c9de560d | 5508 | |
b10a44c3 | 5509 | might_sleep(); |
c9de560d AT |
5510 | sb = ar->inode->i_sb; |
5511 | sbi = EXT4_SB(sb); | |
5512 | ||
9bffad1e | 5513 | trace_ext4_request_blocks(ar); |
8016e29f HS |
5514 | if (sbi->s_mount_state & EXT4_FC_REPLAY) |
5515 | return ext4_mb_new_blocks_simple(handle, ar, errp); | |
ba80b101 | 5516 | |
45dc63e7 | 5517 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 5518 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
5519 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
5520 | ||
e3cf5d5d | 5521 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
5522 | /* Without delayed allocation we need to verify |
5523 | * there is enough free blocks to do block allocation | |
5524 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 5525 | */ |
55f020db | 5526 | while (ar->len && |
e7d5f315 | 5527 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 5528 | |
030ba6bc | 5529 | /* let others to free the space */ |
bb8b20ed | 5530 | cond_resched(); |
030ba6bc AK |
5531 | ar->len = ar->len >> 1; |
5532 | } | |
5533 | if (!ar->len) { | |
bbc4ec77 | 5534 | ext4_mb_show_pa(sb); |
a30d542a AK |
5535 | *errp = -ENOSPC; |
5536 | return 0; | |
5537 | } | |
53accfa9 | 5538 | reserv_clstrs = ar->len; |
55f020db | 5539 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
5540 | dquot_alloc_block_nofail(ar->inode, |
5541 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
5542 | } else { |
5543 | while (ar->len && | |
53accfa9 TT |
5544 | dquot_alloc_block(ar->inode, |
5545 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
5546 | |
5547 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
5548 | ar->len--; | |
5549 | } | |
60e58e0f MC |
5550 | } |
5551 | inquota = ar->len; | |
5552 | if (ar->len == 0) { | |
5553 | *errp = -EDQUOT; | |
6c7a120a | 5554 | goto out; |
60e58e0f | 5555 | } |
07031431 | 5556 | } |
d2a17637 | 5557 | |
85556c9a | 5558 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 5559 | if (!ac) { |
363d4251 | 5560 | ar->len = 0; |
256bdb49 | 5561 | *errp = -ENOMEM; |
6c7a120a | 5562 | goto out; |
256bdb49 ES |
5563 | } |
5564 | ||
256bdb49 | 5565 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
5566 | if (*errp) { |
5567 | ar->len = 0; | |
6c7a120a | 5568 | goto out; |
c9de560d AT |
5569 | } |
5570 | ||
256bdb49 | 5571 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
81198536 | 5572 | seq = this_cpu_read(discard_pa_seq); |
256bdb49 | 5573 | if (!ext4_mb_use_preallocated(ac)) { |
256bdb49 ES |
5574 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
5575 | ext4_mb_normalize_request(ac, ar); | |
53f86b17 RH |
5576 | |
5577 | *errp = ext4_mb_pa_alloc(ac); | |
5578 | if (*errp) | |
5579 | goto errout; | |
c9de560d AT |
5580 | repeat: |
5581 | /* allocate space in core */ | |
6c7a120a | 5582 | *errp = ext4_mb_regular_allocator(ac); |
53f86b17 RH |
5583 | /* |
5584 | * pa allocated above is added to grp->bb_prealloc_list only | |
5585 | * when we were able to allocate some block i.e. when | |
5586 | * ac->ac_status == AC_STATUS_FOUND. | |
5587 | * And error from above mean ac->ac_status != AC_STATUS_FOUND | |
5588 | * So we have to free this pa here itself. | |
5589 | */ | |
2c00ef3e | 5590 | if (*errp) { |
53f86b17 | 5591 | ext4_mb_pa_free(ac); |
2c00ef3e AK |
5592 | ext4_discard_allocated_blocks(ac); |
5593 | goto errout; | |
5594 | } | |
53f86b17 RH |
5595 | if (ac->ac_status == AC_STATUS_FOUND && |
5596 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) | |
5597 | ext4_mb_pa_free(ac); | |
c9de560d | 5598 | } |
256bdb49 | 5599 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 5600 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 5601 | if (*errp) { |
b844167e | 5602 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
5603 | goto errout; |
5604 | } else { | |
519deca0 AK |
5605 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
5606 | ar->len = ac->ac_b_ex.fe_len; | |
5607 | } | |
c9de560d | 5608 | } else { |
07b5b8e1 | 5609 | if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) |
c9de560d | 5610 | goto repeat; |
53f86b17 RH |
5611 | /* |
5612 | * If block allocation fails then the pa allocated above | |
5613 | * needs to be freed here itself. | |
5614 | */ | |
5615 | ext4_mb_pa_free(ac); | |
c9de560d | 5616 | *errp = -ENOSPC; |
6c7a120a AK |
5617 | } |
5618 | ||
6d138ced | 5619 | errout: |
6c7a120a | 5620 | if (*errp) { |
256bdb49 | 5621 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 5622 | ar->len = 0; |
256bdb49 | 5623 | ext4_mb_show_ac(ac); |
c9de560d | 5624 | } |
256bdb49 | 5625 | ext4_mb_release_context(ac); |
6c7a120a AK |
5626 | out: |
5627 | if (ac) | |
5628 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 5629 | if (inquota && ar->len < inquota) |
53accfa9 | 5630 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 5631 | if (!ar->len) { |
e3cf5d5d | 5632 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 5633 | /* release all the reserved blocks if non delalloc */ |
57042651 | 5634 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 5635 | reserv_clstrs); |
0087d9fb | 5636 | } |
c9de560d | 5637 | |
