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> |
5229a658 | 19 | #include <linux/freezer.h> |
9bffad1e | 20 | #include <trace/events/ext4.h> |
bdefd689 | 21 | #include <kunit/static_stub.h> |
9bffad1e | 22 | |
c9de560d AT |
23 | /* |
24 | * MUSTDO: | |
25 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
26 | * - search for metadata in few groups | |
27 | * | |
28 | * TODO v4: | |
29 | * - normalization should take into account whether file is still open | |
30 | * - discard preallocations if no free space left (policy?) | |
31 | * - don't normalize tails | |
32 | * - quota | |
33 | * - reservation for superuser | |
34 | * | |
35 | * TODO v3: | |
36 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
37 | * - track min/max extents in each group for better group selection | |
38 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
39 | * - tree of groups sorted by number of free blocks | |
40 | * - error handling | |
41 | */ | |
42 | ||
43 | /* | |
44 | * The allocation request involve request for multiple number of blocks | |
45 | * near to the goal(block) value specified. | |
46 | * | |
b713a5ec TT |
47 | * During initialization phase of the allocator we decide to use the |
48 | * group preallocation or inode preallocation depending on the size of | |
49 | * the file. The size of the file could be the resulting file size we | |
50 | * would have after allocation, or the current file size, which ever | |
51 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
52 | * select to use the group preallocation. The default value of | |
53 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
54 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
55 | * terms of number of blocks. | |
c9de560d AT |
56 | * |
57 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 58 | * ensure that we have small files closer together on the disk. |
c9de560d | 59 | * |
b713a5ec TT |
60 | * First stage the allocator looks at the inode prealloc list, |
61 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
62 | * spaces for this particular inode. The inode prealloc space is | |
63 | * represented as: | |
c9de560d AT |
64 | * |
65 | * pa_lstart -> the logical start block for this prealloc space | |
66 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
67 | * pa_len -> length for this prealloc space (in clusters) |
68 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
69 | * |
70 | * The inode preallocation space is used looking at the _logical_ start | |
71 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
72 | * space we will consume the particular prealloc space. This makes sure that |
73 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
74 | * |
75 | * The important thing to be noted in case of inode prealloc space is that | |
76 | * we don't modify the values associated to inode prealloc space except | |
77 | * pa_free. | |
78 | * | |
79 | * If we are not able to find blocks in the inode prealloc space and if we | |
80 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 81 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
82 | * |
83 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
84 | * | |
85 | * The reason for having a per cpu locality group is to reduce the contention | |
86 | * between CPUs. It is possible to get scheduled at this point. | |
87 | * | |
88 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 89 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
90 | * |
91 | * If we can't allocate blocks via inode prealloc or/and locality group | |
92 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
93 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
94 | * mapped to the buddy and bitmap information regarding different | |
95 | * groups. The buddy information is attached to buddy cache inode so that | |
96 | * we can access them through the page cache. The information regarding | |
97 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
98 | * block bitmap and buddy information. The information are stored in the | |
99 | * inode as: | |
100 | * | |
101 | * { page } | |
c3a326a6 | 102 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
103 | * |
104 | * | |
105 | * one block each for bitmap and buddy information. So for each group we | |
ea1754a0 | 106 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / |
c9de560d AT |
107 | * blocksize) blocks. So it can have information regarding groups_per_page |
108 | * which is blocks_per_page/2 | |
109 | * | |
110 | * The buddy cache inode is not stored on disk. The inode is thrown | |
111 | * away when the filesystem is unmounted. | |
112 | * | |
113 | * We look for count number of blocks in the buddy cache. If we were able | |
114 | * to locate that many free blocks we return with additional information | |
115 | * regarding rest of the contiguous physical block available | |
116 | * | |
117 | * Before allocating blocks via buddy cache we normalize the request | |
118 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
119 | * blocks that we get after allocation is added to the respective prealloc | |
120 | * list. In case of inode preallocation we follow a list of heuristics | |
121 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
122 | * we are doing a group prealloc we try to normalize the request to | |
27baebb8 TT |
123 | * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is |
124 | * dependent on the cluster size; for non-bigalloc file systems, it is | |
c9de560d | 125 | * 512 blocks. This can be tuned via |
d7a1fee1 | 126 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
127 | * terms of number of blocks. If we have mounted the file system with -O |
128 | * stripe=<value> option the group prealloc request is normalized to the | |
b483bb77 | 129 | * smallest multiple of the stripe value (sbi->s_stripe) which is |
d7a1fee1 | 130 | * greater than the default mb_group_prealloc. |
c9de560d | 131 | * |
196e402a HS |
132 | * If "mb_optimize_scan" mount option is set, we maintain in memory group info |
133 | * structures in two data structures: | |
134 | * | |
135 | * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders) | |
136 | * | |
137 | * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks) | |
138 | * | |
139 | * This is an array of lists where the index in the array represents the | |
140 | * largest free order in the buddy bitmap of the participating group infos of | |
141 | * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total | |
142 | * number of buddy bitmap orders possible) number of lists. Group-infos are | |
143 | * placed in appropriate lists. | |
144 | * | |
83e80a6e | 145 | * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size) |
196e402a | 146 | * |
83e80a6e | 147 | * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks) |
196e402a | 148 | * |
83e80a6e JK |
149 | * This is an array of lists where in the i-th list there are groups with |
150 | * average fragment size >= 2^i and < 2^(i+1). The average fragment size | |
151 | * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments. | |
152 | * Note that we don't bother with a special list for completely empty groups | |
153 | * so we only have MB_NUM_ORDERS(sb) lists. | |
196e402a HS |
154 | * |
155 | * When "mb_optimize_scan" mount option is set, mballoc consults the above data | |
156 | * structures to decide the order in which groups are to be traversed for | |
157 | * fulfilling an allocation request. | |
158 | * | |
f52f3d2b OM |
159 | * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order |
160 | * >= the order of the request. We directly look at the largest free order list | |
161 | * in the data structure (1) above where largest_free_order = order of the | |
162 | * request. If that list is empty, we look at remaining list in the increasing | |
163 | * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED | |
164 | * lookup in O(1) time. | |
165 | * | |
166 | * At CR_GOAL_LEN_FAST, we only consider groups where | |
167 | * average fragment size > request size. So, we lookup a group which has average | |
168 | * fragment size just above or equal to request size using our average fragment | |
169 | * size group lists (data structure 2) in O(1) time. | |
170 | * | |
171 | * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied | |
172 | * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in | |
173 | * CR_GOAL_LEN_FAST suggests that there is no BG that has avg | |
174 | * fragment size > goal length. So before falling to the slower | |
175 | * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and | |
176 | * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big | |
177 | * enough average fragment size. This increases the chances of finding a | |
178 | * suitable block group in O(1) time and results in faster allocation at the | |
179 | * cost of reduced size of allocation. | |
7e170922 | 180 | * |
196e402a | 181 | * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in |
f52f3d2b OM |
182 | * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and |
183 | * CR_GOAL_LEN_FAST phase. | |
196e402a | 184 | * |
d7a1fee1 | 185 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 186 | * |
b713a5ec TT |
187 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
188 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
189 | * /sys/fs/ext4/<partition>/mb_order2_req | |
196e402a | 190 | * /sys/fs/ext4/<partition>/mb_linear_limit |
c9de560d | 191 | * |
b713a5ec | 192 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
193 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
194 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 195 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 196 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
197 | * stripe size. This should result in better allocation on RAID setups. If |
198 | * not, we search in the specific group using bitmap for best extents. The | |
199 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 200 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 201 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
202 | * best extent in the found extents. Searching for the blocks starts with |
203 | * the group specified as the goal value in allocation context via | |
204 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 205 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
206 | * checked. |
207 | * | |
196e402a HS |
208 | * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not |
209 | * get traversed linearly. That may result in subsequent allocations being not | |
210 | * close to each other. And so, the underlying device may get filled up in a | |
211 | * non-linear fashion. While that may not matter on non-rotational devices, for | |
212 | * rotational devices that may result in higher seek times. "mb_linear_limit" | |
213 | * tells mballoc how many groups mballoc should search linearly before | |
214 | * performing consulting above data structures for more efficient lookups. For | |
215 | * non rotational devices, this value defaults to 0 and for rotational devices | |
216 | * this is set to MB_DEFAULT_LINEAR_LIMIT. | |
217 | * | |
c9de560d AT |
218 | * Both the prealloc space are getting populated as above. So for the first |
219 | * request we will hit the buddy cache which will result in this prealloc | |
220 | * space getting filled. The prealloc space is then later used for the | |
221 | * subsequent request. | |
222 | */ | |
223 | ||
224 | /* | |
225 | * mballoc operates on the following data: | |
226 | * - on-disk bitmap | |
227 | * - in-core buddy (actually includes buddy and bitmap) | |
228 | * - preallocation descriptors (PAs) | |
229 | * | |
230 | * there are two types of preallocations: | |
231 | * - inode | |
232 | * assiged to specific inode and can be used for this inode only. | |
233 | * it describes part of inode's space preallocated to specific | |
234 | * physical blocks. any block from that preallocated can be used | |
235 | * independent. the descriptor just tracks number of blocks left | |
236 | * unused. so, before taking some block from descriptor, one must | |
237 | * make sure corresponded logical block isn't allocated yet. this | |
238 | * also means that freeing any block within descriptor's range | |
239 | * must discard all preallocated blocks. | |
240 | * - locality group | |
241 | * assigned to specific locality group which does not translate to | |
242 | * permanent set of inodes: inode can join and leave group. space | |
243 | * from this type of preallocation can be used for any inode. thus | |
244 | * it's consumed from the beginning to the end. | |
245 | * | |
246 | * relation between them can be expressed as: | |
247 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
248 | * | |
249 | * this mean blocks mballoc considers used are: | |
250 | * - allocated blocks (persistent) | |
251 | * - preallocated blocks (non-persistent) | |
252 | * | |
253 | * consistency in mballoc world means that at any time a block is either | |
254 | * free or used in ALL structures. notice: "any time" should not be read | |
255 | * literally -- time is discrete and delimited by locks. | |
256 | * | |
257 | * to keep it simple, we don't use block numbers, instead we count number of | |
258 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
259 | * | |
260 | * all operations can be expressed as: | |
261 | * - init buddy: buddy = on-disk + PAs | |
262 | * - new PA: buddy += N; PA = N | |
263 | * - use inode PA: on-disk += N; PA -= N | |
264 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
265 | * - use locality group PA on-disk += N; PA -= N | |
266 | * - discard locality group PA buddy -= PA; PA = 0 | |
267 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
268 | * is used in real operation because we can't know actual used | |
269 | * bits from PA, only from on-disk bitmap | |
270 | * | |
271 | * if we follow this strict logic, then all operations above should be atomic. | |
272 | * given some of them can block, we'd have to use something like semaphores | |
273 | * killing performance on high-end SMP hardware. let's try to relax it using | |
274 | * the following knowledge: | |
275 | * 1) if buddy is referenced, it's already initialized | |
276 | * 2) while block is used in buddy and the buddy is referenced, | |
277 | * nobody can re-allocate that block | |
278 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
279 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
280 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
281 | * block | |
282 | * | |
283 | * so, now we're building a concurrency table: | |
284 | * - init buddy vs. | |
285 | * - new PA | |
286 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
287 | * until PA is linked to allocation group to avoid concurrent buddy init | |
288 | * - use inode PA | |
289 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
290 | * given (3) we care that PA-=N operation doesn't interfere with init | |
291 | * - discard inode PA | |
292 | * the simplest way would be to have buddy initialized by the discard | |
293 | * - use locality group PA | |
294 | * again PA-=N must be serialized with init | |
295 | * - discard locality group PA | |
296 | * the simplest way would be to have buddy initialized by the discard | |
297 | * - new PA vs. | |
298 | * - use inode PA | |
299 | * i_data_sem serializes them | |
300 | * - discard inode PA | |
301 | * discard process must wait until PA isn't used by another process | |
302 | * - use locality group PA | |
303 | * some mutex should serialize them | |
304 | * - discard locality group PA | |
305 | * discard process must wait until PA isn't used by another process | |
306 | * - use inode PA | |
307 | * - use inode PA | |
308 | * i_data_sem or another mutex should serializes them | |
309 | * - discard inode PA | |
310 | * discard process must wait until PA isn't used by another process | |
311 | * - use locality group PA | |
312 | * nothing wrong here -- they're different PAs covering different blocks | |
313 | * - discard locality group PA | |
314 | * discard process must wait until PA isn't used by another process | |
315 | * | |
316 | * now we're ready to make few consequences: | |
317 | * - PA is referenced and while it is no discard is possible | |
318 | * - PA is referenced until block isn't marked in on-disk bitmap | |
319 | * - PA changes only after on-disk bitmap | |
320 | * - discard must not compete with init. either init is done before | |
321 | * any discard or they're serialized somehow | |
322 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
323 | * | |
324 | * a special case when we've used PA to emptiness. no need to modify buddy | |
325 | * in this case, but we should care about concurrent init | |
326 | * | |
327 | */ | |
328 | ||
329 | /* | |
330 | * Logic in few words: | |
331 | * | |
332 | * - allocation: | |
333 | * load group | |
334 | * find blocks | |
335 | * mark bits in on-disk bitmap | |
336 | * release group | |
337 | * | |
338 | * - use preallocation: | |
339 | * find proper PA (per-inode or group) | |
340 | * load group | |
341 | * mark bits in on-disk bitmap | |
342 | * release group | |
343 | * release PA | |
344 | * | |
345 | * - free: | |
346 | * load group | |
347 | * mark bits in on-disk bitmap | |
348 | * release group | |
349 | * | |
350 | * - discard preallocations in group: | |
351 | * mark PAs deleted | |
352 | * move them onto local list | |
353 | * load on-disk bitmap | |
354 | * load group | |
355 | * remove PA from object (inode or locality group) | |
356 | * mark free blocks in-core | |
357 | * | |
358 | * - discard inode's preallocations: | |
359 | */ | |
360 | ||
361 | /* | |
362 | * Locking rules | |
363 | * | |
364 | * Locks: | |
365 | * - bitlock on a group (group) | |
366 | * - object (inode/locality) (object) | |
367 | * - per-pa lock (pa) | |
f52f3d2b OM |
368 | * - cr_power2_aligned lists lock (cr_power2_aligned) |
369 | * - cr_goal_len_fast lists lock (cr_goal_len_fast) | |
c9de560d AT |
370 | * |
371 | * Paths: | |
372 | * - new pa | |
373 | * object | |
374 | * group | |
375 | * | |
376 | * - find and use pa: | |
377 | * pa | |
378 | * | |
379 | * - release consumed pa: | |
380 | * pa | |
381 | * group | |
382 | * object | |
383 | * | |
384 | * - generate in-core bitmap: | |
385 | * group | |
386 | * pa | |
387 | * | |
388 | * - discard all for given object (inode, locality group): | |
389 | * object | |
390 | * pa | |
391 | * group | |
392 | * | |
393 | * - discard all for given group: | |
394 | * group | |
395 | * pa | |
396 | * group | |
397 | * object | |
398 | * | |
196e402a HS |
399 | * - allocation path (ext4_mb_regular_allocator) |
400 | * group | |
f52f3d2b | 401 | * cr_power2_aligned/cr_goal_len_fast |
c9de560d | 402 | */ |
c3a326a6 AK |
403 | static struct kmem_cache *ext4_pspace_cachep; |
404 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 405 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
406 | |
407 | /* We create slab caches for groupinfo data structures based on the | |
408 | * superblock block size. There will be one per mounted filesystem for | |
409 | * each unique s_blocksize_bits */ | |
2892c15d | 410 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
411 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
412 | ||
d6006186 | 413 | static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
2892c15d ES |
414 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", |
415 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
416 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
417 | }; | |
418 | ||
c3a326a6 AK |
419 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
420 | ext4_group_t group); | |
53f86b17 | 421 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); |
c3a326a6 | 422 | |
196e402a | 423 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
4eb7a4a1 | 424 | ext4_group_t group, enum criteria cr); |
196e402a | 425 | |
55cdd0af WJ |
426 | static int ext4_try_to_trim_range(struct super_block *sb, |
427 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
428 | ext4_grpblk_t max, ext4_grpblk_t minblocks); | |
429 | ||
07b5b8e1 RH |
430 | /* |
431 | * The algorithm using this percpu seq counter goes below: | |
432 | * 1. We sample the percpu discard_pa_seq counter before trying for block | |
433 | * allocation in ext4_mb_new_blocks(). | |
434 | * 2. We increment this percpu discard_pa_seq counter when we either allocate | |
435 | * or free these blocks i.e. while marking those blocks as used/free in | |
436 | * mb_mark_used()/mb_free_blocks(). | |
437 | * 3. We also increment this percpu seq counter when we successfully identify | |
438 | * that the bb_prealloc_list is not empty and hence proceed for discarding | |
439 | * of those PAs inside ext4_mb_discard_group_preallocations(). | |
440 | * | |
441 | * Now to make sure that the regular fast path of block allocation is not | |
442 | * affected, as a small optimization we only sample the percpu seq counter | |
443 | * on that cpu. Only when the block allocation fails and when freed blocks | |
444 | * found were 0, that is when we sample percpu seq counter for all cpus using | |
445 | * below function ext4_get_discard_pa_seq_sum(). This happens after making | |
446 | * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. | |
447 | */ | |
448 | static DEFINE_PER_CPU(u64, discard_pa_seq); | |
449 | static inline u64 ext4_get_discard_pa_seq_sum(void) | |
450 | { | |
451 | int __cpu; | |
452 | u64 __seq = 0; | |
453 | ||
454 | for_each_possible_cpu(__cpu) | |
455 | __seq += per_cpu(discard_pa_seq, __cpu); | |
456 | return __seq; | |
457 | } | |
458 | ||
ffad0a44 AK |
459 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
460 | { | |
c9de560d | 461 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
462 | *bit += ((unsigned long) addr & 7UL) << 3; |
463 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 464 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
465 | *bit += ((unsigned long) addr & 3UL) << 3; |
466 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
467 | #else |
468 | #error "how many bits you are?!" | |
469 | #endif | |
ffad0a44 AK |
470 | return addr; |
471 | } | |
c9de560d AT |
472 | |
473 | static inline int mb_test_bit(int bit, void *addr) | |
474 | { | |
475 | /* | |
476 | * ext4_test_bit on architecture like powerpc | |
477 | * needs unsigned long aligned address | |
478 | */ | |
ffad0a44 | 479 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
480 | return ext4_test_bit(bit, addr); |
481 | } | |
482 | ||
483 | static inline void mb_set_bit(int bit, void *addr) | |
484 | { | |
ffad0a44 | 485 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
486 | ext4_set_bit(bit, addr); |
487 | } | |
488 | ||
c9de560d AT |
489 | static inline void mb_clear_bit(int bit, void *addr) |
490 | { | |
ffad0a44 | 491 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
492 | ext4_clear_bit(bit, addr); |
493 | } | |
494 | ||
eabe0444 AS |
495 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
496 | { | |
497 | addr = mb_correct_addr_and_bit(&bit, addr); | |
498 | return ext4_test_and_clear_bit(bit, addr); | |
499 | } | |
500 | ||
ffad0a44 AK |
501 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
502 | { | |
e7dfb246 | 503 | int fix = 0, ret, tmpmax; |
ffad0a44 | 504 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 505 | tmpmax = max + fix; |
ffad0a44 AK |
506 | start += fix; |
507 | ||
e7dfb246 AK |
508 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
509 | if (ret > max) | |
510 | return max; | |
511 | return ret; | |
ffad0a44 AK |
512 | } |
513 | ||
514 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
515 | { | |
e7dfb246 | 516 | int fix = 0, ret, tmpmax; |
ffad0a44 | 517 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 518 | tmpmax = max + fix; |
ffad0a44 AK |
519 | start += fix; |
520 | ||
e7dfb246 AK |
521 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
522 | if (ret > max) | |
523 | return max; | |
524 | return ret; | |
ffad0a44 AK |
525 | } |
526 | ||
c9de560d AT |
527 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
528 | { | |
529 | char *bb; | |
530 | ||
c5e8f3f3 | 531 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
532 | BUG_ON(max == NULL); |
533 | ||
534 | if (order > e4b->bd_blkbits + 1) { | |
535 | *max = 0; | |
536 | return NULL; | |
537 | } | |
538 | ||
539 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
540 | if (order == 0) { |
541 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 542 | return e4b->bd_bitmap; |
84b775a3 | 543 | } |
c9de560d | 544 | |
c5e8f3f3 | 545 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
546 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
547 | ||
548 | return bb; | |
549 | } | |
550 | ||
551 | #ifdef DOUBLE_CHECK | |
552 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
553 | int first, int count) | |
554 | { | |
555 | int i; | |
556 | struct super_block *sb = e4b->bd_sb; | |
557 | ||
558 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
559 | return; | |
bc8e6740 | 560 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
561 | for (i = 0; i < count; i++) { |
562 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
563 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
564 | |
565 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 566 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
c5f3a382 BL |
567 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
568 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
5d1b1b3f | 569 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
570 | inode ? inode->i_ino : 0, |
571 | blocknr, | |
572 | "freeing block already freed " | |
573 | "(bit %u)", | |
574 | first + i); | |
c9de560d AT |
575 | } |
576 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
577 | } | |
578 | } | |
579 | ||
580 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
581 | { | |
582 | int i; | |
583 | ||
584 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
585 | return; | |
bc8e6740 | 586 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
587 | for (i = 0; i < count; i++) { |
588 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
589 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
590 | } | |
591 | } | |
592 | ||
593 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
594 | { | |
eb2b8ebb RH |
595 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
596 | return; | |
c9de560d AT |
597 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { |
598 | unsigned char *b1, *b2; | |
599 | int i; | |
600 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
601 | b2 = (unsigned char *) bitmap; | |
602 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
603 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
604 | ext4_msg(e4b->bd_sb, KERN_ERR, |
605 | "corruption in group %u " | |
606 | "at byte %u(%u): %x in copy != %x " | |
607 | "on disk/prealloc", | |
608 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
609 | BUG(); |
610 | } | |
611 | } | |
612 | } | |
613 | } | |
614 | ||
a3450215 RH |
615 | static void mb_group_bb_bitmap_alloc(struct super_block *sb, |
616 | struct ext4_group_info *grp, ext4_group_t group) | |
617 | { | |
618 | struct buffer_head *bh; | |
619 | ||
620 | grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); | |
eb2b8ebb RH |
621 | if (!grp->bb_bitmap) |
622 | return; | |
a3450215 RH |
623 | |
624 | bh = ext4_read_block_bitmap(sb, group); | |
eb2b8ebb RH |
625 | if (IS_ERR_OR_NULL(bh)) { |
626 | kfree(grp->bb_bitmap); | |
627 | grp->bb_bitmap = NULL; | |
628 | return; | |
629 | } | |
a3450215 RH |
630 | |
631 | memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); | |
632 | put_bh(bh); | |
633 | } | |
634 | ||
635 | static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
636 | { | |
637 | kfree(grp->bb_bitmap); | |
638 | } | |
639 | ||
c9de560d AT |
640 | #else |
641 | static inline void mb_free_blocks_double(struct inode *inode, | |
642 | struct ext4_buddy *e4b, int first, int count) | |
643 | { | |
644 | return; | |
645 | } | |
646 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
647 | int first, int count) | |
648 | { | |
649 | return; | |
650 | } | |
651 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
652 | { | |
653 | return; | |
654 | } | |
a3450215 RH |
655 | |
656 | static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, | |
657 | struct ext4_group_info *grp, ext4_group_t group) | |
658 | { | |
659 | return; | |
660 | } | |
661 | ||
662 | static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) | |
663 | { | |
664 | return; | |
665 | } | |
c9de560d AT |
666 | #endif |
667 | ||
668 | #ifdef AGGRESSIVE_CHECK | |
669 | ||
670 | #define MB_CHECK_ASSERT(assert) \ | |
671 | do { \ | |
672 | if (!(assert)) { \ | |
673 | printk(KERN_EMERG \ | |
674 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
675 | function, file, line, # assert); \ | |
676 | BUG(); \ | |
677 | } \ | |
678 | } while (0) | |
679 | ||
133de5a0 | 680 | static void __mb_check_buddy(struct ext4_buddy *e4b, char *file, |
c9de560d AT |
681 | const char *function, int line) |
682 | { | |
683 | struct super_block *sb = e4b->bd_sb; | |
684 | int order = e4b->bd_blkbits + 1; | |
685 | int max; | |
686 | int max2; | |
687 | int i; | |
688 | int j; | |
689 | int k; | |
690 | int count; | |
691 | struct ext4_group_info *grp; | |
692 | int fragments = 0; | |
693 | int fstart; | |
694 | struct list_head *cur; | |
695 | void *buddy; | |
696 | void *buddy2; | |
697 | ||
addd752c | 698 | if (e4b->bd_info->bb_check_counter++ % 10) |
133de5a0 | 699 | return; |
c9de560d AT |
700 | |
701 | while (order > 1) { | |
702 | buddy = mb_find_buddy(e4b, order, &max); | |
703 | MB_CHECK_ASSERT(buddy); | |
704 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
705 | MB_CHECK_ASSERT(buddy2); | |
706 | MB_CHECK_ASSERT(buddy != buddy2); | |
707 | MB_CHECK_ASSERT(max * 2 == max2); | |
708 | ||
709 | count = 0; | |
710 | for (i = 0; i < max; i++) { | |
711 | ||
712 | if (mb_test_bit(i, buddy)) { | |
af2b3275 | 713 | /* only single bit in buddy2 may be 0 */ |
c9de560d AT |
714 | if (!mb_test_bit(i << 1, buddy2)) { |
715 | MB_CHECK_ASSERT( | |
716 | mb_test_bit((i<<1)+1, buddy2)); | |
c9de560d AT |
717 | } |
718 | continue; | |
719 | } | |
720 | ||
0a10da73 | 721 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
722 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
723 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
724 | ||
725 | for (j = 0; j < (1 << order); j++) { | |
726 | k = (i * (1 << order)) + j; | |
727 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 728 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
729 | } |
730 | count++; | |
731 | } | |
732 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
733 | order--; | |
734 | } | |
735 | ||
736 | fstart = -1; | |
737 | buddy = mb_find_buddy(e4b, 0, &max); | |
738 | for (i = 0; i < max; i++) { | |
739 | if (!mb_test_bit(i, buddy)) { | |
740 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
741 | if (fstart == -1) { | |
742 | fragments++; | |
743 | fstart = i; | |
744 | } | |
745 | continue; | |
746 | } | |
747 | fstart = -1; | |
748 | /* check used bits only */ | |
749 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
750 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
751 | k = i >> j; | |
752 | MB_CHECK_ASSERT(k < max2); | |
753 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
754 | } | |
755 | } | |
756 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
757 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
758 | ||
759 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
5354b2af | 760 | if (!grp) |
133de5a0 | 761 | return; |
c9de560d AT |
762 | list_for_each(cur, &grp->bb_prealloc_list) { |
763 | ext4_group_t groupnr; | |
764 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
765 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
766 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 767 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 768 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
769 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
770 | } | |
c9de560d AT |
771 | } |
772 | #undef MB_CHECK_ASSERT | |
773 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 774 | __FILE__, __func__, __LINE__) |
c9de560d AT |
775 | #else |
776 | #define mb_check_buddy(e4b) | |
777 | #endif | |
778 | ||
7c786059 CL |
779 | /* |
780 | * Divide blocks started from @first with length @len into | |
781 | * smaller chunks with power of 2 blocks. | |
782 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
783 | * then increase bb_counters[] for corresponded chunk size. | |
784 | */ | |
c9de560d | 785 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 786 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
787 | struct ext4_group_info *grp) |
788 | { | |
789 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
790 | ext4_grpblk_t min; |
791 | ext4_grpblk_t max; | |
792 | ext4_grpblk_t chunk; | |
69e43e8c | 793 | unsigned int border; |
c9de560d | 794 | |
7137d7a4 | 795 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
796 | |
797 | border = 2 << sb->s_blocksize_bits; | |
798 | ||
799 | while (len > 0) { | |
800 | /* find how many blocks can be covered since this position */ | |
801 | max = ffs(first | border) - 1; | |
802 | ||
803 | /* find how many blocks of power 2 we need to mark */ | |
804 | min = fls(len) - 1; | |
805 | ||