9bffad1e | 5638 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 5639 | |
c9de560d AT |
5640 | return block; |
5641 | } | |
c9de560d | 5642 | |
c894058d AK |
5643 | /* |
5644 | * We can merge two free data extents only if the physical blocks | |
5645 | * are contiguous, AND the extents were freed by the same transaction, | |
5646 | * AND the blocks are associated with the same group. | |
5647 | */ | |
a0154344 DJ |
5648 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
5649 | struct ext4_free_data *entry, | |
5650 | struct ext4_free_data *new_entry, | |
5651 | struct rb_root *entry_rb_root) | |
c894058d | 5652 | { |
a0154344 DJ |
5653 | if ((entry->efd_tid != new_entry->efd_tid) || |
5654 | (entry->efd_group != new_entry->efd_group)) | |
5655 | return; | |
5656 | if (entry->efd_start_cluster + entry->efd_count == | |
5657 | new_entry->efd_start_cluster) { | |
5658 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
5659 | new_entry->efd_count += entry->efd_count; | |
5660 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
5661 | entry->efd_start_cluster) { | |
5662 | new_entry->efd_count += entry->efd_count; | |
5663 | } else | |
5664 | return; | |
5665 | spin_lock(&sbi->s_md_lock); | |
5666 | list_del(&entry->efd_list); | |
5667 | spin_unlock(&sbi->s_md_lock); | |
5668 | rb_erase(&entry->efd_node, entry_rb_root); | |
5669 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
5670 | } |
5671 | ||
4ddfef7b ES |
5672 | static noinline_for_stack int |
5673 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 5674 | struct ext4_free_data *new_entry) |
c9de560d | 5675 | { |
e29136f8 | 5676 | ext4_group_t group = e4b->bd_group; |
84130193 | 5677 | ext4_grpblk_t cluster; |
d08854f5 | 5678 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 5679 | struct ext4_free_data *entry; |
c9de560d AT |
5680 | struct ext4_group_info *db = e4b->bd_info; |
5681 | struct super_block *sb = e4b->bd_sb; | |
5682 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
5683 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
5684 | struct rb_node *parent = NULL, *new_node; | |
5685 | ||
0390131b | 5686 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
5687 | BUG_ON(e4b->bd_bitmap_page == NULL); |
5688 | BUG_ON(e4b->bd_buddy_page == NULL); | |
5689 | ||
18aadd47 BJ |
5690 | new_node = &new_entry->efd_node; |
5691 | cluster = new_entry->efd_start_cluster; | |
c894058d | 5692 | |
c894058d AK |
5693 | if (!*n) { |
5694 | /* first free block exent. We need to | |
5695 | protect buddy cache from being freed, | |
5696 | * otherwise we'll refresh it from | |
5697 | * on-disk bitmap and lose not-yet-available | |
5698 | * blocks */ | |
09cbfeaf KS |
5699 | get_page(e4b->bd_buddy_page); |
5700 | get_page(e4b->bd_bitmap_page); | |
c894058d AK |
5701 | } |
5702 | while (*n) { | |
5703 | parent = *n; | |
18aadd47 BJ |
5704 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
5705 | if (cluster < entry->efd_start_cluster) | |
c894058d | 5706 | n = &(*n)->rb_left; |
18aadd47 | 5707 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
5708 | n = &(*n)->rb_right; |
5709 | else { | |
e29136f8 | 5710 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
5711 | ext4_group_first_block_no(sb, group) + |
5712 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 5713 | "Block already on to-be-freed list"); |
cca41553 | 5714 | kmem_cache_free(ext4_free_data_cachep, new_entry); |
c894058d | 5715 | return 0; |
c9de560d | 5716 | } |
c894058d | 5717 | } |
c9de560d | 5718 | |
c894058d AK |
5719 | rb_link_node(new_node, parent, n); |
5720 | rb_insert_color(new_node, &db->bb_free_root); | |
5721 | ||
5722 | /* Now try to see the extent can be merged to left and right */ | |
5723 | node = rb_prev(new_node); | |
5724 | if (node) { | |
18aadd47 | 5725 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5726 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5727 | &(db->bb_free_root)); | |
c894058d | 5728 | } |
c9de560d | 5729 | |
c894058d AK |
5730 | node = rb_next(new_node); |
5731 | if (node) { | |
18aadd47 | 5732 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
5733 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
5734 | &(db->bb_free_root)); | |
c9de560d | 5735 | } |
a0154344 | 5736 | |
d08854f5 | 5737 | spin_lock(&sbi->s_md_lock); |
a0154344 | 5738 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); |
d08854f5 TT |
5739 | sbi->s_mb_free_pending += clusters; |
5740 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
5741 | return 0; |
5742 | } | |
5743 | ||
8016e29f HS |
5744 | /* |
5745 | * Simple allocator for Ext4 fast commit replay path. It searches for blocks | |
5746 | * linearly starting at the goal block and also excludes the blocks which | |
5747 | * are going to be in use after fast commit replay. | |
5748 | */ | |
5749 | static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle, | |
5750 | struct ext4_allocation_request *ar, int *errp) | |
5751 | { | |
5752 | struct buffer_head *bitmap_bh; | |
5753 | struct super_block *sb = ar->inode->i_sb; | |
5754 | ext4_group_t group; | |
5755 | ext4_grpblk_t blkoff; | |