806 | if (max < min) | |
807 | min = max; | |
808 | chunk = 1 << min; | |
809 | ||
810 | /* mark multiblock chunks only */ | |
811 | grp->bb_counters[min]++; | |
812 | if (min > 0) | |
813 | mb_clear_bit(first >> min, | |
814 | buddy + sbi->s_mb_offsets[min]); | |
815 | ||
816 | len -= chunk; | |
817 | first += chunk; | |
818 | } | |
819 | } | |
820 | ||
83e80a6e | 821 | static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) |
196e402a | 822 | { |
83e80a6e | 823 | int order; |
196e402a | 824 | |
83e80a6e JK |
825 | /* |
826 | * We don't bother with a special lists groups with only 1 block free | |
827 | * extents and for completely empty groups. | |
828 | */ | |
829 | order = fls(len) - 2; | |
830 | if (order < 0) | |
831 | return 0; | |
832 | if (order == MB_NUM_ORDERS(sb)) | |
833 | order--; | |
13df4d44 BL |
834 | if (WARN_ON_ONCE(order > MB_NUM_ORDERS(sb))) |
835 | order = MB_NUM_ORDERS(sb) - 1; | |
83e80a6e | 836 | return order; |
196e402a HS |
837 | } |
838 | ||
83e80a6e | 839 | /* Move group to appropriate avg_fragment_size list */ |
196e402a HS |
840 | static void |
841 | mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) | |
842 | { | |
843 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
83e80a6e | 844 | int new_order; |
196e402a | 845 | |
993bf0f4 | 846 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_fragments == 0) |
196e402a HS |
847 | return; |
848 | ||
83e80a6e JK |
849 | new_order = mb_avg_fragment_size_order(sb, |
850 | grp->bb_free / grp->bb_fragments); | |
851 | if (new_order == grp->bb_avg_fragment_size_order) | |
852 | return; | |
196e402a | 853 | |
83e80a6e JK |
854 | if (grp->bb_avg_fragment_size_order != -1) { |
855 | write_lock(&sbi->s_mb_avg_fragment_size_locks[ | |
856 | grp->bb_avg_fragment_size_order]); | |
857 | list_del(&grp->bb_avg_fragment_size_node); | |
858 | write_unlock(&sbi->s_mb_avg_fragment_size_locks[ | |
859 | grp->bb_avg_fragment_size_order]); | |
860 | } | |
861 | grp->bb_avg_fragment_size_order = new_order; | |
862 | write_lock(&sbi->s_mb_avg_fragment_size_locks[ | |
863 | grp->bb_avg_fragment_size_order]); | |
864 | list_add_tail(&grp->bb_avg_fragment_size_node, | |
865 | &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]); | |
866 | write_unlock(&sbi->s_mb_avg_fragment_size_locks[ | |
867 | grp->bb_avg_fragment_size_order]); | |
196e402a HS |
868 | } |
869 | ||
870 | /* | |
871 | * Choose next group by traversing largest_free_order lists. Updates *new_cr if | |
872 | * cr level needs an update. | |
873 | */ | |
f52f3d2b | 874 | static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac, |
438a35e7 | 875 | enum criteria *new_cr, ext4_group_t *group) |
196e402a HS |
876 | { |
877 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
919eb90c | 878 | struct ext4_group_info *iter; |
196e402a HS |
879 | int i; |
880 | ||
881 | if (ac->ac_status == AC_STATUS_FOUND) | |
882 | return; | |
883 | ||
f52f3d2b OM |
884 | if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED)) |
885 | atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions); | |
196e402a | 886 | |
196e402a HS |
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 | } | |
196e402a HS |
895 | list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], |
896 | bb_largest_free_order_node) { | |
897 | if (sbi->s_mb_stats) | |
f52f3d2b OM |
898 | atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]); |
899 | if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) { | |
919eb90c KS |
900 | *group = iter->bb_group; |
901 | ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED; | |
902 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
903 | return; | |
196e402a HS |
904 | } |
905 | } | |
906 | read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); | |
196e402a HS |
907 | } |
908 | ||
919eb90c KS |
909 | /* Increment cr and search again if no group is found */ |
910 | *new_cr = CR_GOAL_LEN_FAST; | |
196e402a HS |
911 | } |
912 | ||
856d865c OM |
913 | /* |
914 | * Find a suitable group of given order from the average fragments list. | |
915 | */ | |
916 | static struct ext4_group_info * | |
917 | ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order) | |
918 | { | |
919 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
920 | struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order]; | |
921 | rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order]; | |
922 | struct ext4_group_info *grp = NULL, *iter; | |
923 | enum criteria cr = ac->ac_criteria; | |
924 | ||
925 | if (list_empty(frag_list)) | |
926 | return NULL; | |
927 | read_lock(frag_list_lock); | |
928 | if (list_empty(frag_list)) { | |
929 | read_unlock(frag_list_lock); | |
930 | return NULL; | |
931 | } | |
932 | list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) { | |
933 | if (sbi->s_mb_stats) | |
934 | atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]); | |
935 | if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) { | |
936 | grp = iter; | |
937 | break; | |
938 | } | |
939 | } | |
940 | read_unlock(frag_list_lock); | |
941 | return grp; | |
942 | } | |
943 | ||
196e402a | 944 | /* |
83e80a6e JK |
945 | * Choose next group by traversing average fragment size list of suitable |
946 | * order. Updates *new_cr if cr level needs an update. | |
196e402a | 947 | */ |
f52f3d2b | 948 | static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac, |
438a35e7 | 949 | enum criteria *new_cr, ext4_group_t *group) |
196e402a HS |
950 | { |
951 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
856d865c | 952 | struct ext4_group_info *grp = NULL; |
83e80a6e | 953 | int i; |
196e402a | 954 | |
f52f3d2b | 955 | if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) { |
196e402a | 956 | if (sbi->s_mb_stats) |
f52f3d2b | 957 | atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions); |
83e80a6e JK |
958 | } |
959 | ||
960 | for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); | |
961 | i < MB_NUM_ORDERS(ac->ac_sb); i++) { | |
856d865c | 962 | grp = ext4_mb_find_good_group_avg_frag_lists(ac, i); |
b50675a4 KS |
963 | if (grp) { |
964 | *group = grp->bb_group; | |
965 | ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED; | |
966 | return; | |
967 | } | |
196e402a HS |
968 | } |
969 | ||
772c9f69 RH |
970 | /* |
971 | * CR_BEST_AVAIL_LEN works based on the concept that we have | |
972 | * a larger normalized goal len request which can be trimmed to | |
973 | * a smaller goal len such that it can still satisfy original | |
974 | * request len. However, allocation request for non-regular | |
975 | * files never gets normalized. | |
976 | * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA). | |
977 | */ | |
978 | if (ac->ac_flags & EXT4_MB_HINT_DATA) | |
979 | *new_cr = CR_BEST_AVAIL_LEN; | |
980 | else | |
981 | *new_cr = CR_GOAL_LEN_SLOW; | |
7e170922 OM |
982 | } |
983 | ||
984 | /* | |
f52f3d2b | 985 | * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment |
7e170922 OM |
986 | * order we have and proactively trim the goal request length to that order to |
987 | * find a suitable group faster. | |
988 | * | |
989 | * This optimizes allocation speed at the cost of slightly reduced | |
990 | * preallocations. However, we make sure that we don't trim the request too | |
f52f3d2b | 991 | * much and fall to CR_GOAL_LEN_SLOW in that case. |
7e170922 | 992 | */ |
f52f3d2b | 993 | static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac, |
438a35e7 | 994 | enum criteria *new_cr, ext4_group_t *group) |
7e170922 OM |
995 | { |
996 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
997 | struct ext4_group_info *grp = NULL; | |
998 | int i, order, min_order; | |
999 | unsigned long num_stripe_clusters = 0; | |
1000 | ||
f52f3d2b | 1001 | if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) { |
7e170922 | 1002 | if (sbi->s_mb_stats) |
f52f3d2b | 1003 | atomic_inc(&sbi->s_bal_best_avail_bad_suggestions); |
7e170922 OM |
1004 | } |
1005 | ||
1006 | /* | |
1007 | * mb_avg_fragment_size_order() returns order in a way that makes | |
1008 | * retrieving back the length using (1 << order) inaccurate. Hence, use | |
1009 | * fls() instead since we need to know the actual length while modifying | |
1010 | * goal length. | |
1011 | */ | |
5d5460fa | 1012 | order = fls(ac->ac_g_ex.fe_len) - 1; |
13df4d44 BL |
1013 | if (WARN_ON_ONCE(order - 1 > MB_NUM_ORDERS(ac->ac_sb))) |
1014 | order = MB_NUM_ORDERS(ac->ac_sb); | |
f52f3d2b | 1015 | min_order = order - sbi->s_mb_best_avail_max_trim_order; |
7e170922 OM |
1016 | if (min_order < 0) |
1017 | min_order = 0; | |
1018 | ||
7e170922 OM |
1019 | if (sbi->s_stripe > 0) { |
1020 | /* | |
1021 | * We are assuming that stripe size is always a multiple of | |
1022 | * cluster ratio otherwise __ext4_fill_super exists early. | |
1023 | */ | |
1024 | num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe); | |
1025 | if (1 << min_order < num_stripe_clusters) | |
5d5460fa OM |
1026 | /* |
1027 | * We consider 1 order less because later we round | |
1028 | * up the goal len to num_stripe_clusters | |
1029 | */ | |
1030 | min_order = fls(num_stripe_clusters) - 1; | |
7e170922 OM |
1031 | } |
1032 | ||
5d5460fa OM |
1033 | if (1 << min_order < ac->ac_o_ex.fe_len) |
1034 | min_order = fls(ac->ac_o_ex.fe_len); | |
1035 | ||
7e170922 OM |
1036 | for (i = order; i >= min_order; i--) { |
1037 | int frag_order; | |
1038 | /* | |
1039 | * Scale down goal len to make sure we find something | |
1040 | * in the free fragments list. Basically, reduce | |
1041 | * preallocations. | |
1042 | */ | |
1043 | ac->ac_g_ex.fe_len = 1 << i; | |
1044 | ||
1045 | if (num_stripe_clusters > 0) { | |
1046 | /* | |
4c0cfebd TT |
1047 | * Try to round up the adjusted goal length to |
1048 | * stripe size (in cluster units) multiple for | |
1049 | * efficiency. | |
7e170922 OM |
1050 | */ |
1051 | ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len, | |
1052 | num_stripe_clusters); | |
1053 | } | |
1054 | ||
1055 | frag_order = mb_avg_fragment_size_order(ac->ac_sb, | |
1056 | ac->ac_g_ex.fe_len); | |
1057 | ||
1058 | grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order); | |
bcb123ac KS |
1059 | if (grp) { |
1060 | *group = grp->bb_group; | |
1061 | ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED; | |
1062 | return; | |
1063 | } | |
7e170922 OM |
1064 | } |
1065 | ||
bcb123ac KS |
1066 | /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */ |
1067 | ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; | |
1068 | *new_cr = CR_GOAL_LEN_SLOW; | |
196e402a HS |
1069 | } |
1070 | ||
1071 | static inline int should_optimize_scan(struct ext4_allocation_context *ac) | |
1072 | { | |
1073 | if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) | |
1074 | return 0; | |
f52f3d2b | 1075 | if (ac->ac_criteria >= CR_GOAL_LEN_SLOW) |
196e402a | 1076 | return 0; |
077d0c2c | 1077 | if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) |
196e402a HS |
1078 | return 0; |
1079 | return 1; | |
1080 | } | |
1081 | ||
1082 | /* | |
da5704ee | 1083 | * Return next linear group for allocation. |
196e402a | 1084 | */ |
60c672b7 | 1085 | static ext4_group_t |
da5704ee | 1086 | next_linear_group(ext4_group_t group, ext4_group_t ngroups) |
196e402a | 1087 | { |
196e402a HS |
1088 | /* |
1089 | * Artificially restricted ngroups for non-extent | |
1090 | * files makes group > ngroups possible on first loop. | |
1091 | */ | |
1092 | return group + 1 >= ngroups ? 0 : group + 1; | |
1093 | } | |
1094 | ||
1095 | /* | |
1096 | * ext4_mb_choose_next_group: choose next group for allocation. | |
1097 | * | |
1098 | * @ac Allocation Context | |
1099 | * @new_cr This is an output parameter. If the there is no good group | |
1100 | * available at current CR level, this field is updated to indicate | |
1101 | * the new cr level that should be used. | |
1102 | * @group This is an input / output parameter. As an input it indicates the | |
1103 | * next group that the allocator intends to use for allocation. As | |
1104 | * output, this field indicates the next group that should be used as | |
1105 | * determined by the optimization functions. | |
1106 | * @ngroups Total number of groups | |
1107 | */ | |
1108 | static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, | |
4eb7a4a1 | 1109 | enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) |
196e402a HS |
1110 | { |
1111 | *new_cr = ac->ac_criteria; | |
1112 | ||
da5704ee KS |
1113 | if (!should_optimize_scan(ac)) { |
1114 | *group = next_linear_group(*group, ngroups); | |
1115 | return; | |
1116 | } | |
1117 | ||
1118 | /* | |
1119 | * Optimized scanning can return non adjacent groups which can cause | |
1120 | * seek overhead for rotational disks. So try few linear groups before | |
1121 | * trying optimized scan. | |
1122 | */ | |
1123 | if (ac->ac_groups_linear_remaining) { | |
1124 | *group = next_linear_group(*group, ngroups); | |
1125 | ac->ac_groups_linear_remaining--; | |
196e402a | 1126 | return; |
4fca50d4 | 1127 | } |
196e402a | 1128 | |
f52f3d2b | 1129 | if (*new_cr == CR_POWER2_ALIGNED) { |
438a35e7 | 1130 | ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group); |
f52f3d2b | 1131 | } else if (*new_cr == CR_GOAL_LEN_FAST) { |
438a35e7 | 1132 | ext4_mb_choose_next_group_goal_fast(ac, new_cr, group); |
f52f3d2b | 1133 | } else if (*new_cr == CR_BEST_AVAIL_LEN) { |
438a35e7 | 1134 | ext4_mb_choose_next_group_best_avail(ac, new_cr, group); |
196e402a HS |
1135 | } else { |
1136 | /* | |
2caffb6a KS |
1137 | * TODO: For CR_GOAL_LEN_SLOW, we can arrange groups in an |
1138 | * rb tree sorted by bb_free. But until that happens, we should | |
1139 | * never come here. | |
196e402a HS |
1140 | */ |
1141 | WARN_ON(1); | |
1142 | } | |
1143 | } | |
1144 | ||
8a57d9d6 CW |
1145 | /* |
1146 | * Cache the order of the largest free extent we have available in this block | |
1147 | * group. | |
1148 | */ | |
1149 | static void | |
1150 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
1151 | { | |
196e402a | 1152 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
8a57d9d6 | 1153 | int i; |
8a57d9d6 | 1154 | |
1940265e JK |
1155 | for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) |
1156 | if (grp->bb_counters[i] > 0) | |
1157 | break; | |
1158 | /* No need to move between order lists? */ | |
1159 | if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || | |
1160 | i == grp->bb_largest_free_order) { | |
1161 | grp->bb_largest_free_order = i; | |
1162 | return; | |
1163 | } | |
1164 | ||
1165 | if (grp->bb_largest_free_order >= 0) { | |
196e402a HS |
1166 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1167 | grp->bb_largest_free_order]); | |
1168 | list_del_init(&grp->bb_largest_free_order_node); | |
1169 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1170 | grp->bb_largest_free_order]); | |
1171 | } | |
1940265e JK |
1172 | grp->bb_largest_free_order = i; |
1173 | if (grp->bb_largest_free_order >= 0 && grp->bb_free) { | |
196e402a HS |
1174 | write_lock(&sbi->s_mb_largest_free_orders_locks[ |
1175 | grp->bb_largest_free_order]); | |
1176 | list_add_tail(&grp->bb_largest_free_order_node, | |
1177 | &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); | |
1178 | write_unlock(&sbi->s_mb_largest_free_orders_locks[ | |
1179 | grp->bb_largest_free_order]); | |
1180 | } | |
8a57d9d6 CW |
1181 | } |
1182 | ||
089ceecc ES |
1183 | static noinline_for_stack |
1184 | void ext4_mb_generate_buddy(struct super_block *sb, | |
5354b2af TT |
1185 | void *buddy, void *bitmap, ext4_group_t group, |
1186 | struct ext4_group_info *grp) | |
c9de560d | 1187 | { |
e43bb4e6 | 1188 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 1189 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
1190 | ext4_grpblk_t i = 0; |
1191 | ext4_grpblk_t first; | |
1192 | ext4_grpblk_t len; | |
c9de560d AT |
1193 | unsigned free = 0; |
1194 | unsigned fragments = 0; | |
1195 | unsigned long long period = get_cycles(); | |
1196 | ||
1197 | /* initialize buddy from bitmap which is aggregation | |
1198 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 1199 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
1200 | grp->bb_first_free = i; |
1201 | while (i < max) { | |
1202 | fragments++; | |
1203 | first = i; | |
ffad0a44 | 1204 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
1205 | len = i - first; |
1206 | free += len; | |
1207 | if (len > 1) | |
1208 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
1209 | else | |
1210 | grp->bb_counters[0]++; | |
1211 | if (i < max) | |
ffad0a44 | 1212 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
1213 | } |
1214 | grp->bb_fragments = fragments; | |
1215 | ||
1216 | if (free != grp->bb_free) { | |
e29136f8 | 1217 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
1218 | "block bitmap and bg descriptor " |
1219 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 1220 | free, grp->bb_free); |
e56eb659 | 1221 | /* |
163a203d | 1222 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
1223 | * corrupt and update bb_free using bitmap value |
1224 | */ | |
c9de560d | 1225 | grp->bb_free = free; |
db79e6d1 WS |
1226 | ext4_mark_group_bitmap_corrupted(sb, group, |
1227 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
c9de560d | 1228 | } |
8a57d9d6 | 1229 | mb_set_largest_free_order(sb, grp); |
83e80a6e | 1230 | mb_update_avg_fragment_size(sb, grp); |
c9de560d AT |
1231 | |
1232 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
1233 | ||
1234 | period = get_cycles() - period; | |
67d25186 HS |
1235 | atomic_inc(&sbi->s_mb_buddies_generated); |
1236 | atomic64_add(period, &sbi->s_mb_generation_time); | |
c9de560d AT |
1237 | } |
1238 | ||
c9b528c3 BL |
1239 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
1240 | { | |
1241 | int count; | |
1242 | int order = 1; | |
1243 | void *buddy; | |
1244 | ||
1245 | while ((buddy = mb_find_buddy(e4b, order++, &count))) | |
1246 | mb_set_bits(buddy, 0, count); | |
1247 | ||
1248 | e4b->bd_info->bb_fragments = 0; | |
1249 | memset(e4b->bd_info->bb_counters, 0, | |
1250 | sizeof(*e4b->bd_info->bb_counters) * | |
1251 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
1252 | ||
1253 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
1254 | e4b->bd_bitmap, e4b->bd_group, e4b->bd_info); | |
1255 | } | |
1256 | ||
c9de560d AT |
1257 | /* The buddy information is attached the buddy cache inode |
1258 | * for convenience. The information regarding each group | |
1259 | * is loaded via ext4_mb_load_buddy. The information involve | |
1260 | * block bitmap and buddy information. The information are | |
1261 | * stored in the inode as | |
1262 | * | |
1263 | * { page } | |
c3a326a6 | 1264 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
1265 | * |
1266 | * | |
1267 | * one block each for bitmap and buddy information. | |
1268 | * So for each group we take up 2 blocks. A page can | |
ea1754a0 | 1269 | * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. |
c9de560d AT |
1270 | * So it can have information regarding groups_per_page which |
1271 | * is blocks_per_page/2 | |
8a57d9d6 CW |
1272 | * |
1273 | * Locking note: This routine takes the block group lock of all groups | |
1274 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
1275 | */ |
1276 | ||
e1622a0d | 1277 | static int ext4_mb_init_cache(struct folio *folio, char *incore, gfp_t gfp) |
c9de560d | 1278 | { |
8df9675f | 1279 | ext4_group_t ngroups; |
89cadf6e | 1280 | unsigned int blocksize; |
c9de560d AT |
1281 | int blocks_per_page; |
1282 | int groups_per_page; | |
1283 | int err = 0; | |
1284 | int i; | |
813e5727 | 1285 | ext4_group_t first_group, group; |
c9de560d AT |
1286 | int first_block; |
1287 | struct super_block *sb; | |
1288 | struct buffer_head *bhs; | |
fa77dcfa | 1289 | struct buffer_head **bh = NULL; |
c9de560d AT |
1290 | struct inode *inode; |
1291 | char *data; | |
1292 | char *bitmap; | |
9b8b7d35 | 1293 | struct ext4_group_info *grinfo; |
c9de560d | 1294 | |
e1622a0d | 1295 | inode = folio->mapping->host; |
c9de560d | 1296 | sb = inode->i_sb; |
8df9675f | 1297 | ngroups = ext4_get_groups_count(sb); |
93407472 | 1298 | blocksize = i_blocksize(inode); |
09cbfeaf | 1299 | blocks_per_page = PAGE_SIZE / blocksize; |
c9de560d | 1300 | |
e1622a0d | 1301 | mb_debug(sb, "init folio %lu\n", folio->index); |
d3df1453 | 1302 | |
c9de560d AT |
1303 | groups_per_page = blocks_per_page >> 1; |
1304 | if (groups_per_page == 0) | |
1305 | groups_per_page = 1; | |
1306 | ||
1307 | /* allocate buffer_heads to read bitmaps */ | |
1308 | if (groups_per_page > 1) { | |
c9de560d | 1309 | i = sizeof(struct buffer_head *) * groups_per_page; |
adb7ef60 | 1310 | bh = kzalloc(i, gfp); |
139f46d3 KS |
1311 | if (bh == NULL) |
1312 | return -ENOMEM; | |
c9de560d AT |
1313 | } else |
1314 | bh = &bhs; | |
1315 | ||
e1622a0d | 1316 | first_group = folio->index * blocks_per_page / 2; |
c9de560d | 1317 | |
e1622a0d | 1318 | /* read all groups the folio covers into the cache */ |
813e5727 TT |
1319 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
1320 | if (group >= ngroups) | |
c9de560d AT |
1321 | break; |
1322 | ||
813e5727 | 1323 | grinfo = ext4_get_group_info(sb, group); |
5354b2af TT |
1324 | if (!grinfo) |
1325 | continue; | |
9b8b7d35 AG |
1326 | /* |
1327 | * If page is uptodate then we came here after online resize | |
1328 | * which added some new uninitialized group info structs, so | |
e1622a0d | 1329 | * we must skip all initialized uptodate buddies on the folio, |
9b8b7d35 AG |
1330 | * which may be currently in use by an allocating task. |
1331 | */ | |
e1622a0d MWO |
1332 | if (folio_test_uptodate(folio) && |
1333 | !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
9b8b7d35 AG |
1334 | bh[i] = NULL; |
1335 | continue; | |
1336 | } | |
cfd73237 | 1337 | bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
9008a58e DW |
1338 | if (IS_ERR(bh[i])) { |
1339 | err = PTR_ERR(bh[i]); | |
1340 | bh[i] = NULL; | |
c9de560d | 1341 | goto out; |
2ccb5fb9 | 1342 | } |
d3df1453 | 1343 | mb_debug(sb, "read bitmap for group %u\n", group); |
c9de560d AT |
1344 | } |
1345 | ||
1346 | /* wait for I/O completion */ | |
813e5727 | 1347 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
9008a58e DW |
1348 | int err2; |
1349 | ||
1350 | if (!bh[i]) | |
1351 | continue; | |
1352 | err2 = ext4_wait_block_bitmap(sb, group, bh[i]); | |
1353 | if (!err) | |
1354 | err = err2; | |
813e5727 | 1355 | } |
c9de560d | 1356 | |
e1622a0d | 1357 | first_block = folio->index * blocks_per_page; |
c9de560d | 1358 | for (i = 0; i < blocks_per_page; i++) { |
c9de560d | 1359 | group = (first_block + i) >> 1; |
8df9675f | 1360 | if (group >= ngroups) |
c9de560d AT |
1361 | break; |
1362 | ||
9b8b7d35 AG |
1363 | if (!bh[group - first_group]) |
1364 | /* skip initialized uptodate buddy */ | |
1365 | continue; | |
1366 | ||
bbdc322f LC |
1367 | if (!buffer_verified(bh[group - first_group])) |
1368 | /* Skip faulty bitmaps */ | |
1369 | continue; | |
1370 | err = 0; | |
1371 | ||
c9de560d AT |
1372 | /* |
1373 | * data carry information regarding this | |
1374 | * particular group in the format specified | |
1375 | * above | |
1376 | * | |
1377 | */ | |
e1622a0d | 1378 | data = folio_address(folio) + (i * blocksize); |
c9de560d AT |
1379 | bitmap = bh[group - first_group]->b_data; |
1380 | ||
1381 | /* | |
1382 | * We place the buddy block and bitmap block | |
1383 | * close together | |
1384 | */ | |
ebf6cb7c WJ |
1385 | grinfo = ext4_get_group_info(sb, group); |
1386 | if (!grinfo) { | |
1387 | err = -EFSCORRUPTED; | |
1388 | goto out; | |
1389 | } | |
c9de560d AT |
1390 | if ((first_block + i) & 1) { |
1391 | /* this is block of buddy */ | |
1392 | BUG_ON(incore == NULL); | |
e1622a0d MWO |
1393 | mb_debug(sb, "put buddy for group %u in folio %lu/%x\n", |
1394 | group, folio->index, i * blocksize); | |
f307333e | 1395 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
1396 | grinfo->bb_fragments = 0; |
1397 | memset(grinfo->bb_counters, 0, | |
1927805e | 1398 | sizeof(*grinfo->bb_counters) * |
4b68f6df | 1399 | (MB_NUM_ORDERS(sb))); |
c9de560d AT |
1400 | /* |
1401 | * incore got set to the group block bitmap below | |
1402 | */ | |
7a2fcbf7 | 1403 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
1404 | /* init the buddy */ |
1405 | memset(data, 0xff, blocksize); | |
5354b2af | 1406 | ext4_mb_generate_buddy(sb, data, incore, group, grinfo); |
7a2fcbf7 | 1407 | ext4_unlock_group(sb, group); |
c9de560d AT |
1408 | incore = NULL; |
1409 | } else { | |
1410 | /* this is block of bitmap */ | |
1411 | BUG_ON(incore != NULL); | |
e1622a0d MWO |
1412 | mb_debug(sb, "put bitmap for group %u in folio %lu/%x\n", |
1413 | group, folio->index, i * blocksize); | |
f307333e | 1414 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
1415 | |
1416 | /* see comments in ext4_mb_put_pa() */ | |
1417 | ext4_lock_group(sb, group); | |
1418 | memcpy(data, bitmap, blocksize); | |
1419 | ||
1420 | /* mark all preallocated blks used in in-core bitmap */ | |
1421 | ext4_mb_generate_from_pa(sb, data, group); | |
ebf6cb7c | 1422 | WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root)); |
c9de560d AT |
1423 | ext4_unlock_group(sb, group); |
1424 | ||
1425 | /* set incore so that the buddy information can be | |
1426 | * generated using this | |
1427 | */ | |
1428 | incore = data; | |
1429 | } | |
1430 | } | |
e1622a0d | 1431 | folio_mark_uptodate(folio); |
c9de560d AT |
1432 | |
1433 | out: | |
1434 | if (bh) { | |
9b8b7d35 | 1435 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
1436 | brelse(bh[i]); |
1437 | if (bh != &bhs) | |
1438 | kfree(bh); | |
1439 | } | |
1440 | return err; | |
1441 | } | |
1442 | ||
eee4adc7 | 1443 | /* |
2de8807b AG |
1444 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
1445 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
1446 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
5eea586b | 1447 | * are on the same page e4b->bd_buddy_folio is NULL and return value is 0. |
eee4adc7 | 1448 | */ |
2de8807b | 1449 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
adb7ef60 | 1450 | ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
eee4adc7 | 1451 | { |
2de8807b AG |
1452 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
1453 | int block, pnum, poff; | |
eee4adc7 | 1454 | int blocks_per_page; |
99b150d8 | 1455 | struct folio *folio; |
2de8807b | 1456 | |
5eea586b | 1457 | e4b->bd_buddy_folio = NULL; |
99b150d8 | 1458 | e4b->bd_bitmap_folio = NULL; |
eee4adc7 | 1459 | |
09cbfeaf | 1460 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
eee4adc7 ES |
1461 | /* |
1462 | * the buddy cache inode stores the block bitmap | |
1463 | * and buddy information in consecutive blocks. | |
1464 | * So for each group we need two blocks. | |
1465 | */ | |
1466 | block = group * 2; | |
1467 | pnum = block / blocks_per_page; | |
2de8807b | 1468 | poff = block % blocks_per_page; |
99b150d8 MWO |
1469 | folio = __filemap_get_folio(inode->i_mapping, pnum, |
1470 | FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); | |
1471 | if (IS_ERR(folio)) | |
1472 | return PTR_ERR(folio); | |
1473 | BUG_ON(folio->mapping != inode->i_mapping); | |
1474 | e4b->bd_bitmap_folio = folio; | |
1475 | e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize); | |
2de8807b AG |
1476 | |
1477 | if (blocks_per_page >= 2) { | |
1478 | /* buddy and bitmap are on the same page */ | |
1479 | return 0; | |
eee4adc7 | 1480 | } |
2de8807b | 1481 | |
f2fec3e9 | 1482 | /* blocks_per_page == 1, hence we need another page for the buddy */ |
5eea586b MWO |
1483 | folio = __filemap_get_folio(inode->i_mapping, block + 1, |
1484 | FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); | |
1485 | if (IS_ERR(folio)) | |
1486 | return PTR_ERR(folio); | |
1487 | BUG_ON(folio->mapping != inode->i_mapping); | |
1488 | e4b->bd_buddy_folio = folio; | |
2de8807b | 1489 | return 0; |
eee4adc7 ES |
1490 | } |
1491 | ||
2de8807b | 1492 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1493 | { |
99b150d8 MWO |
1494 | if (e4b->bd_bitmap_folio) { |
1495 | folio_unlock(e4b->bd_bitmap_folio); | |
1496 | folio_put(e4b->bd_bitmap_folio); | |
2de8807b | 1497 | } |
5eea586b MWO |
1498 | if (e4b->bd_buddy_folio) { |
1499 | folio_unlock(e4b->bd_buddy_folio); | |
1500 | folio_put(e4b->bd_buddy_folio); | |
eee4adc7 | 1501 | } |
eee4adc7 ES |
1502 | } |
1503 | ||
8a57d9d6 CW |
1504 | /* |
1505 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1506 | * block group lock of all groups for this page; do not hold the BG lock when | |
1507 | * calling this routine! | |
1508 | */ | |
b6a758ec | 1509 | static noinline_for_stack |
adb7ef60 | 1510 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
b6a758ec AK |
1511 | { |
1512 | ||
b6a758ec | 1513 | struct ext4_group_info *this_grp; |
2de8807b | 1514 | struct ext4_buddy e4b; |
99b150d8 | 1515 | struct folio *folio; |
2de8807b | 1516 | int ret = 0; |
b6a758ec | 1517 | |
b10a44c3 | 1518 | might_sleep(); |
d3df1453 | 1519 | mb_debug(sb, "init group %u\n", group); |
b6a758ec | 1520 | this_grp = ext4_get_group_info(sb, group); |
5354b2af TT |
1521 | if (!this_grp) |
1522 | return -EFSCORRUPTED; | |
1523 | ||
b6a758ec | 1524 | /* |
08c3a813 AK |
1525 | * This ensures that we don't reinit the buddy cache |
1526 | * page which map to the group from which we are already | |
1527 | * allocating. If we are looking at the buddy cache we would | |
1528 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1529 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1530 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1531 | * page accessed. | |
b6a758ec | 1532 | */ |
adb7ef60 | 1533 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); |
2de8807b | 1534 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
b6a758ec AK |
1535 | /* |
1536 | * somebody initialized the group | |
1537 | * return without doing anything | |
1538 | */ | |
b6a758ec AK |
1539 | goto err; |
1540 | } | |
2de8807b | 1541 | |
99b150d8 | 1542 | folio = e4b.bd_bitmap_folio; |
e1622a0d | 1543 | ret = ext4_mb_init_cache(folio, NULL, gfp); |
2de8807b AG |
1544 | if (ret) |
1545 | goto err; | |
99b150d8 | 1546 | if (!folio_test_uptodate(folio)) { |
b6a758ec AK |
1547 | ret = -EIO; |
1548 | goto err; | |
1549 | } | |
b6a758ec | 1550 | |
5eea586b | 1551 | if (e4b.bd_buddy_folio == NULL) { |
b6a758ec AK |
1552 | /* |
1553 | * If both the bitmap and buddy are in | |
1554 | * the same page we don't need to force | |
1555 | * init the buddy | |
1556 | */ | |
2de8807b AG |
1557 | ret = 0; |
1558 | goto err; | |
b6a758ec | 1559 | } |
2de8807b | 1560 | /* init buddy cache */ |
5eea586b | 1561 | folio = e4b.bd_buddy_folio; |
e1622a0d | 1562 | ret = ext4_mb_init_cache(folio, e4b.bd_bitmap, gfp); |
2de8807b AG |
1563 | if (ret) |
1564 | goto err; | |
5eea586b | 1565 | if (!folio_test_uptodate(folio)) { |
b6a758ec AK |
1566 | ret = -EIO; |
1567 | goto err; | |
1568 | } | |
b6a758ec | 1569 | err: |
2de8807b | 1570 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1571 | return ret; |
1572 | } | |
1573 | ||
8a57d9d6 CW |
1574 | /* |
1575 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1576 | * block group lock of all groups for this page; do not hold the BG lock when | |
1577 | * calling this routine! | |
1578 | */ | |
4ddfef7b | 1579 | static noinline_for_stack int |
adb7ef60 KK |
1580 | ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
1581 | struct ext4_buddy *e4b, gfp_t gfp) | |
c9de560d | 1582 | { |
c9de560d AT |
1583 | int blocks_per_page; |
1584 | int block; | |
1585 | int pnum; | |
1586 | int poff; | |
99b150d8 | 1587 | struct folio *folio; |
fdf6c7a7 | 1588 | int ret; |
920313a7 AK |
1589 | struct ext4_group_info *grp; |
1590 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1591 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1592 | |
b10a44c3 | 1593 | might_sleep(); |
d3df1453 | 1594 | mb_debug(sb, "load group %u\n", group); |
c9de560d | 1595 | |
09cbfeaf | 1596 | blocks_per_page = PAGE_SIZE / sb->s_blocksize; |
920313a7 | 1597 | grp = ext4_get_group_info(sb, group); |
5354b2af TT |
1598 | if (!grp) |
1599 | return -EFSCORRUPTED; | |
c9de560d AT |
1600 | |
1601 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1602 | e4b->bd_info = grp; |
c9de560d AT |
1603 | e4b->bd_sb = sb; |
1604 | e4b->bd_group = group; | |
5eea586b | 1605 | e4b->bd_buddy_folio = NULL; |
99b150d8 | 1606 | e4b->bd_bitmap_folio = NULL; |
c9de560d | 1607 | |
f41c0750 | 1608 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1609 | /* |
1610 | * we need full data about the group | |
1611 | * to make a good selection | |
1612 | */ | |
adb7ef60 | 1613 | ret = ext4_mb_init_group(sb, group, gfp); |
f41c0750 AK |
1614 | if (ret) |
1615 | return ret; | |
f41c0750 AK |
1616 | } |
1617 | ||
c9de560d AT |
1618 | /* |
1619 | * the buddy cache inode stores the block bitmap | |
1620 | * and buddy information in consecutive blocks. | |
1621 | * So for each group we need two blocks. | |
1622 | */ | |
1623 | block = group * 2; | |
1624 | pnum = block / blocks_per_page; | |
1625 | poff = block % blocks_per_page; | |
1626 | ||
99b150d8 MWO |
1627 | /* Avoid locking the folio in the fast path ... */ |
1628 | folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0); | |
1629 | if (IS_ERR(folio) || !folio_test_uptodate(folio)) { | |
1630 | if (!IS_ERR(folio)) | |
920313a7 | 1631 | /* |
99b150d8 MWO |
1632 | * drop the folio reference and try |
1633 | * to get the folio with lock. If we | |