e121bd48 | 5756 | int i = sb->s_blocksize; |
8016e29f HS |
5757 | ext4_fsblk_t goal, block; |
5758 | struct ext4_super_block *es = EXT4_SB(sb)->s_es; | |
5759 | ||
5760 | goal = ar->goal; | |
5761 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5762 | goal >= ext4_blocks_count(es)) | |
5763 | goal = le32_to_cpu(es->s_first_data_block); | |
5764 | ||
5765 | ar->len = 0; | |
5766 | ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); | |
5767 | for (; group < ext4_get_groups_count(sb); group++) { | |
5768 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
5769 | if (IS_ERR(bitmap_bh)) { | |
5770 | *errp = PTR_ERR(bitmap_bh); | |
5771 | pr_warn("Failed to read block bitmap\n"); | |
5772 | return 0; | |
5773 | } | |
5774 | ||
5775 | ext4_get_group_no_and_offset(sb, | |
5776 | max(ext4_group_first_block_no(sb, group), goal), | |
5777 | NULL, &blkoff); | |
5778 | i = mb_find_next_zero_bit(bitmap_bh->b_data, sb->s_blocksize, | |
5779 | blkoff); | |
5780 | brelse(bitmap_bh); | |
5781 | if (i >= sb->s_blocksize) | |
5782 | continue; | |
5783 | if (ext4_fc_replay_check_excluded(sb, | |
5784 | ext4_group_first_block_no(sb, group) + i)) | |
5785 | continue; | |
5786 | break; | |
5787 | } | |
5788 | ||
5789 | if (group >= ext4_get_groups_count(sb) && i >= sb->s_blocksize) | |
5790 | return 0; | |
5791 | ||
5792 | block = ext4_group_first_block_no(sb, group) + i; | |
5793 | ext4_mb_mark_bb(sb, block, 1, 1); | |
5794 | ar->len = 1; | |
5795 | ||
5796 | return block; | |
5797 | } | |
5798 | ||
5799 | static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, | |
5800 | unsigned long count) | |
5801 | { | |
5802 | struct buffer_head *bitmap_bh; | |
5803 | struct super_block *sb = inode->i_sb; | |
5804 | struct ext4_group_desc *gdp; | |
5805 | struct buffer_head *gdp_bh; | |
5806 | ext4_group_t group; | |
5807 | ext4_grpblk_t blkoff; | |
5808 | int already_freed = 0, err, i; | |
5809 | ||
5810 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
5811 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
5812 | if (IS_ERR(bitmap_bh)) { | |
5813 | err = PTR_ERR(bitmap_bh); | |
5814 | pr_warn("Failed to read block bitmap\n"); | |
5815 | return; | |
5816 | } | |
5817 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
5818 | if (!gdp) | |
5819 | return; | |
5820 | ||
5821 | for (i = 0; i < count; i++) { | |
5822 | if (!mb_test_bit(blkoff + i, bitmap_bh->b_data)) | |
5823 | already_freed++; | |
5824 | } | |
5825 | mb_clear_bits(bitmap_bh->b_data, blkoff, count); | |
5826 | err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); | |
5827 | if (err) | |
5828 | return; | |
5829 | ext4_free_group_clusters_set( | |
5830 | sb, gdp, ext4_free_group_clusters(sb, gdp) + | |
5831 | count - already_freed); | |
5832 | ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh); | |
5833 | ext4_group_desc_csum_set(sb, group, gdp); | |
5834 | ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); | |
5835 | sync_dirty_buffer(bitmap_bh); | |
5836 | sync_dirty_buffer(gdp_bh); | |
5837 | brelse(bitmap_bh); | |
5838 | } | |
5839 | ||
44338711 TT |
5840 | /** |
5841 | * ext4_free_blocks() -- Free given blocks and update quota | |
5842 | * @handle: handle for this transaction | |
5843 | * @inode: inode | |
c60990b3 TT |
5844 | * @bh: optional buffer of the block to be freed |
5845 | * @block: starting physical block to be freed | |
5846 | * @count: number of blocks to be freed | |
5def1360 | 5847 | * @flags: flags used by ext4_free_blocks |
c9de560d | 5848 | */ |
44338711 | 5849 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
5850 | struct buffer_head *bh, ext4_fsblk_t block, |
5851 | unsigned long count, int flags) | |
c9de560d | 5852 | { |
26346ff6 | 5853 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 5854 | struct super_block *sb = inode->i_sb; |
c9de560d | 5855 | struct ext4_group_desc *gdp; |
498e5f24 | 5856 | unsigned int overflow; |
c9de560d AT |
5857 | ext4_grpblk_t bit; |
5858 | struct buffer_head *gd_bh; | |
5859 | ext4_group_t block_group; | |
5860 | struct ext4_sb_info *sbi; | |
5861 | struct ext4_buddy e4b; | |
84130193 | 5862 | unsigned int count_clusters; |
c9de560d AT |
5863 | int err = 0; |
5864 | int ret; | |
5865 | ||
8016e29f HS |
5866 | sbi = EXT4_SB(sb); |
5867 | ||
5868 | if (sbi->s_mount_state & EXT4_FC_REPLAY) { | |
5869 | ext4_free_blocks_simple(inode, block, count); | |
5870 | return; | |
5871 | } | |
5872 | ||
b10a44c3 | 5873 | might_sleep(); |
e6362609 TT |
5874 | if (bh) { |
5875 | if (block) | |
5876 | BUG_ON(block != bh->b_blocknr); | |
5877 | else | |
5878 | block = bh->b_blocknr; | |
5879 | } | |
c9de560d | 5880 | |
1f2acb60 | 5881 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
ce9f24cc | 5882 | !ext4_inode_block_valid(inode, block, count)) { |
12062ddd | 5883 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 5884 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
5885 | goto error_return; |
5886 | } | |
5887 | ||
0610b6e9 | 5888 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
5889 | trace_ext4_free_blocks(inode, block, count, flags); |
5890 | ||
9c02ac97 DJ |
5891 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
5892 | BUG_ON(count > 1); | |
e6362609 | 5893 | |
9c02ac97 DJ |
5894 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
5895 | inode, bh, block); | |
e6362609 TT |
5896 | } |
5897 | ||
84130193 TT |
5898 | /* |
5899 | * If the extent to be freed does not begin on a cluster | |
5900 | * boundary, we need to deal with partial clusters at the | |
5901 | * beginning and end of the extent. Normally we will free | |
5902 | * blocks at the beginning or the end unless we are explicitly | |
5903 | * requested to avoid doing so. | |
5904 | */ | |
f5a44db5 | 5905 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
5906 | if (overflow) { |
5907 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
5908 | overflow = sbi->s_cluster_ratio - overflow; | |
5909 | block += overflow; | |
5910 | if (count > overflow) | |
5911 | count -= overflow; | |
5912 | else | |
5913 | return; | |
5914 | } else { | |
5915 | block -= overflow; | |
5916 | count += overflow; | |
5917 | } | |
5918 | } | |
f5a44db5 | 5919 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
5920 | if (overflow) { |
5921 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
5922 | if (count > overflow) | |
5923 | count -= overflow; | |
5924 | else | |
5925 | return; | |
5926 | } else | |
5927 | count += sbi->s_cluster_ratio - overflow; | |
5928 | } | |
5929 | ||
9c02ac97 DJ |
5930 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
5931 | int i; | |
f96c450d | 5932 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; |
9c02ac97 DJ |
5933 | |
5934 | for (i = 0; i < count; i++) { | |
5935 | cond_resched(); | |
f96c450d DJ |
5936 | if (is_metadata) |
5937 | bh = sb_find_get_block(inode->i_sb, block + i); | |
5938 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
9c02ac97 DJ |
5939 | } |
5940 | } | |
5941 | ||
c9de560d AT |
5942 | do_more: |
5943 | overflow = 0; | |
5944 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
5945 | ||
163a203d DW |
5946 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
5947 | ext4_get_group_info(sb, block_group)))) | |
5948 | return; | |
5949 | ||
c9de560d AT |
5950 | /* |
5951 | * Check to see if we are freeing blocks across a group | |
5952 | * boundary. | |
5953 | */ | |
84130193 TT |
5954 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
5955 | overflow = EXT4_C2B(sbi, bit) + count - | |
5956 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
5957 | count -= overflow; |
5958 | } | |
810da240 | 5959 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 5960 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
5961 | if (IS_ERR(bitmap_bh)) { |
5962 | err = PTR_ERR(bitmap_bh); | |
5963 | bitmap_bh = NULL; | |
c9de560d | 5964 | goto error_return; |
ce89f46c | 5965 | } |
c9de560d | 5966 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
5967 | if (!gdp) { |
5968 | err = -EIO; | |
c9de560d | 5969 | goto error_return; |
ce89f46c | 5970 | } |
c9de560d AT |
5971 | |
5972 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
5973 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
5974 | in_range(block, ext4_inode_table(sb, gdp), | |
49598e04 | 5975 | sbi->s_itb_per_group) || |
c9de560d | 5976 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
49598e04 | 5977 | sbi->s_itb_per_group)) { |
c9de560d | 5978 | |
12062ddd | 5979 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 5980 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
5981 | /* err = 0. ext4_std_error should be a no op */ |
5982 | goto error_return; | |
c9de560d AT |
5983 | } |
5984 | ||
5985 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
188c299e JK |
5986 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
5987 | EXT4_JTR_NONE); | |
c9de560d AT |
5988 | if (err) |
5989 | goto error_return; | |
5990 | ||
5991 | /* | |
5992 | * We are about to modify some metadata. Call the journal APIs | |
5993 | * to unshare ->b_data if a currently-committing transaction is | |
5994 | * using it | |
5995 | */ | |
5996 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 5997 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
c9de560d AT |
5998 | if (err) |
5999 | goto error_return; | |
c9de560d AT |
6000 | #ifdef AGGRESSIVE_CHECK |
6001 | { | |
6002 | int i; | |
84130193 | 6003 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
6004 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
6005 | } | |
6006 | #endif | |
84130193 | 6007 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 6008 | |
adb7ef60 KK |
6009 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
6010 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
6011 | GFP_NOFS|__GFP_NOFAIL); | |
920313a7 AK |
6012 | if (err) |
6013 | goto error_return; | |
e6362609 | 6014 | |
f96c450d DJ |
6015 | /* |
6016 | * We need to make sure we don't reuse the freed block until after the | |
6017 | * transaction is committed. We make an exception if the inode is to be | |
6018 | * written in writeback mode since writeback mode has weak data | |
6019 | * consistency guarantees. | |
6020 | */ | |
6021 | if (ext4_handle_valid(handle) && | |