920313a7 | 1634 | * are not uptodate that implies |
99b150d8 MWO |
1635 | * somebody just created the folio but |
1636 | * is yet to initialize it. So | |
920313a7 AK |
1637 | * wait for it to initialize. |
1638 | */ | |
99b150d8 MWO |
1639 | folio_put(folio); |
1640 | folio = __filemap_get_folio(inode->i_mapping, pnum, | |
1641 | FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); | |
1642 | if (!IS_ERR(folio)) { | |
1643 | if (WARN_RATELIMIT(folio->mapping != inode->i_mapping, | |
1644 | "ext4: bitmap's mapping != inode->i_mapping\n")) { | |
19b8b035 | 1645 | /* should never happen */ |
99b150d8 | 1646 | folio_unlock(folio); |
19b8b035 TT |
1647 | ret = -EINVAL; |
1648 | goto err; | |
1649 | } | |
99b150d8 | 1650 | if (!folio_test_uptodate(folio)) { |
e1622a0d | 1651 | ret = ext4_mb_init_cache(folio, NULL, gfp); |
fdf6c7a7 | 1652 | if (ret) { |
99b150d8 | 1653 | folio_unlock(folio); |
fdf6c7a7 SF |
1654 | goto err; |
1655 | } | |
99b150d8 | 1656 | mb_cmp_bitmaps(e4b, folio_address(folio) + |
c9de560d AT |
1657 | (poff * sb->s_blocksize)); |
1658 | } | |
99b150d8 | 1659 | folio_unlock(folio); |
c9de560d AT |
1660 | } |
1661 | } | |
99b150d8 MWO |
1662 | if (IS_ERR(folio)) { |
1663 | ret = PTR_ERR(folio); | |
c57ab39b YL |
1664 | goto err; |
1665 | } | |
99b150d8 | 1666 | if (!folio_test_uptodate(folio)) { |
fdf6c7a7 | 1667 | ret = -EIO; |
c9de560d | 1668 | goto err; |
fdf6c7a7 | 1669 | } |
2457aec6 | 1670 | |
5eea586b | 1671 | /* Folios marked accessed already */ |
99b150d8 MWO |
1672 | e4b->bd_bitmap_folio = folio; |
1673 | e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize); | |
c9de560d AT |
1674 | |
1675 | block++; | |
1676 | pnum = block / blocks_per_page; | |
1677 | poff = block % blocks_per_page; | |
1678 | ||
5eea586b MWO |
1679 | folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0); |
1680 | if (IS_ERR(folio) || !folio_test_uptodate(folio)) { | |
1681 | if (!IS_ERR(folio)) | |
1682 | folio_put(folio); | |
1683 | folio = __filemap_get_folio(inode->i_mapping, pnum, | |
1684 | FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); | |
1685 | if (!IS_ERR(folio)) { | |
1686 | if (WARN_RATELIMIT(folio->mapping != inode->i_mapping, | |
1687 | "ext4: buddy bitmap's mapping != inode->i_mapping\n")) { | |
19b8b035 | 1688 | /* should never happen */ |
5eea586b | 1689 | folio_unlock(folio); |
19b8b035 TT |
1690 | ret = -EINVAL; |
1691 | goto err; | |
1692 | } | |
5eea586b | 1693 | if (!folio_test_uptodate(folio)) { |
e1622a0d | 1694 | ret = ext4_mb_init_cache(folio, e4b->bd_bitmap, |
adb7ef60 | 1695 | gfp); |
fdf6c7a7 | 1696 | if (ret) { |
5eea586b | 1697 | folio_unlock(folio); |
fdf6c7a7 SF |
1698 | goto err; |
1699 | } | |
1700 | } | |
5eea586b | 1701 | folio_unlock(folio); |
c9de560d AT |
1702 | } |
1703 | } | |
5eea586b MWO |
1704 | if (IS_ERR(folio)) { |
1705 | ret = PTR_ERR(folio); | |
c57ab39b YL |
1706 | goto err; |
1707 | } | |
5eea586b | 1708 | if (!folio_test_uptodate(folio)) { |
fdf6c7a7 | 1709 | ret = -EIO; |
c9de560d | 1710 | goto err; |
fdf6c7a7 | 1711 | } |
2457aec6 | 1712 | |
5eea586b MWO |
1713 | /* Folios marked accessed already */ |
1714 | e4b->bd_buddy_folio = folio; | |
1715 | e4b->bd_buddy = folio_address(folio) + (poff * sb->s_blocksize); | |
c9de560d | 1716 | |
c9de560d AT |
1717 | return 0; |
1718 | ||
1719 | err: | |
c6a6c969 | 1720 | if (!IS_ERR_OR_NULL(folio)) |
5eea586b | 1721 | folio_put(folio); |
99b150d8 MWO |
1722 | if (e4b->bd_bitmap_folio) |
1723 | folio_put(e4b->bd_bitmap_folio); | |
285164b8 | 1724 | |
c9de560d AT |
1725 | e4b->bd_buddy = NULL; |
1726 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1727 | return ret; |
c9de560d AT |
1728 | } |
1729 | ||
adb7ef60 KK |
1730 | static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
1731 | struct ext4_buddy *e4b) | |
1732 | { | |
1733 | return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); | |
1734 | } | |
1735 | ||
e39e07fd | 1736 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d | 1737 | { |
99b150d8 MWO |
1738 | if (e4b->bd_bitmap_folio) |
1739 | folio_put(e4b->bd_bitmap_folio); | |
5eea586b MWO |
1740 | if (e4b->bd_buddy_folio) |
1741 | folio_put(e4b->bd_buddy_folio); | |
c9de560d AT |
1742 | } |
1743 | ||
1744 | ||
1745 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1746 | { | |
ce3cca33 | 1747 | int order = 1, max; |
c9de560d AT |
1748 | void *bb; |
1749 | ||
c5e8f3f3 | 1750 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1751 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1752 | ||
c9de560d | 1753 | while (order <= e4b->bd_blkbits + 1) { |
ce3cca33 CX |
1754 | bb = mb_find_buddy(e4b, order, &max); |
1755 | if (!mb_test_bit(block >> order, bb)) { | |
c9de560d AT |
1756 | /* this block is part of buddy of order 'order' */ |
1757 | return order; | |
1758 | } | |
c9de560d AT |
1759 | order++; |
1760 | } | |
1761 | return 0; | |
1762 | } | |
1763 | ||
955ce5f5 | 1764 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1765 | { |
1766 | __u32 *addr; | |
1767 | ||
1768 | len = cur + len; | |
1769 | while (cur < len) { | |
1770 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1771 | /* fast path: clear whole word at once */ | |
1772 | addr = bm + (cur >> 3); | |
1773 | *addr = 0; | |
1774 | cur += 32; | |
1775 | continue; | |
1776 | } | |
955ce5f5 | 1777 | mb_clear_bit(cur, bm); |
c9de560d AT |
1778 | cur++; |
1779 | } | |
1780 | } | |
1781 | ||
eabe0444 AS |
1782 | /* clear bits in given range |
1783 | * will return first found zero bit if any, -1 otherwise | |
1784 | */ | |
1785 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1786 | { | |
1787 | __u32 *addr; | |
1788 | int zero_bit = -1; | |
1789 | ||
1790 | len = cur + len; | |
1791 | while (cur < len) { | |
1792 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1793 | /* fast path: clear whole word at once */ | |
1794 | addr = bm + (cur >> 3); | |
1795 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1796 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1797 | *addr = 0; | |
1798 | cur += 32; | |
1799 | continue; | |
1800 | } | |
1801 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1802 | zero_bit = cur; | |
1803 | cur++; | |
1804 | } | |
1805 | ||
1806 | return zero_bit; | |
1807 | } | |
1808 | ||
123e3016 | 1809 | void mb_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1810 | { |
1811 | __u32 *addr; | |
1812 | ||
1813 | len = cur + len; | |
1814 | while (cur < len) { | |
1815 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1816 | /* fast path: set whole word at once */ | |
1817 | addr = bm + (cur >> 3); | |
1818 | *addr = 0xffffffff; | |
1819 | cur += 32; | |
1820 | continue; | |
1821 | } | |
955ce5f5 | 1822 | mb_set_bit(cur, bm); |
c9de560d AT |
1823 | cur++; |
1824 | } | |
1825 | } | |
1826 | ||
eabe0444 AS |
1827 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
1828 | { | |
1829 | if (mb_test_bit(*bit + side, bitmap)) { | |
1830 | mb_clear_bit(*bit, bitmap); | |
1831 | (*bit) -= side; | |
1832 | return 1; | |
1833 | } | |
1834 | else { | |
1835 | (*bit) += side; | |
1836 | mb_set_bit(*bit, bitmap); | |
1837 | return -1; | |
1838 | } | |
1839 | } | |
1840 | ||
1841 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1842 | { | |
1843 | int max; | |
1844 | int order = 1; | |
1845 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1846 | ||
1847 | while (buddy) { | |
1848 | void *buddy2; | |
1849 | ||
1850 | /* Bits in range [first; last] are known to be set since | |
1851 | * corresponding blocks were allocated. Bits in range | |
1852 | * (first; last) will stay set because they form buddies on | |
1853 | * upper layer. We just deal with borders if they don't | |
1854 | * align with upper layer and then go up. | |
1855 | * Releasing entire group is all about clearing | |
1856 | * single bit of highest order buddy. | |
1857 | */ | |
1858 | ||
1859 | /* Example: | |
1860 | * --------------------------------- | |
1861 | * | 1 | 1 | 1 | 1 | | |
1862 | * --------------------------------- | |
1863 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1864 | * --------------------------------- | |
1865 | * 0 1 2 3 4 5 6 7 | |
1866 | * \_____________________/ | |
1867 | * | |
1868 | * Neither [1] nor [6] is aligned to above layer. | |
1869 | * Left neighbour [0] is free, so mark it busy, | |
1870 | * decrease bb_counters and extend range to | |
1871 | * [0; 6] | |
1872 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1873 | * mark [6] free, increase bb_counters and shrink range to | |
1874 | * [0; 5]. | |
1875 | * Then shift range to [0; 2], go up and do the same. | |
1876 | */ | |
1877 | ||
1878 | ||
1879 | if (first & 1) | |
1880 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1881 | if (!(last & 1)) | |
1882 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1883 | if (first > last) | |
1884 | break; | |
1885 | order++; | |
1886 | ||
976620bd KS |
1887 | buddy2 = mb_find_buddy(e4b, order, &max); |
1888 | if (!buddy2) { | |
eabe0444 AS |
1889 | mb_clear_bits(buddy, first, last - first + 1); |
1890 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1891 | break; | |
1892 | } | |
1893 | first >>= 1; | |
1894 | last >>= 1; | |
1895 | buddy = buddy2; | |
1896 | } | |
1897 | } | |
1898 | ||
7e5a8cdd | 1899 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1900 | int first, int count) |
c9de560d | 1901 | { |
eabe0444 AS |
1902 | int left_is_free = 0; |
1903 | int right_is_free = 0; | |
1904 | int block; | |
1905 | int last = first + count - 1; | |
c9de560d AT |
1906 | struct super_block *sb = e4b->bd_sb; |
1907 | ||
c99d1e6e TT |
1908 | if (WARN_ON(count == 0)) |
1909 | return; | |
eabe0444 | 1910 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1911 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1912 | /* Don't bother if the block group is corrupt. */ |
1913 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1914 | return; | |
1915 | ||
c9de560d AT |
1916 | mb_check_buddy(e4b); |
1917 | mb_free_blocks_double(inode, e4b, first, count); | |
1918 | ||
eabe0444 AS |
1919 | /* access memory sequentially: check left neighbour, |
1920 | * clear range and then check right neighbour | |
1921 | */ | |
c9de560d | 1922 | if (first != 0) |
eabe0444 AS |
1923 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1924 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1925 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1926 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1927 | ||
1928 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1929 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1930 | ext4_fsblk_t blocknr; |
1931 | ||
2331fd4a BL |
1932 | /* |
1933 | * Fastcommit replay can free already freed blocks which | |
1934 | * corrupts allocation info. Regenerate it. | |
1935 | */ | |
1936 | if (sbi->s_mount_state & EXT4_FC_REPLAY) { | |
1937 | mb_regenerate_buddy(e4b); | |
1938 | goto check; | |
1939 | } | |
1940 | ||
eabe0444 | 1941 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); |
49598e04 | 1942 | blocknr += EXT4_C2B(sbi, block); |
c5f3a382 BL |
1943 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
1944 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
2331fd4a BL |
1945 | ext4_grp_locked_error(sb, e4b->bd_group, |
1946 | inode ? inode->i_ino : 0, blocknr, | |
1947 | "freeing already freed block (bit %u); block bitmap corrupt.", | |
1948 | block); | |
2331fd4a | 1949 | return; |
eabe0444 AS |
1950 | } |
1951 | ||
2331fd4a BL |
1952 | this_cpu_inc(discard_pa_seq); |
1953 | e4b->bd_info->bb_free += count; | |
1954 | if (first < e4b->bd_info->bb_first_free) | |
1955 | e4b->bd_info->bb_first_free = first; | |
1956 | ||
eabe0444 AS |
1957 | /* let's maintain fragments counter */ |
1958 | if (left_is_free && right_is_free) | |
c9de560d | 1959 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1960 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1961 | e4b->bd_info->bb_fragments++; |
1962 | ||
eabe0444 AS |
1963 | /* buddy[0] == bd_bitmap is a special case, so handle |
1964 | * it right away and let mb_buddy_mark_free stay free of | |
1965 | * zero order checks. | |
1966 | * Check if neighbours are to be coaleasced, | |
1967 | * adjust bitmap bb_counters and borders appropriately. | |
1968 | */ | |
1969 | if (first & 1) { | |
1970 | first += !left_is_free; | |
1971 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1972 | } | |
1973 | if (!(last & 1)) { | |
1974 | last -= !right_is_free; | |
1975 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1976 | } | |
c9de560d | 1977 | |
eabe0444 AS |
1978 | if (first <= last) |
1979 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1980 | |
8a57d9d6 | 1981 | mb_set_largest_free_order(sb, e4b->bd_info); |
196e402a | 1982 | mb_update_avg_fragment_size(sb, e4b->bd_info); |
2331fd4a | 1983 | check: |
c9de560d | 1984 | mb_check_buddy(e4b); |
c9de560d AT |
1985 | } |
1986 | ||
15c006a2 | 1987 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1988 | int needed, struct ext4_free_extent *ex) |
1989 | { | |
2bf5eb2a | 1990 | int max, order, next; |
c9de560d AT |
1991 | void *buddy; |
1992 | ||
bc8e6740 | 1993 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1994 | BUG_ON(ex == NULL); |
1995 | ||
15c006a2 | 1996 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1997 | BUG_ON(buddy == NULL); |
1998 | BUG_ON(block >= max); | |
1999 | if (mb_test_bit(block, buddy)) { | |
2000 | ex->fe_len = 0; | |
2001 | ex->fe_start = 0; | |
2002 | ex->fe_group = 0; | |
2003 | return 0; | |
2004 | } | |
2005 | ||
15c006a2 RD |
2006 | /* find actual order */ |
2007 | order = mb_find_order_for_block(e4b, block); | |
c9de560d | 2008 | |
2bf5eb2a GH |
2009 | ex->fe_len = (1 << order) - (block & ((1 << order) - 1)); |
2010 | ex->fe_start = block; | |
c9de560d AT |
2011 | ex->fe_group = e4b->bd_group; |
2012 | ||
2bf5eb2a | 2013 | block = block >> order; |
c9de560d AT |
2014 | |
2015 | while (needed > ex->fe_len && | |
d8ec0c39 | 2016 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
2017 | |
2018 | if (block + 1 >= max) | |
2019 | break; | |
2020 | ||
2021 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 2022 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
2023 | break; |
2024 | ||
b051d8dc | 2025 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 2026 | |
c9de560d AT |
2027 | block = next >> order; |
2028 | ex->fe_len += 1 << order; | |
2029 | } | |
2030 | ||
31562b95 | 2031 | if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
43c73221 TT |
2032 | /* Should never happen! (but apparently sometimes does?!?) */ |
2033 | WARN_ON(1); | |
cd84bbba SB |
2034 | ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, |
2035 | "corruption or bug in mb_find_extent " | |
2036 | "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", | |
2037 | block, order, needed, ex->fe_group, ex->fe_start, | |
2038 | ex->fe_len, ex->fe_logical); | |
43c73221 TT |
2039 | ex->fe_len = 0; |
2040 | ex->fe_start = 0; | |
2041 | ex->fe_group = 0; | |
2042 | } | |
c9de560d AT |
2043 | return ex->fe_len; |
2044 | } | |
2045 | ||
2046 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
2047 | { | |
2048 | int ord; | |
2049 | int mlen = 0; | |
2050 | int max = 0; | |
c9de560d AT |
2051 | int start = ex->fe_start; |
2052 | int len = ex->fe_len; | |
2053 | unsigned ret = 0; | |
2054 | int len0 = len; | |
2055 | void *buddy; | |
d1a3924e | 2056 | int ord_start, ord_end; |
c9de560d AT |
2057 | |
2058 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
2059 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 2060 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
2061 | mb_check_buddy(e4b); |
2062 | mb_mark_used_double(e4b, start, len); | |
2063 | ||
07b5b8e1 | 2064 | this_cpu_inc(discard_pa_seq); |
c9de560d AT |
2065 | e4b->bd_info->bb_free -= len; |
2066 | if (e4b->bd_info->bb_first_free == start) | |
2067 | e4b->bd_info->bb_first_free += len; | |
2068 | ||
2069 | /* let's maintain fragments counter */ | |
2070 | if (start != 0) | |
c5e8f3f3 | 2071 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 2072 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 2073 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
2074 | if (mlen && max) |
2075 | e4b->bd_info->bb_fragments++; | |
2076 | else if (!mlen && !max) | |
2077 | e4b->bd_info->bb_fragments--; | |
2078 | ||
2079 | /* let's maintain buddy itself */ | |
2080 | while (len) { | |
d1a3924e | 2081 | ord = mb_find_order_for_block(e4b, start); |
c9de560d AT |
2082 | |
2083 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
2084 | /* the whole chunk may be allocated at once! */ | |
2085 | mlen = 1 << ord; | |
d1a3924e | 2086 | buddy = mb_find_buddy(e4b, ord, &max); |
c9de560d AT |
2087 | BUG_ON((start >> ord) >= max); |
2088 | mb_set_bit(start >> ord, buddy); | |