6022 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
6023 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
6024 | struct ext4_free_data *new_entry; |
6025 | /* | |
7444a072 MH |
6026 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
6027 | * to fail. | |
7a2fcbf7 | 6028 | */ |
7444a072 MH |
6029 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
6030 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
6031 | new_entry->efd_start_cluster = bit; |
6032 | new_entry->efd_group = block_group; | |
6033 | new_entry->efd_count = count_clusters; | |
6034 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 6035 | |
7a2fcbf7 | 6036 | ext4_lock_group(sb, block_group); |
84130193 | 6037 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 6038 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 6039 | } else { |
7a2fcbf7 AK |
6040 | /* need to update group_info->bb_free and bitmap |
6041 | * with group lock held. generate_buddy look at | |
6042 | * them with group lock_held | |
6043 | */ | |
d71c1ae2 | 6044 | if (test_opt(sb, DISCARD)) { |
a0154344 DJ |
6045 | err = ext4_issue_discard(sb, block_group, bit, count, |
6046 | NULL); | |
d71c1ae2 LC |
6047 | if (err && err != -EOPNOTSUPP) |
6048 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
6049 | " group:%d block:%d count:%lu failed" | |
6050 | " with %d", block_group, bit, count, | |
6051 | err); | |
8f9ff189 LC |
6052 | } else |
6053 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 6054 | |
955ce5f5 | 6055 | ext4_lock_group(sb, block_group); |
84130193 TT |
6056 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
6057 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
6058 | } |
6059 | ||
021b65bb TT |
6060 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
6061 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 6062 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 6063 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 6064 | ext4_unlock_group(sb, block_group); |
c9de560d | 6065 | |
772cb7c8 JS |
6066 | if (sbi->s_log_groups_per_flex) { |
6067 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f | 6068 | atomic64_add(count_clusters, |
7c990728 SJS |
6069 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6070 | flex_group)->free_clusters); | |
772cb7c8 JS |
6071 | } |
6072 | ||
9fe67149 EW |
6073 | /* |
6074 | * on a bigalloc file system, defer the s_freeclusters_counter | |
6075 | * update to the caller (ext4_remove_space and friends) so they | |
6076 | * can determine if a cluster freed here should be rereserved | |
6077 | */ | |
6078 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
6079 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
6080 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
6081 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
6082 | count_clusters); | |
6083 | } | |
7d734532 JK |
6084 | |
6085 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 6086 | |
7a2fcbf7 AK |
6087 | /* We dirtied the bitmap block */ |
6088 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6089 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6090 | ||
c9de560d AT |
6091 | /* And the group descriptor block */ |
6092 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 6093 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
6094 | if (!err) |
6095 | err = ret; | |
6096 | ||
6097 | if (overflow && !err) { | |
6098 | block += count; | |
6099 | count = overflow; | |
6100 | put_bh(bitmap_bh); | |
6101 | goto do_more; | |
6102 | } | |
c9de560d AT |
6103 | error_return: |
6104 | brelse(bitmap_bh); | |
6105 | ext4_std_error(sb, err); | |
6106 | return; | |
6107 | } | |
7360d173 | 6108 | |
2846e820 | 6109 | /** |
0529155e | 6110 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
6111 | * @handle: handle to this transaction |
6112 | * @sb: super block | |
4907cb7b | 6113 | * @block: start physical block to add to the block group |
2846e820 AG |
6114 | * @count: number of blocks to free |
6115 | * | |
e73a347b | 6116 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 6117 | */ |
cc7365df | 6118 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
6119 | ext4_fsblk_t block, unsigned long count) |
6120 | { | |
6121 | struct buffer_head *bitmap_bh = NULL; | |
6122 | struct buffer_head *gd_bh; | |
6123 | ext4_group_t block_group; | |
6124 | ext4_grpblk_t bit; | |
6125 | unsigned int i; | |
6126 | struct ext4_group_desc *desc; | |
6127 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 6128 | struct ext4_buddy e4b; |
d77147ff | 6129 | int err = 0, ret, free_clusters_count; |
6130 | ext4_grpblk_t clusters_freed; | |
6131 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); | |
6132 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
6133 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
2846e820 AG |
6134 | |
6135 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
6136 | ||
4740b830 YY |
6137 | if (count == 0) |
6138 | return 0; | |
6139 | ||
2846e820 | 6140 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
6141 | /* |
6142 | * Check to see if we are freeing blocks across a group | |
6143 | * boundary. | |
6144 | */ | |
d77147ff | 6145 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
6146 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
6147 | block_group); |
6148 | err = -EINVAL; | |
2846e820 | 6149 | goto error_return; |
cc7365df | 6150 | } |
2cd05cc3 | 6151 | |
2846e820 | 6152 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
9008a58e DW |
6153 | if (IS_ERR(bitmap_bh)) { |
6154 | err = PTR_ERR(bitmap_bh); | |
6155 | bitmap_bh = NULL; | |
2846e820 | 6156 | goto error_return; |
cc7365df YY |
6157 | } |
6158 | ||
2846e820 | 6159 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
6160 | if (!desc) { |
6161 | err = -EIO; | |
2846e820 | 6162 | goto error_return; |
cc7365df | 6163 | } |
2846e820 AG |
6164 | |
6165 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
6166 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
6167 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
6168 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
6169 | sbi->s_itb_per_group)) { | |
6170 | ext4_error(sb, "Adding blocks in system zones - " | |
6171 | "Block = %llu, count = %lu", | |
6172 | block, count); | |
cc7365df | 6173 | err = -EINVAL; |
2846e820 AG |
6174 | goto error_return; |
6175 | } | |
6176 | ||
2cd05cc3 | 6177 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
188c299e JK |
6178 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
6179 | EXT4_JTR_NONE); | |
2846e820 AG |
6180 | if (err) |
6181 | goto error_return; | |
6182 | ||
6183 | /* | |
6184 | * We are about to modify some metadata. Call the journal APIs | |
6185 | * to unshare ->b_data if a currently-committing transaction is | |
6186 | * using it | |
6187 | */ | |
6188 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
188c299e | 6189 | err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); |
2846e820 AG |
6190 | if (err) |
6191 | goto error_return; | |
e73a347b | 6192 | |
d77147ff | 6193 | for (i = 0, clusters_freed = 0; i < cluster_count; i++) { |
2846e820 | 6194 | BUFFER_TRACE(bitmap_bh, "clear bit"); |
e73a347b | 6195 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
6196 | ext4_error(sb, "bit already cleared for block %llu", |
6197 | (ext4_fsblk_t)(block + i)); | |
6198 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
6199 | } else { | |
d77147ff | 6200 | clusters_freed++; |
2846e820 AG |
6201 | } |
6202 | } | |
e73a347b AG |
6203 | |
6204 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
6205 | if (err) | |
6206 | goto error_return; | |
6207 | ||
6208 | /* | |
6209 | * need to update group_info->bb_free and bitmap | |
6210 | * with group lock held. generate_buddy look at | |
6211 | * them with group lock_held | |
6212 | */ | |
2846e820 | 6213 | ext4_lock_group(sb, block_group); |
d77147ff | 6214 | mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); |
6215 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
6216 | free_clusters_count = clusters_freed + | |
6217 | ext4_free_group_clusters(sb, desc); | |
6218 | ext4_free_group_clusters_set(sb, desc, free_clusters_count); | |
79f1ba49 | 6219 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 6220 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 6221 | ext4_unlock_group(sb, block_group); |
57042651 | 6222 | percpu_counter_add(&sbi->s_freeclusters_counter, |
d77147ff | 6223 | clusters_freed); |
2846e820 AG |
6224 | |
6225 | if (sbi->s_log_groups_per_flex) { | |
6226 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
d77147ff | 6227 | atomic64_add(clusters_freed, |
7c990728 SJS |
6228 | &sbi_array_rcu_deref(sbi, s_flex_groups, |
6229 | flex_group)->free_clusters); | |
2846e820 | 6230 | } |
e73a347b AG |
6231 | |
6232 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
6233 | |
6234 | /* We dirtied the bitmap block */ | |
6235 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
6236 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
6237 | ||
6238 | /* And the group descriptor block */ | |
6239 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
6240 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
6241 | if (!err) | |
6242 | err = ret; | |
6243 | ||
6244 | error_return: | |
6245 | brelse(bitmap_bh); | |
6246 | ext4_std_error(sb, err); | |
cc7365df | 6247 | return err; |
2846e820 AG |
6248 | } |
6249 | ||
7360d173 LC |
6250 | /** |
6251 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
6252 | * @sb: super block for the file system | |
6253 | * @start: starting block of the free extent in the alloc. group | |
6254 | * @count: number of blocks to TRIM | |
7360d173 LC |
6255 | * @e4b: ext4 buddy for the group |
6256 | * | |
6257 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
6258 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
6259 | * be called with under the group lock. | |
6260 | */ | |
bd2eea8d WJ |
6261 | static int ext4_trim_extent(struct super_block *sb, |
6262 | int start, int count, struct ext4_buddy *e4b) | |
e2cbd587 | 6263 | __releases(bitlock) |
6264 | __acquires(bitlock) | |
7360d173 LC |
6265 | { |
6266 | struct ext4_free_extent ex; | |
bd2eea8d | 6267 | ext4_group_t group = e4b->bd_group; |
d71c1ae2 | 6268 | int ret = 0; |