2089 | e4b->bd_info->bb_counters[ord]--; | |
2090 | start += mlen; | |
2091 | len -= mlen; | |
2092 | BUG_ON(len < 0); | |
2093 | continue; | |
2094 | } | |
2095 | ||
2096 | /* store for history */ | |
2097 | if (ret == 0) | |
2098 | ret = len | (ord << 16); | |
2099 | ||
c9de560d AT |
2100 | BUG_ON(ord <= 0); |
2101 | buddy = mb_find_buddy(e4b, ord, &max); | |
2102 | mb_set_bit(start >> ord, buddy); | |
2103 | e4b->bd_info->bb_counters[ord]--; | |
2104 | ||
d1a3924e KS |
2105 | ord_start = (start >> ord) << ord; |
2106 | ord_end = ord_start + (1 << ord); | |
2107 | /* first chunk */ | |
2108 | if (start > ord_start) | |
2109 | ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy, | |
2110 | ord_start, start - ord_start, | |
2111 | e4b->bd_info); | |
2112 | ||
2113 | /* last chunk */ | |
2114 | if (start + len < ord_end) { | |
2115 | ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy, | |
2116 | start + len, | |
2117 | ord_end - (start + len), | |
2118 | e4b->bd_info); | |
2119 | break; | |
2120 | } | |
2121 | len = start + len - ord_end; | |
2122 | start = ord_end; | |
c9de560d | 2123 | } |
8a57d9d6 | 2124 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 2125 | |
196e402a | 2126 | mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); |
123e3016 | 2127 | mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
2128 | mb_check_buddy(e4b); |
2129 | ||
2130 | return ret; | |
2131 | } | |
2132 | ||
2133 | /* | |
2134 | * Must be called under group lock! | |
2135 | */ | |
2136 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
2137 | struct ext4_buddy *e4b) | |
2138 | { | |
2139 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2140 | int ret; | |
2141 | ||
2142 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
2143 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
2144 | ||
2145 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
2146 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
2147 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
2148 | ||
2149 | /* preallocation can change ac_b_ex, thus we store actually | |
2150 | * allocated blocks for history */ | |
2151 | ac->ac_f_ex = ac->ac_b_ex; | |
2152 | ||
2153 | ac->ac_status = AC_STATUS_FOUND; | |
2154 | ac->ac_tail = ret & 0xffff; | |
2155 | ac->ac_buddy = ret >> 16; | |
2156 | ||
c3a326a6 AK |
2157 | /* |
2158 | * take the page reference. We want the page to be pinned | |
2159 | * so that we don't get a ext4_mb_init_cache_call for this | |
2160 | * group until we update the bitmap. That would mean we | |
2161 | * double allocate blocks. The reference is dropped | |
2162 | * in ext4_mb_release_context | |
2163 | */ | |
ccedf35b MWO |
2164 | ac->ac_bitmap_folio = e4b->bd_bitmap_folio; |
2165 | folio_get(ac->ac_bitmap_folio); | |
c84f1510 MWO |
2166 | ac->ac_buddy_folio = e4b->bd_buddy_folio; |
2167 | folio_get(ac->ac_buddy_folio); | |
c9de560d | 2168 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 2169 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
2170 | spin_lock(&sbi->s_md_lock); |
2171 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
2172 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
2173 | spin_unlock(&sbi->s_md_lock); | |
2174 | } | |
53f86b17 RH |
2175 | /* |
2176 | * As we've just preallocated more space than | |
2177 | * user requested originally, we store allocated | |
2178 | * space in a special descriptor. | |
2179 | */ | |
2180 | if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) | |
2181 | ext4_mb_new_preallocation(ac); | |
2182 | ||
c9de560d AT |
2183 | } |
2184 | ||
c9de560d AT |
2185 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
2186 | struct ext4_buddy *e4b, | |
2187 | int finish_group) | |
2188 | { | |
2189 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
2190 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2191 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
c9de560d | 2192 | |
032115fc AK |
2193 | if (ac->ac_status == AC_STATUS_FOUND) |
2194 | return; | |
c9de560d AT |
2195 | /* |
2196 | * We don't want to scan for a whole year | |
2197 | */ | |
2198 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
2199 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2200 | ac->ac_status = AC_STATUS_BREAK; | |
2201 | return; | |
2202 | } | |
2203 | ||
2204 | /* | |
2205 | * Haven't found good chunk so far, let's continue | |
2206 | */ | |
2207 | if (bex->fe_len < gex->fe_len) | |
2208 | return; | |
2209 | ||
3582e745 | 2210 | if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan) |
78dc9f84 | 2211 | ext4_mb_use_best_found(ac, e4b); |
c9de560d AT |
2212 | } |
2213 | ||
2214 | /* | |
2215 | * The routine checks whether found extent is good enough. If it is, | |
2216 | * then the extent gets marked used and flag is set to the context | |
2217 | * to stop scanning. Otherwise, the extent is compared with the | |
2218 | * previous found extent and if new one is better, then it's stored | |
2219 | * in the context. Later, the best found extent will be used, if | |
2220 | * mballoc can't find good enough extent. | |
2221 | * | |
3582e745 OM |
2222 | * The algorithm used is roughly as follows: |
2223 | * | |
2224 | * * If free extent found is exactly as big as goal, then | |
2225 | * stop the scan and use it immediately | |
2226 | * | |
2227 | * * If free extent found is smaller than goal, then keep retrying | |
2228 | * upto a max of sbi->s_mb_max_to_scan times (default 200). After | |
2229 | * that stop scanning and use whatever we have. | |
2230 | * | |
2231 | * * If free extent found is bigger than goal, then keep retrying | |
2232 | * upto a max of sbi->s_mb_min_to_scan times (default 10) before | |
2233 | * stopping the scan and using the extent. | |
2234 | * | |
2235 | * | |
c9de560d AT |
2236 | * FIXME: real allocation policy is to be designed yet! |
2237 | */ | |
2238 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
2239 | struct ext4_free_extent *ex, | |
2240 | struct ext4_buddy *e4b) | |
2241 | { | |
2242 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
2243 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
2244 | ||
2245 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
2246 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
2247 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
2248 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
2249 | ||
2250 | ac->ac_found++; | |
fdd9a009 | 2251 | ac->ac_cX_found[ac->ac_criteria]++; |
c9de560d AT |
2252 | |
2253 | /* | |
2254 | * The special case - take what you catch first | |
2255 | */ | |
2256 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2257 | *bex = *ex; | |
2258 | ext4_mb_use_best_found(ac, e4b); | |
2259 | return; | |
2260 | } | |
2261 | ||
2262 | /* | |
2263 | * Let's check whether the chuck is good enough | |
2264 | */ | |
2265 | if (ex->fe_len == gex->fe_len) { | |
2266 | *bex = *ex; | |
2267 | ext4_mb_use_best_found(ac, e4b); | |
2268 | return; | |
2269 | } | |
2270 | ||
2271 | /* | |
2272 | * If this is first found extent, just store it in the context | |
2273 | */ | |
2274 | if (bex->fe_len == 0) { | |
2275 | *bex = *ex; | |
2276 | return; | |
2277 | } | |
2278 | ||
2279 | /* | |
2280 | * If new found extent is better, store it in the context | |
2281 | */ | |
2282 | if (bex->fe_len < gex->fe_len) { | |
2283 | /* if the request isn't satisfied, any found extent | |
2284 | * larger than previous best one is better */ | |
2285 | if (ex->fe_len > bex->fe_len) | |
2286 | *bex = *ex; | |
2287 | } else if (ex->fe_len > gex->fe_len) { | |
2288 | /* if the request is satisfied, then we try to find | |
2289 | * an extent that still satisfy the request, but is | |
2290 | * smaller than previous one */ | |
2291 | if (ex->fe_len < bex->fe_len) | |
2292 | *bex = *ex; | |
2293 | } | |
2294 | ||
2295 | ext4_mb_check_limits(ac, e4b, 0); | |
2296 | } | |
2297 | ||
089ceecc | 2298 | static noinline_for_stack |
85b67ffb | 2299 | void ext4_mb_try_best_found(struct ext4_allocation_context *ac, |
c9de560d AT |
2300 | struct ext4_buddy *e4b) |
2301 | { | |
2302 | struct ext4_free_extent ex = ac->ac_b_ex; | |
2303 | ext4_group_t group = ex.fe_group; | |
2304 | int max; | |
2305 | int err; | |
2306 | ||
2307 | BUG_ON(ex.fe_len <= 0); | |
2308 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2309 | if (err) | |
85b67ffb | 2310 | return; |
c9de560d AT |
2311 | |
2312 | ext4_lock_group(ac->ac_sb, group); | |
4530b366 BL |
2313 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
2314 | goto out; | |
2315 | ||
15c006a2 | 2316 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
2317 | |
2318 | if (max > 0) { | |
2319 | ac->ac_b_ex = ex; | |
2320 | ext4_mb_use_best_found(ac, e4b); | |
2321 | } | |
2322 | ||
4530b366 | 2323 | out: |
c9de560d | 2324 | ext4_unlock_group(ac->ac_sb, group); |
e39e07fd | 2325 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2326 | } |
2327 | ||
089ceecc ES |
2328 | static noinline_for_stack |
2329 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
2330 | struct ext4_buddy *e4b) |
2331 | { | |
2332 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
2333 | int max; | |
2334 | int err; | |
2335 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 2336 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
2337 | struct ext4_free_extent ex; |
2338 | ||
5354b2af TT |
2339 | if (!grp) |
2340 | return -EFSCORRUPTED; | |
01e4ca29 | 2341 | if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY))) |
c9de560d | 2342 | return 0; |
838cd0cf YY |
2343 | if (grp->bb_free == 0) |
2344 | return 0; | |
c9de560d AT |
2345 | |
2346 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
2347 | if (err) | |
2348 | return err; | |
2349 | ||
2350 | ext4_lock_group(ac->ac_sb, group); | |
83269837 BL |
2351 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
2352 | goto out; | |
2353 | ||
15c006a2 | 2354 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 2355 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 2356 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d | 2357 | |
c3defd99 KS |
2358 | if (max >= ac->ac_g_ex.fe_len && |
2359 | ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) { | |
c9de560d AT |
2360 | ext4_fsblk_t start; |
2361 | ||
99c515e3 | 2362 | start = ext4_grp_offs_to_block(ac->ac_sb, &ex); |
c9de560d AT |
2363 | /* use do_div to get remainder (would be 64-bit modulo) */ |
2364 | if (do_div(start, sbi->s_stripe) == 0) { | |
2365 | ac->ac_found++; | |
2366 | ac->ac_b_ex = ex; | |
2367 | ext4_mb_use_best_found(ac, e4b); | |
2368 | } | |
2369 | } else if (max >= ac->ac_g_ex.fe_len) { | |
2370 | BUG_ON(ex.fe_len <= 0); | |
2371 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2372 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2373 | ac->ac_found++; | |
2374 | ac->ac_b_ex = ex; | |
2375 | ext4_mb_use_best_found(ac, e4b); | |
2376 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
2377 | /* Sometimes, caller may want to merge even small | |
2378 | * number of blocks to an existing extent */ | |
2379 | BUG_ON(ex.fe_len <= 0); | |
2380 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
2381 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
2382 | ac->ac_found++; | |
2383 | ac->ac_b_ex = ex; | |
2384 | ext4_mb_use_best_found(ac, e4b); | |
2385 | } | |
83269837 | 2386 | out: |
c9de560d | 2387 | ext4_unlock_group(ac->ac_sb, group); |
e39e07fd | 2388 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
2389 | |
2390 | return 0; | |
2391 | } | |
2392 | ||
2393 | /* | |
2394 | * The routine scans buddy structures (not bitmap!) from given order | |
2395 | * to max order and tries to find big enough chunk to satisfy the req | |
2396 | */ | |
089ceecc ES |
2397 | static noinline_for_stack |
2398 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2399 | struct ext4_buddy *e4b) |
2400 | { | |
2401 | struct super_block *sb = ac->ac_sb; | |
2402 | struct ext4_group_info *grp = e4b->bd_info; | |
2403 | void *buddy; | |
2404 | int i; | |
2405 | int k; | |
2406 | int max; | |
2407 | ||
2408 | BUG_ON(ac->ac_2order <= 0); | |
4b68f6df | 2409 | for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { |
c9de560d AT |
2410 | if (grp->bb_counters[i] == 0) |
2411 | continue; | |
2412 | ||
2413 | buddy = mb_find_buddy(e4b, i, &max); | |
19b8b035 TT |
2414 | if (WARN_RATELIMIT(buddy == NULL, |
2415 | "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i)) | |
2416 | continue; | |
c9de560d | 2417 | |
ffad0a44 | 2418 | k = mb_find_next_zero_bit(buddy, max, 0); |
eb576086 | 2419 | if (k >= max) { |
c5f3a382 BL |
2420 | ext4_mark_group_bitmap_corrupted(ac->ac_sb, |
2421 | e4b->bd_group, | |
2422 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
eb576086 DM |
2423 | ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, |
2424 | "%d free clusters of order %d. But found 0", | |
2425 | grp->bb_counters[i], i); | |
eb576086 DM |
2426 | break; |
2427 | } | |
c9de560d | 2428 | ac->ac_found++; |
fdd9a009 | 2429 | ac->ac_cX_found[ac->ac_criteria]++; |
c9de560d AT |
2430 | |
2431 | ac->ac_b_ex.fe_len = 1 << i; | |
2432 | ac->ac_b_ex.fe_start = k << i; | |
2433 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
2434 | ||
2435 | ext4_mb_use_best_found(ac, e4b); | |
2436 | ||
53f86b17 | 2437 | BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
c9de560d AT |
2438 | |
2439 | if (EXT4_SB(sb)->s_mb_stats) | |
2440 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
2441 | ||
2442 | break; | |
2443 | } | |
2444 | } | |
2445 | ||
2446 | /* | |
2447 | * The routine scans the group and measures all found extents. | |
2448 | * In order to optimize scanning, caller must pass number of | |
2449 | * free blocks in the group, so the routine can know upper limit. | |
2450 | */ | |
089ceecc ES |
2451 | static noinline_for_stack |
2452 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
2453 | struct ext4_buddy *e4b) |
2454 | { | |
2455 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 2456 | void *bitmap = e4b->bd_bitmap; |
c9de560d | 2457 | struct ext4_free_extent ex; |
1b420011 | 2458 | int i, j, freelen; |
c9de560d AT |
2459 | int free; |
2460 | ||
2461 | free = e4b->bd_info->bb_free; | |
907ea529 TT |
2462 | if (WARN_ON(free <= 0)) |
2463 | return; | |
c9de560d AT |
2464 | |
2465 | i = e4b->bd_info->bb_first_free; | |
2466 | ||
2467 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 2468 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
2469 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
2470 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 2471 | /* |
e56eb659 | 2472 | * IF we have corrupt bitmap, we won't find any |
26346ff6 | 2473 | * free blocks even though group info says we |
b483bb77 | 2474 | * have free blocks |
26346ff6 | 2475 | */ |
c5f3a382 BL |
2476 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2477 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e29136f8 | 2478 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2479 | "%d free clusters as per " |
fde4d95a | 2480 | "group info. But bitmap says 0", |
26346ff6 | 2481 | free); |
c9de560d AT |
2482 | break; |
2483 | } | |
2484 | ||
304749c0 | 2485 | if (!ext4_mb_cr_expensive(ac->ac_criteria)) { |
1b420011 | 2486 | /* |
f52f3d2b OM |
2487 | * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are |
2488 | * sure that this group will have a large enough | |
2489 | * continuous free extent, so skip over the smaller free | |
2490 | * extents | |
1b420011 OM |
2491 | */ |
2492 | j = mb_find_next_bit(bitmap, | |
2493 | EXT4_CLUSTERS_PER_GROUP(sb), i); | |
2494 | freelen = j - i; | |
2495 | ||
2496 | if (freelen < ac->ac_g_ex.fe_len) { | |
2497 | i = j; | |
2498 | free -= freelen; | |
2499 | continue; | |
2500 | } | |
2501 | } | |
2502 | ||
15c006a2 | 2503 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
907ea529 TT |
2504 | if (WARN_ON(ex.fe_len <= 0)) |
2505 | break; | |
26346ff6 | 2506 | if (free < ex.fe_len) { |
c5f3a382 BL |
2507 | ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
2508 | EXT4_GROUP_INFO_BBITMAP_CORRUPT); | |
e29136f8 | 2509 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 2510 | "%d free clusters as per " |
fde4d95a | 2511 | "group info. But got %d blocks", |
26346ff6 | 2512 | free, ex.fe_len); |
e56eb659 AK |
2513 | /* |
2514 | * The number of free blocks differs. This mostly | |
2515 | * indicate that the bitmap is corrupt. So exit | |
2516 | * without claiming the space. | |
2517 | */ | |
2518 | break; | |
26346ff6 | 2519 | } |
ab0c00fc | 2520 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
2521 | ext4_mb_measure_extent(ac, &ex, e4b); |
2522 | ||
2523 | i += ex.fe_len; | |
2524 | free -= ex.fe_len; | |
2525 | } | |
2526 | ||
2527 | ext4_mb_check_limits(ac, e4b, 1); | |
2528 | } | |
2529 | ||
2530 | /* | |
2531 | * This is a special case for storages like raid5 | |
506bf2d8 | 2532 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 2533 | */ |
089ceecc ES |
2534 | static noinline_for_stack |
2535 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
2536 | struct ext4_buddy *e4b) |
2537 | { | |
2538 | struct super_block *sb = ac->ac_sb; | |
2539 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 2540 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
2541 | struct ext4_free_extent ex; |
2542 | ext4_fsblk_t first_group_block; | |
2543 | ext4_fsblk_t a; | |
c3defd99 | 2544 | ext4_grpblk_t i, stripe; |
c9de560d AT |
2545 | int max; |
2546 | ||
2547 | BUG_ON(sbi->s_stripe == 0); | |
2548 | ||
2549 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
2550 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
2551 | ||
c9de560d AT |
2552 | a = first_group_block + sbi->s_stripe - 1; |
2553 | do_div(a, sbi->s_stripe); | |
2554 | i = (a * sbi->s_stripe) - first_group_block; | |
2555 | ||
c3defd99 KS |
2556 | stripe = EXT4_B2C(sbi, sbi->s_stripe); |
2557 | i = EXT4_B2C(sbi, i); | |
7137d7a4 | 2558 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 2559 | if (!mb_test_bit(i, bitmap)) { |
c3defd99 KS |
2560 | max = mb_find_extent(e4b, i, stripe, &ex); |
2561 | if (max >= stripe) { | |
c9de560d | 2562 | ac->ac_found++; |
fdd9a009 | 2563 | ac->ac_cX_found[ac->ac_criteria]++; |
ab0c00fc | 2564 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2565 | ac->ac_b_ex = ex; |
2566 | ext4_mb_use_best_found(ac, e4b); | |
2567 | break; | |
2568 | } | |
2569 | } | |
c3defd99 | 2570 | i += stripe; |
c9de560d AT |
2571 | } |
2572 | } | |
2573 | ||
42ac1848 | 2574 | /* |
8ef123fe | 2575 | * This is also called BEFORE we load the buddy bitmap. |
42ac1848 | 2576 | * Returns either 1 or 0 indicating that the group is either suitable |
8ef123fe | 2577 | * for the allocation or not. |
42ac1848 | 2578 | */ |
8ef123fe | 2579 | static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
4eb7a4a1 | 2580 | ext4_group_t group, enum criteria cr) |
c9de560d | 2581 | { |
8ef123fe | 2582 | ext4_grpblk_t free, fragments; |
a4912123 | 2583 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2584 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2585 | ||
f52f3d2b | 2586 | BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS); |
8a57d9d6 | 2587 | |
a9ce5993 | 2588 | if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
8ef123fe | 2589 | return false; |
01fc48e8 | 2590 | |
dddcd2f9 | 2591 | free = grp->bb_free; |
2592 | if (free == 0) | |
8ef123fe | 2593 | return false; |
c9de560d | 2594 | |
c9de560d | 2595 | fragments = grp->bb_fragments; |
c9de560d | 2596 | if (fragments == 0) |
8ef123fe | 2597 | return false; |
c9de560d AT |
2598 | |
2599 | switch (cr) { | |
f52f3d2b | 2600 | case CR_POWER2_ALIGNED: |
c9de560d | 2601 | BUG_ON(ac->ac_2order == 0); |
c9de560d | 2602 | |
a4912123 TT |
2603 | /* Avoid using the first bg of a flexgroup for data files */ |
2604 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2605 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2606 | ((group % flex_size) == 0)) | |
8ef123fe | 2607 | return false; |
a4912123 | 2608 | |
dddcd2f9 | 2609 | if (free < ac->ac_g_ex.fe_len) |
2610 | return false; | |
2611 | ||
4b68f6df | 2612 | if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) |
8ef123fe | 2613 | return true; |
40ae3487 TT |
2614 | |
2615 | if (grp->bb_largest_free_order < ac->ac_2order) | |
8ef123fe | 2616 | return false; |
40ae3487 | 2617 | |
8ef123fe | 2618 | return true; |
f52f3d2b OM |
2619 | case CR_GOAL_LEN_FAST: |
2620 | case CR_BEST_AVAIL_LEN: | |
c9de560d | 2621 | if ((free / fragments) >= ac->ac_g_ex.fe_len) |
8ef123fe | 2622 | return true; |
c9de560d | 2623 | break; |
f52f3d2b | 2624 | case CR_GOAL_LEN_SLOW: |
c9de560d | 2625 | if (free >= ac->ac_g_ex.fe_len) |
8ef123fe | 2626 | return true; |
c9de560d | 2627 | break; |
f52f3d2b | 2628 | case CR_ANY_FREE: |
8ef123fe | 2629 | return true; |
c9de560d AT |
2630 | default: |
2631 | BUG(); | |
2632 | } | |
2633 | ||
8ef123fe RH |
2634 | return false; |
2635 | } | |
2636 | ||
2637 | /* | |
2638 | * This could return negative error code if something goes wrong | |
2639 | * during ext4_mb_init_group(). This should not be called with | |
2640 | * ext4_lock_group() held. | |
a5fda113 TT |
2641 | * |
2642 | * Note: because we are conditionally operating with the group lock in | |
2643 | * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this | |
2644 | * function using __acquire and __release. This means we need to be | |
2645 | * super careful before messing with the error path handling via "goto | |
2646 | * out"! | |
8ef123fe RH |
2647 | */ |
2648 | static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, | |
4eb7a4a1 | 2649 | ext4_group_t group, enum criteria cr) |
8ef123fe RH |
2650 | { |
2651 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); | |
99377830 | 2652 | struct super_block *sb = ac->ac_sb; |
c1d2c7d4 | 2653 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
99377830 | 2654 | bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
8ef123fe RH |
2655 | ext4_grpblk_t free; |
2656 | int ret = 0; | |
2657 | ||
5354b2af TT |
2658 | if (!grp) |
2659 | return -EFSCORRUPTED; | |
a6c75eaf HS |
2660 | if (sbi->s_mb_stats) |
2661 | atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); | |
a5fda113 | 2662 | if (should_lock) { |
99377830 | 2663 | ext4_lock_group(sb, group); |
a5fda113 TT |
2664 | __release(ext4_group_lock_ptr(sb, group)); |
2665 | } | |
8ef123fe RH |
2666 | free = grp->bb_free; |
2667 | if (free == 0) | |
2668 | goto out; | |
304749c0 OM |
2669 | /* |
2670 | * In all criterias except CR_ANY_FREE we try to avoid groups that | |
2671 | * can't possibly satisfy the full goal request due to insufficient | |
2672 | * free blocks. | |
2673 | */ | |
2674 | if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len) | |
8ef123fe RH |
2675 | goto out; |
2676 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
2677 | goto out; | |
a5fda113 TT |
2678 | if (should_lock) { |
2679 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2680 | ext4_unlock_group(sb, group); |
a5fda113 | 2681 | } |
8ef123fe RH |
2682 | |
2683 | /* We only do this if the grp has never been initialized */ | |
2684 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
c1d2c7d4 AZ |
2685 | struct ext4_group_desc *gdp = |
2686 | ext4_get_group_desc(sb, group, NULL); | |
2687 | int ret; | |
2688 | ||
f52f3d2b | 2689 | /* |
2caffb6a | 2690 | * CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic |
f52f3d2b OM |
2691 | * search to find large good chunks almost for free. If buddy |
2692 | * data is not ready, then this optimization makes no sense. But | |
2693 | * we never skip the first block group in a flex_bg, since this | |
2694 | * gets used for metadata block allocation, and we want to make | |
2695 | * sure we locate metadata blocks in the first block group in | |
2696 | * the flex_bg if possible. | |
c1d2c7d4 | 2697 | */ |
304749c0 | 2698 | if (!ext4_mb_cr_expensive(cr) && |
c1d2c7d4 AZ |
2699 | (!sbi->s_log_groups_per_flex || |
2700 | ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && | |
2701 | !(ext4_has_group_desc_csum(sb) && | |
2702 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) | |
2703 | return 0; | |
2704 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); | |
8ef123fe RH |
2705 | if (ret) |
2706 | return ret; | |
2707 | } | |
2708 | ||
a5fda113 | 2709 | if (should_lock) { |
99377830 | 2710 | ext4_lock_group(sb, group); |
a5fda113 TT |
2711 | __release(ext4_group_lock_ptr(sb, group)); |
2712 | } | |
8ef123fe RH |
2713 | ret = ext4_mb_good_group(ac, group, cr); |
2714 | out: | |
a5fda113 TT |
2715 | if (should_lock) { |
2716 | __acquire(ext4_group_lock_ptr(sb, group)); | |
99377830 | 2717 | ext4_unlock_group(sb, group); |
a5fda113 | 2718 | } |
8ef123fe | 2719 | return ret; |
c9de560d AT |
2720 | } |
2721 | ||
cfd73237 AZ |
2722 | /* |
2723 | * Start prefetching @nr block bitmaps starting at @group. | |
2724 | * Return the next group which needs to be prefetched. | |
2725 | */ | |
3d392b26 TT |
2726 | ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
2727 | unsigned int nr, int *cnt) | |
cfd73237 AZ |
2728 | { |
2729 | ext4_group_t ngroups = ext4_get_groups_count(sb); | |
2730 | struct buffer_head *bh; | |
2731 | struct blk_plug plug; | |
2732 | ||
2733 | blk_start_plug(&plug); | |
2734 | while (nr-- > 0) { | |
2735 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, | |
2736 | NULL); | |
2737 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
2738 | ||
2739 | /* | |
2740 | * Prefetch block groups with free blocks; but don't | |
2741 | * bother if it is marked uninitialized on disk, since | |
2742 | * it won't require I/O to read. Also only try to | |
2743 | * prefetch once, so we avoid getblk() call, which can | |
2744 | * be expensive. | |
2745 | */ | |
5354b2af | 2746 | if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) && |
cfd73237 | 2747 | EXT4_MB_GRP_NEED_INIT(grp) && |
3c629604 | 2748 | ext4_free_group_clusters(sb, gdp) > 0 ) { |
cfd73237 AZ |
2749 | bh = ext4_read_block_bitmap_nowait(sb, group, true); |
2750 | if (bh && !IS_ERR(bh)) { | |
2751 | if (!buffer_uptodate(bh) && cnt) | |
2752 | (*cnt)++; | |
2753 | brelse(bh); | |
2754 | } | |
2755 | } | |
2756 | if (++group >= ngroups) | |
2757 | group = 0; | |
2758 | } | |
2759 | blk_finish_plug(&plug); | |
2760 | return group; | |
2761 | } | |
2762 | ||
2763 | /* | |
2764 | * Prefetching reads the block bitmap into the buffer cache; but we | |
2765 | * need to make sure that the buddy bitmap in the page cache has been | |
2766 | * initialized. Note that ext4_mb_init_group() will block if the I/O | |
2767 | * is not yet completed, or indeed if it was not initiated by | |
2768 | * ext4_mb_prefetch did not start the I/O. | |
2769 | * | |
2770 | * TODO: We should actually kick off the buddy bitmap setup in a work | |
2771 | * queue when the buffer I/O is completed, so that we don't block | |
2772 | * waiting for the block allocation bitmap read to finish when | |
2773 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). | |
2774 | */ | |
3d392b26 TT |
2775 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
2776 | unsigned int nr) | |
cfd73237 | 2777 | { |
22fab984 KS |
2778 | struct ext4_group_desc *gdp; |
2779 | struct ext4_group_info *grp; | |
cfd73237 | 2780 | |
22fab984 | 2781 | while (nr-- > 0) { |
cfd73237 AZ |
2782 | if (!group) |
2783 | group = ext4_get_groups_count(sb); | |
2784 | group--; | |
22fab984 | 2785 | gdp = ext4_get_group_desc(sb, group, NULL); |
cfd73237 AZ |
2786 | grp = ext4_get_group_info(sb, group); |
2787 | ||
5354b2af | 2788 | if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) && |
3c629604 | 2789 | ext4_free_group_clusters(sb, gdp) > 0) { |
cfd73237 AZ |
2790 | if (ext4_mb_init_group(sb, group, GFP_NOFS)) |
2791 | break; | |
2792 | } | |
2793 | } | |
2794 | } | |
2795 | ||
4ddfef7b ES |
2796 | static noinline_for_stack int |
2797 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2798 | { |
cfd73237 | 2799 | ext4_group_t prefetch_grp = 0, ngroups, group, i; |
4c0cfebd | 2800 | enum criteria new_cr, cr = CR_GOAL_LEN_FAST; |
42ac1848 | 2801 | int err = 0, first_err = 0; |
cfd73237 | 2802 | unsigned int nr = 0, prefetch_ios = 0; |
c9de560d AT |
2803 | struct ext4_sb_info *sbi; |
2804 | struct super_block *sb; | |
2805 | struct ext4_buddy e4b; | |
66d5e027 | 2806 | int lost; |
c9de560d AT |
2807 | |
2808 | sb = ac->ac_sb; | |
2809 | sbi = EXT4_SB(sb); | |
8df9675f | 2810 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2811 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2812 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2813 | ngroups = sbi->s_blockfile_groups; |
2814 | ||
c9de560d AT |
2815 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2816 | ||
2817 | /* first, try the goal */ | |
2818 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2819 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2820 | goto out; | |
2821 | ||
2822 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2823 | goto out; | |
2824 | ||
2825 | /* | |
e9a3cd48 | 2826 | * ac->ac_2order is set only if the fe_len is a power of 2 |
4eea9fbe KS |
2827 | * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED |
2828 | * so that we try exact allocation using buddy. | |
c9de560d AT |
2829 | */ |
2830 | i = fls(ac->ac_g_ex.fe_len); | |
2831 | ac->ac_2order = 0; | |
2832 | /* | |
2833 | * We search using buddy data only if the order of the request | |
2834 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2835 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
d9b22cf9 JK |
2836 | * We also support searching for power-of-two requests only for |
2837 | * requests upto maximum buddy size we have constructed. | |
c9de560d | 2838 | */ |
4b68f6df | 2839 | if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { |
bb60caa2 | 2840 | if (is_power_of_2(ac->ac_g_ex.fe_len)) |
1a5d5e5d | 2841 | ac->ac_2order = array_index_nospec(i - 1, |
4b68f6df | 2842 | MB_NUM_ORDERS(sb)); |
c9de560d AT |
2843 | } |
2844 | ||
4ba74d00 TT |
2845 | /* if stream allocation is enabled, use global goal */ |
2846 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2847 | /* TBD: may be hot point */ |
2848 | spin_lock(&sbi->s_md_lock); | |
2849 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2850 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2851 | spin_unlock(&sbi->s_md_lock); | |
2852 | } | |
4ba74d00 | 2853 | |
c9de560d | 2854 | /* |
4c0cfebd TT |
2855 | * Let's just scan groups to find more-less suitable blocks We |
2856 | * start with CR_GOAL_LEN_FAST, unless it is power of 2 | |
2857 | * aligned, in which case let's do that faster approach first. | |
c9de560d | 2858 | */ |
4c0cfebd TT |
2859 | if (ac->ac_2order) |
2860 | cr = CR_POWER2_ALIGNED; | |
c9de560d | 2861 | repeat: |
4eb7a4a1 | 2862 | for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) { |
c9de560d | 2863 | ac->ac_criteria = cr; |
ed8f9c75 AK |
2864 | /* |
2865 | * searching for the right group start | |
2866 | * from the goal value specified | |
2867 | */ | |
2868 | group = ac->ac_g_ex.fe_group; | |
196e402a | 2869 | ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; |
cfd73237 | 2870 | prefetch_grp = group; |
9c97c34a | 2871 | nr = 0; |
ed8f9c75 | 2872 | |
4fca50d4 JK |
2873 | for (i = 0, new_cr = cr; i < ngroups; i++, |
2874 | ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { | |
2875 | int ret = 0; | |
196e402a | 2876 | |
2ed5724d | 2877 | cond_resched(); |
196e402a HS |
2878 | if (new_cr != cr) { |
2879 | cr = new_cr; | |
2880 | goto repeat; | |
2881 | } | |
c9de560d | 2882 | |
cfd73237 AZ |
2883 | /* |
2884 | * Batch reads of the block allocation bitmaps | |
2885 | * to get multiple READs in flight; limit | |
4eea9fbe KS |
2886 | * prefetching at inexpensive CR, otherwise mballoc |
2887 | * can spend a lot of time loading imperfect groups | |
cfd73237 AZ |
2888 | */ |
2889 | if ((prefetch_grp == group) && | |
304749c0 | 2890 | (ext4_mb_cr_expensive(cr) || |
cfd73237 | 2891 | prefetch_ios < sbi->s_mb_prefetch_limit)) { |
cfd73237 AZ |
2892 | nr = sbi->s_mb_prefetch; |
2893 | if (ext4_has_feature_flex_bg(sb)) { | |
82ef1370 CX |
2894 | nr = 1 << sbi->s_log_groups_per_flex; |
2895 | nr -= group & (nr - 1); | |
2896 | nr = min(nr, sbi->s_mb_prefetch); | |
cfd73237 AZ |
2897 | } |
2898 | prefetch_grp = ext4_mb_prefetch(sb, group, | |
2899 | nr, &prefetch_ios); | |
cfd73237 AZ |
2900 | } |
2901 | ||
8a57d9d6 | 2902 | /* This now checks without needing the buddy page */ |
8ef123fe | 2903 | ret = ext4_mb_good_group_nolock(ac, group, cr); |
42ac1848 LC |
2904 | if (ret <= 0) { |
2905 | if (!first_err) | |
2906 | first_err = ret; | |
c9de560d | 2907 | continue; |
42ac1848 | 2908 | } |
c9de560d | 2909 | |
c9de560d AT |
2910 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2911 | if (err) | |
2912 | goto out; | |
2913 | ||
2914 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2915 | |
2916 | /* | |
2917 | * We need to check again after locking the | |
2918 | * block group | |
2919 | */ | |
42ac1848 | 2920 | ret = ext4_mb_good_group(ac, group, cr); |
8ef123fe | 2921 | if (ret == 0) { |
c9de560d | 2922 | ext4_unlock_group(sb, group); |
e39e07fd | 2923 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2924 | continue; |
2925 | } | |
2926 | ||
2927 | ac->ac_groups_scanned++; | |
f52f3d2b | 2928 | if (cr == CR_POWER2_ALIGNED) |
c9de560d | 2929 | ext4_mb_simple_scan_group(ac, &e4b); |
1f6bc02f OM |
2930 | else { |
2931 | bool is_stripe_aligned = sbi->s_stripe && | |
2932 | !(ac->ac_g_ex.fe_len % | |
2933 | EXT4_B2C(sbi, sbi->s_stripe)); | |
2934 | ||