7360d173 | 6269 | |
b3d4c2b1 TM |
6270 | trace_ext4_trim_extent(sb, group, start, count); |
6271 | ||
7360d173 LC |
6272 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
6273 | ||
6274 | ex.fe_start = start; | |
6275 | ex.fe_group = group; | |
6276 | ex.fe_len = count; | |
6277 | ||
6278 | /* | |
6279 | * Mark blocks used, so no one can reuse them while | |
6280 | * being trimmed. | |
6281 | */ | |
6282 | mb_mark_used(e4b, &ex); | |
6283 | ext4_unlock_group(sb, group); | |
a0154344 | 6284 | ret = ext4_issue_discard(sb, group, start, count, NULL); |
7360d173 LC |
6285 | ext4_lock_group(sb, group); |
6286 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 6287 | return ret; |
7360d173 LC |
6288 | } |
6289 | ||
6920b391 WJ |
6290 | static int ext4_try_to_trim_range(struct super_block *sb, |
6291 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
6292 | ext4_grpblk_t max, ext4_grpblk_t minblocks) | |
a5fda113 TT |
6293 | __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) |
6294 | __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) | |
6920b391 WJ |
6295 | { |
6296 | ext4_grpblk_t next, count, free_count; | |
6297 | void *bitmap; | |
6920b391 WJ |
6298 | |
6299 | bitmap = e4b->bd_bitmap; | |
6300 | start = (e4b->bd_info->bb_first_free > start) ? | |
6301 | e4b->bd_info->bb_first_free : start; | |
6302 | count = 0; | |
6303 | free_count = 0; | |
6304 | ||
6305 | while (start <= max) { | |
6306 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
6307 | if (start > max) | |
6308 | break; | |
6309 | next = mb_find_next_bit(bitmap, max + 1, start); | |
6310 | ||
6311 | if ((next - start) >= minblocks) { | |
afcc4e32 LB |
6312 | int ret = ext4_trim_extent(sb, start, next - start, e4b); |
6313 | ||
6920b391 WJ |
6314 | if (ret && ret != -EOPNOTSUPP) |
6315 | break; | |
6920b391 WJ |
6316 | count += next - start; |
6317 | } | |
6318 | free_count += next - start; | |
6319 | start = next + 1; | |
6320 | ||
6321 | if (fatal_signal_pending(current)) { | |
6322 | count = -ERESTARTSYS; | |
6323 | break; | |
6324 | } | |
6325 | ||
6326 | if (need_resched()) { | |
6327 | ext4_unlock_group(sb, e4b->bd_group); | |
6328 | cond_resched(); | |
6329 | ext4_lock_group(sb, e4b->bd_group); | |
6330 | } | |
6331 | ||
6332 | if ((e4b->bd_info->bb_free - free_count) < minblocks) | |
6333 | break; | |
6334 | } | |
6335 | ||
6336 | return count; | |
6337 | } | |
6338 | ||
7360d173 LC |
6339 | /** |
6340 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
6341 | * @sb: super block for file system | |
22612283 | 6342 | * @group: group to be trimmed |
7360d173 LC |
6343 | * @start: first group block to examine |
6344 | * @max: last group block to examine | |
6345 | * @minblocks: minimum extent block count | |
6346 | * | |
7360d173 LC |
6347 | * ext4_trim_all_free walks through group's block bitmap searching for free |
6348 | * extents. When the free extent is found, mark it as used in group buddy | |
6349 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
b6f5558c | 6350 | * the group buddy bitmap. |
7360d173 | 6351 | */ |
0b75a840 | 6352 | static ext4_grpblk_t |
78944086 LC |
6353 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
6354 | ext4_grpblk_t start, ext4_grpblk_t max, | |
6355 | ext4_grpblk_t minblocks) | |
7360d173 | 6356 | { |
78944086 | 6357 | struct ext4_buddy e4b; |
6920b391 | 6358 | int ret; |
7360d173 | 6359 | |
b3d4c2b1 TM |
6360 | trace_ext4_trim_all_free(sb, group, start, max); |
6361 | ||
78944086 LC |
6362 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
6363 | if (ret) { | |
9651e6b2 KK |
6364 | ext4_warning(sb, "Error %d loading buddy information for %u", |
6365 | ret, group); | |
78944086 LC |
6366 | return ret; |
6367 | } | |
28739eea LC |
6368 | |
6369 | ext4_lock_group(sb, group); | |
7360d173 | 6370 | |
6920b391 | 6371 | if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || |
2327fb2e | 6372 | minblocks < EXT4_SB(sb)->s_last_trim_minblks) { |
6920b391 WJ |
6373 | ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); |
6374 | if (ret >= 0) | |
6375 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); | |
6376 | } else { | |
6377 | ret = 0; | |
7360d173 | 6378 | } |
3d56b8d2 | 6379 | |
7360d173 | 6380 | ext4_unlock_group(sb, group); |
78944086 | 6381 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
6382 | |
6383 | ext4_debug("trimmed %d blocks in the group %d\n", | |
6920b391 | 6384 | ret, group); |
7360d173 | 6385 | |
d71c1ae2 | 6386 | return ret; |
7360d173 LC |
6387 | } |
6388 | ||
6389 | /** | |
6390 | * ext4_trim_fs() -- trim ioctl handle function | |
6391 | * @sb: superblock for filesystem | |
6392 | * @range: fstrim_range structure | |
6393 | * | |
6394 | * start: First Byte to trim | |
6395 | * len: number of Bytes to trim from start | |
6396 | * minlen: minimum extent length in Bytes | |
6397 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
6398 | * start to start+len. For each such a group ext4_trim_all_free function | |
6399 | * is invoked to trim all free space. | |
6400 | */ | |
6401 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
6402 | { | |
173b6e38 | 6403 | struct request_queue *q = bdev_get_queue(sb->s_bdev); |
78944086 | 6404 | struct ext4_group_info *grp; |
913eed83 | 6405 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 6406 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 6407 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
6408 | ext4_fsblk_t first_data_blk = |
6409 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 6410 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
6411 | int ret = 0; |
6412 | ||
6413 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 6414 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
6415 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
6416 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 6417 | |
5de35e8d LC |
6418 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
6419 | start >= max_blks || | |
6420 | range->len < sb->s_blocksize) | |
7360d173 | 6421 | return -EINVAL; |
173b6e38 JK |
6422 | /* No point to try to trim less than discard granularity */ |
6423 | if (range->minlen < q->limits.discard_granularity) { | |
6424 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), | |
6425 | q->limits.discard_granularity >> sb->s_blocksize_bits); | |
6426 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) | |
6427 | goto out; | |
6428 | } | |
913eed83 LC |
6429 | if (end >= max_blks) |
6430 | end = max_blks - 1; | |
6431 | if (end <= first_data_blk) | |
22f10457 | 6432 | goto out; |
913eed83 | 6433 | if (start < first_data_blk) |
0f0a25bf | 6434 | start = first_data_blk; |
7360d173 | 6435 | |
913eed83 | 6436 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 6437 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 6438 | &first_group, &first_cluster); |
913eed83 | 6439 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 6440 | &last_group, &last_cluster); |
7360d173 | 6441 | |
913eed83 LC |
6442 | /* end now represents the last cluster to discard in this group */ |
6443 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
6444 | |
6445 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
6446 | grp = ext4_get_group_info(sb, group); |
6447 | /* We only do this if the grp has never been initialized */ | |
6448 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 6449 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
6450 | if (ret) |
6451 | break; | |
7360d173 LC |
6452 | } |
6453 | ||
0ba08517 | 6454 | /* |
913eed83 LC |
6455 | * For all the groups except the last one, last cluster will |
6456 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
6457 | * change it for the last group, note that last_cluster is | |
6458 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 6459 | */ |
913eed83 LC |
6460 | if (group == last_group) |
6461 | end = last_cluster; | |
7360d173 | 6462 | |
78944086 | 6463 | if (grp->bb_free >= minlen) { |
7137d7a4 | 6464 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 6465 | end, minlen); |
7360d173 LC |
6466 | if (cnt < 0) { |
6467 | ret = cnt; | |
7360d173 LC |
6468 | break; |
6469 | } | |
21e7fd22 | 6470 | trimmed += cnt; |
7360d173 | 6471 | } |
913eed83 LC |
6472 | |
6473 | /* | |
6474 | * For every group except the first one, we are sure | |
6475 | * that the first cluster to discard will be cluster #0. | |
6476 | */ | |
7137d7a4 | 6477 | first_cluster = 0; |
7360d173 | 6478 | } |
7360d173 | 6479 | |
3d56b8d2 | 6480 | if (!ret) |
2327fb2e | 6481 | EXT4_SB(sb)->s_last_trim_minblks = minlen; |
3d56b8d2 | 6482 | |
22f10457 | 6483 | out: |
aaf7d73e | 6484 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
6485 | return ret; |
6486 | } | |
0c9ec4be DW |
6487 | |
6488 | /* Iterate all the free extents in the group. */ | |
6489 | int | |
6490 | ext4_mballoc_query_range( | |
6491 | struct super_block *sb, | |
6492 | ext4_group_t group, | |
6493 | ext4_grpblk_t start, | |
6494 | ext4_grpblk_t end, | |
6495 | ext4_mballoc_query_range_fn formatter, | |
6496 | void *priv) | |
6497 | { | |
6498 | void *bitmap; | |
6499 | ext4_grpblk_t next; | |
6500 | struct ext4_buddy e4b; | |
6501 | int error; | |
6502 | ||
6503 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
6504 | if (error) | |
6505 | return error; | |
6506 | bitmap = e4b.bd_bitmap; | |
6507 | ||
6508 | ext4_lock_group(sb, group); | |
6509 | ||
6510 | start = (e4b.bd_info->bb_first_free > start) ? | |
6511 | e4b.bd_info->bb_first_free : start; | |
6512 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) | |
6513 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
6514 | ||
6515 | while (start <= end) { | |
6516 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
6517 | if (start > end) | |
6518 | break; | |
6519 | next = mb_find_next_bit(bitmap, end + 1, start); | |
6520 | ||
6521 | ext4_unlock_group(sb, group); | |
6522 | error = formatter(sb, group, start, next - start, priv); | |
6523 | if (error) | |
6524 | goto out_unload; | |
6525 | ext4_lock_group(sb, group); | |
6526 | ||
6527 | start = next + 1; | |
6528 | } | |
6529 | ||
6530 | ext4_unlock_group(sb, group); | |
6531 | out_unload: | |
6532 | ext4_mb_unload_buddy(&e4b); | |
6533 | ||
6534 | return error; | |
6535 | } |