2935 | if ((cr == CR_GOAL_LEN_FAST || | |
2936 | cr == CR_BEST_AVAIL_LEN) && | |
2937 | is_stripe_aligned) | |
2938 | ext4_mb_scan_aligned(ac, &e4b); | |
2939 | ||
2940 | if (ac->ac_status == AC_STATUS_CONTINUE) | |
2941 | ext4_mb_complex_scan_group(ac, &e4b); | |
2942 | } | |
c9de560d AT |
2943 | |
2944 | ext4_unlock_group(sb, group); | |
e39e07fd | 2945 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2946 | |
2947 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2948 | break; | |
2949 | } | |
a6c75eaf HS |
2950 | /* Processed all groups and haven't found blocks */ |
2951 | if (sbi->s_mb_stats && i == ngroups) | |
2952 | atomic64_inc(&sbi->s_bal_cX_failed[cr]); | |
7e170922 | 2953 | |
f52f3d2b | 2954 | if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN) |
7e170922 | 2955 | /* Reset goal length to original goal length before |
f52f3d2b | 2956 | * falling into CR_GOAL_LEN_SLOW */ |
7e170922 | 2957 | ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; |
c9de560d AT |
2958 | } |
2959 | ||
2960 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2961 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2962 | /* | |
2963 | * We've been searching too long. Let's try to allocate | |
2964 | * the best chunk we've found so far | |
2965 | */ | |
c9de560d AT |
2966 | ext4_mb_try_best_found(ac, &e4b); |
2967 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2968 | /* | |
2969 | * Someone more lucky has already allocated it. | |
2970 | * The only thing we can do is just take first | |
2971 | * found block(s) | |
c9de560d | 2972 | */ |
66d5e027 | 2973 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
2974 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", | |
c55ee7d2 | 2975 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
2976 | ac->ac_b_ex.fe_len, lost); | |
2977 | ||
c9de560d AT |
2978 | ac->ac_b_ex.fe_group = 0; |
2979 | ac->ac_b_ex.fe_start = 0; | |
2980 | ac->ac_b_ex.fe_len = 0; | |
2981 | ac->ac_status = AC_STATUS_CONTINUE; | |
2982 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
f52f3d2b | 2983 | cr = CR_ANY_FREE; |
c9de560d AT |
2984 | goto repeat; |
2985 | } | |
2986 | } | |
a6c75eaf HS |
2987 | |
2988 | if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) | |
2989 | atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); | |
c9de560d | 2990 | out: |
42ac1848 LC |
2991 | if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) |
2992 | err = first_err; | |
bbc4ec77 | 2993 | |
d3df1453 | 2994 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
bbc4ec77 RH |
2995 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
2996 | ac->ac_flags, cr, err); | |
cfd73237 AZ |
2997 | |
2998 | if (nr) | |
2999 | ext4_mb_prefetch_fini(sb, prefetch_grp, nr); | |
3000 | ||
c9de560d AT |
3001 | return err; |
3002 | } | |
3003 | ||
c9de560d AT |
3004 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
3005 | { | |
359745d7 | 3006 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
3007 | ext4_group_t group; |
3008 | ||
8df9675f | 3009 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 3010 | return NULL; |
c9de560d | 3011 | group = *pos + 1; |
a9df9a49 | 3012 | return (void *) ((unsigned long) group); |
c9de560d AT |
3013 | } |
3014 | ||
3015 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
3016 | { | |
359745d7 | 3017 | struct super_block *sb = pde_data(file_inode(seq->file)); |
c9de560d AT |
3018 | ext4_group_t group; |
3019 | ||
3020 | ++*pos; | |
8df9675f | 3021 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
3022 | return NULL; |
3023 | group = *pos + 1; | |
a9df9a49 | 3024 | return (void *) ((unsigned long) group); |
c9de560d AT |
3025 | } |
3026 | ||
3027 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
3028 | { | |
359745d7 | 3029 | struct super_block *sb = pde_data(file_inode(seq->file)); |
a9df9a49 | 3030 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
25044880 | 3031 | int i, err; |
4b55d343 | 3032 | char nbuf[16]; |
c9de560d | 3033 | struct ext4_buddy e4b; |
1c8457ca | 3034 | struct ext4_group_info *grinfo; |
2df2c340 AB |
3035 | unsigned char blocksize_bits = min_t(unsigned char, |
3036 | sb->s_blocksize_bits, | |
3037 | EXT4_MAX_BLOCK_LOG_SIZE); | |
c9de560d AT |
3038 | struct sg { |
3039 | struct ext4_group_info info; | |
b80b32b6 | 3040 | ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; |
c9de560d AT |
3041 | } sg; |
3042 | ||
3043 | group--; | |
3044 | if (group == 0) | |
97b4af2f RV |
3045 | seq_puts(seq, "#group: free frags first [" |
3046 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " | |
802cf1f9 | 3047 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
c9de560d | 3048 | |
b80b32b6 TT |
3049 | i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + |
3050 | sizeof(struct ext4_group_info); | |
3051 | ||
1c8457ca | 3052 | grinfo = ext4_get_group_info(sb, group); |
5354b2af TT |
3053 | if (!grinfo) |
3054 | return 0; | |
1c8457ca AK |
3055 | /* Load the group info in memory only if not already loaded. */ |
3056 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
3057 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
3058 | if (err) { | |
4b55d343 | 3059 | seq_printf(seq, "#%-5u: %s\n", group, ext4_decode_error(NULL, err, nbuf)); |
1c8457ca AK |
3060 | return 0; |
3061 | } | |
25044880 | 3062 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 3063 | } |
1c8457ca | 3064 | |
25044880 | 3065 | /* |
3066 | * We care only about free space counters in the group info and | |
3067 | * these are safe to access even after the buddy has been unloaded | |
3068 | */ | |
5354b2af | 3069 | memcpy(&sg, grinfo, i); |
a9df9a49 | 3070 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
3071 | sg.info.bb_fragments, sg.info.bb_first_free); |
3072 | for (i = 0; i <= 13; i++) | |
2df2c340 | 3073 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
c9de560d | 3074 | sg.info.bb_counters[i] : 0); |
68ee261f ZY |
3075 | seq_puts(seq, " ]"); |
3076 | if (EXT4_MB_GRP_BBITMAP_CORRUPT(&sg.info)) | |
3077 | seq_puts(seq, " Block bitmap corrupted!"); | |
3078 | seq_puts(seq, "\n"); | |
c9de560d AT |
3079 | |
3080 | return 0; | |
3081 | } | |
3082 | ||
3083 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
3084 | { | |
3085 | } | |
3086 | ||
247dbed8 | 3087 | const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
3088 | .start = ext4_mb_seq_groups_start, |
3089 | .next = ext4_mb_seq_groups_next, | |
3090 | .stop = ext4_mb_seq_groups_stop, | |
3091 | .show = ext4_mb_seq_groups_show, | |
3092 | }; | |
3093 | ||
a6c75eaf HS |
3094 | int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) |
3095 | { | |
c30365b9 | 3096 | struct super_block *sb = seq->private; |
a6c75eaf HS |
3097 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3098 | ||
3099 | seq_puts(seq, "mballoc:\n"); | |
3100 | if (!sbi->s_mb_stats) { | |
3101 | seq_puts(seq, "\tmb stats collection turned off.\n"); | |
f52f3d2b OM |
3102 | seq_puts( |
3103 | seq, | |
3104 | "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); | |
a6c75eaf HS |
3105 | return 0; |
3106 | } | |
3107 | seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); | |
3108 | seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); | |
3109 | ||
f52f3d2b OM |
3110 | seq_printf(seq, "\tgroups_scanned: %u\n", |
3111 | atomic_read(&sbi->s_bal_groups_scanned)); | |
3112 | ||
3113 | /* CR_POWER2_ALIGNED stats */ | |
3114 | seq_puts(seq, "\tcr_p2_aligned_stats:\n"); | |
3115 | seq_printf(seq, "\t\thits: %llu\n", | |
3116 | atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED])); | |
3117 | seq_printf( | |
3118 | seq, "\t\tgroups_considered: %llu\n", | |
3119 | atomic64_read( | |
3120 | &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED])); | |
3121 | seq_printf(seq, "\t\textents_scanned: %u\n", | |
3122 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED])); | |
a6c75eaf | 3123 | seq_printf(seq, "\t\tuseless_loops: %llu\n", |
f52f3d2b | 3124 | atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED])); |
196e402a | 3125 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
f52f3d2b | 3126 | atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions)); |
a6c75eaf | 3127 | |
f52f3d2b OM |
3128 | /* CR_GOAL_LEN_FAST stats */ |
3129 | seq_puts(seq, "\tcr_goal_fast_stats:\n"); | |
3130 | seq_printf(seq, "\t\thits: %llu\n", | |
3131 | atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST])); | |
a6c75eaf | 3132 | seq_printf(seq, "\t\tgroups_considered: %llu\n", |
f52f3d2b OM |
3133 | atomic64_read( |
3134 | &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST])); | |
3135 | seq_printf(seq, "\t\textents_scanned: %u\n", | |
3136 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST])); | |
a6c75eaf | 3137 | seq_printf(seq, "\t\tuseless_loops: %llu\n", |
f52f3d2b | 3138 | atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST])); |
196e402a | 3139 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
f52f3d2b OM |
3140 | atomic_read(&sbi->s_bal_goal_fast_bad_suggestions)); |
3141 | ||
3142 | /* CR_BEST_AVAIL_LEN stats */ | |
3143 | seq_puts(seq, "\tcr_best_avail_stats:\n"); | |
3144 | seq_printf(seq, "\t\thits: %llu\n", | |
3145 | atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN])); | |
3146 | seq_printf( | |
3147 | seq, "\t\tgroups_considered: %llu\n", | |
3148 | atomic64_read( | |
3149 | &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN])); | |
3150 | seq_printf(seq, "\t\textents_scanned: %u\n", | |
3151 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN])); | |
7e170922 | 3152 | seq_printf(seq, "\t\tuseless_loops: %llu\n", |
f52f3d2b | 3153 | atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN])); |
7e170922 | 3154 | seq_printf(seq, "\t\tbad_suggestions: %u\n", |
f52f3d2b | 3155 | atomic_read(&sbi->s_bal_best_avail_bad_suggestions)); |
7e170922 | 3156 | |
f52f3d2b OM |
3157 | /* CR_GOAL_LEN_SLOW stats */ |
3158 | seq_puts(seq, "\tcr_goal_slow_stats:\n"); | |
3159 | seq_printf(seq, "\t\thits: %llu\n", | |
3160 | atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW])); | |
a6c75eaf | 3161 | seq_printf(seq, "\t\tgroups_considered: %llu\n", |
f52f3d2b OM |
3162 | atomic64_read( |
3163 | &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW])); | |
3164 | seq_printf(seq, "\t\textents_scanned: %u\n", | |
3165 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW])); | |
a6c75eaf | 3166 | seq_printf(seq, "\t\tuseless_loops: %llu\n", |
f52f3d2b OM |
3167 | atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW])); |
3168 | ||
3169 | /* CR_ANY_FREE stats */ | |
3170 | seq_puts(seq, "\tcr_any_free_stats:\n"); | |
3171 | seq_printf(seq, "\t\thits: %llu\n", | |
3172 | atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE])); | |
3173 | seq_printf( | |
3174 | seq, "\t\tgroups_considered: %llu\n", | |
3175 | atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE])); | |
3176 | seq_printf(seq, "\t\textents_scanned: %u\n", | |
3177 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE])); | |
a6c75eaf | 3178 | seq_printf(seq, "\t\tuseless_loops: %llu\n", |
f52f3d2b OM |
3179 | atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE])); |
3180 | ||
3181 | /* Aggregates */ | |
3182 | seq_printf(seq, "\textents_scanned: %u\n", | |
3183 | atomic_read(&sbi->s_bal_ex_scanned)); | |
a6c75eaf | 3184 | seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); |
f52f3d2b OM |
3185 | seq_printf(seq, "\t\tlen_goal_hits: %u\n", |
3186 | atomic_read(&sbi->s_bal_len_goals)); | |
a6c75eaf HS |
3187 | seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); |
3188 | seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); | |
3189 | seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); | |
a6c75eaf HS |
3190 | seq_printf(seq, "\tbuddies_generated: %u/%u\n", |
3191 | atomic_read(&sbi->s_mb_buddies_generated), | |
3192 | ext4_get_groups_count(sb)); | |
3193 | seq_printf(seq, "\tbuddies_time_used: %llu\n", | |
3194 | atomic64_read(&sbi->s_mb_generation_time)); | |
3195 | seq_printf(seq, "\tpreallocated: %u\n", | |
3196 | atomic_read(&sbi->s_mb_preallocated)); | |
f52f3d2b | 3197 | seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded)); |
a6c75eaf HS |
3198 | return 0; |
3199 | } | |
3200 | ||
f68f4063 HS |
3201 | static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) |
3202 | { | |
359745d7 | 3203 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
3204 | unsigned long position; |
3205 | ||
83e80a6e | 3206 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
f68f4063 HS |
3207 | return NULL; |
3208 | position = *pos + 1; | |
3209 | return (void *) ((unsigned long) position); | |
3210 | } | |
3211 | ||
3212 | static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) | |
3213 | { | |
359745d7 | 3214 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
3215 | unsigned long position; |
3216 | ||
3217 | ++*pos; | |
83e80a6e | 3218 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
f68f4063 HS |
3219 | return NULL; |
3220 | position = *pos + 1; | |
3221 | return (void *) ((unsigned long) position); | |
3222 | } | |
3223 | ||
3224 | static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) | |
3225 | { | |
359745d7 | 3226 | struct super_block *sb = pde_data(file_inode(seq->file)); |
f68f4063 HS |
3227 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3228 | unsigned long position = ((unsigned long) v); | |
3229 | struct ext4_group_info *grp; | |
83e80a6e | 3230 | unsigned int count; |
f68f4063 HS |
3231 | |
3232 | position--; | |
3233 | if (position >= MB_NUM_ORDERS(sb)) { | |
83e80a6e JK |
3234 | position -= MB_NUM_ORDERS(sb); |
3235 | if (position == 0) | |
3236 | seq_puts(seq, "avg_fragment_size_lists:\n"); | |
f68f4063 | 3237 | |
83e80a6e JK |
3238 | count = 0; |
3239 | read_lock(&sbi->s_mb_avg_fragment_size_locks[position]); | |
3240 | list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position], | |
3241 | bb_avg_fragment_size_node) | |
3242 | count++; | |
3243 | read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]); | |
3244 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", | |
3245 | (unsigned int)position, count); | |
f68f4063 HS |
3246 | return 0; |
3247 | } | |
3248 | ||
3249 | if (position == 0) { | |
3250 | seq_printf(seq, "optimize_scan: %d\n", | |
3251 | test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); | |
3252 | seq_puts(seq, "max_free_order_lists:\n"); | |
3253 | } | |
3254 | count = 0; | |
83e80a6e | 3255 | read_lock(&sbi->s_mb_largest_free_orders_locks[position]); |
f68f4063 HS |
3256 | list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], |
3257 | bb_largest_free_order_node) | |
3258 | count++; | |
83e80a6e | 3259 | read_unlock(&sbi->s_mb_largest_free_orders_locks[position]); |
f68f4063 HS |
3260 | seq_printf(seq, "\tlist_order_%u_groups: %u\n", |
3261 | (unsigned int)position, count); | |
3262 | ||
3263 | return 0; | |
3264 | } | |
3265 | ||
3266 | static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) | |
3267 | { | |
f68f4063 HS |
3268 | } |
3269 | ||
3270 | const struct seq_operations ext4_mb_seq_structs_summary_ops = { | |
3271 | .start = ext4_mb_seq_structs_summary_start, | |
3272 | .next = ext4_mb_seq_structs_summary_next, | |
3273 | .stop = ext4_mb_seq_structs_summary_stop, | |
3274 | .show = ext4_mb_seq_structs_summary_show, | |
3275 | }; | |
3276 | ||
fb1813f4 CW |
3277 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
3278 | { | |
3279 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3280 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
3281 | ||
3282 | BUG_ON(!cachep); | |
3283 | return cachep; | |
3284 | } | |
5f21b0e6 | 3285 | |
28623c2f TT |
3286 | /* |
3287 | * Allocate the top-level s_group_info array for the specified number | |
3288 | * of groups | |
3289 | */ | |
3290 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
3291 | { | |
3292 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3293 | unsigned size; | |
df3da4ea | 3294 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
28623c2f TT |
3295 | |
3296 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
3297 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
3298 | if (size <= sbi->s_group_info_size) | |
3299 | return 0; | |
3300 | ||
3301 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
a7c3e901 | 3302 | new_groupinfo = kvzalloc(size, GFP_KERNEL); |
28623c2f TT |
3303 | if (!new_groupinfo) { |
3304 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
3305 | return -ENOMEM; | |
3306 | } | |
df3da4ea SJS |
3307 | rcu_read_lock(); |
3308 | old_groupinfo = rcu_dereference(sbi->s_group_info); | |
3309 | if (old_groupinfo) | |
3310 | memcpy(new_groupinfo, old_groupinfo, | |
28623c2f | 3311 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
df3da4ea SJS |
3312 | rcu_read_unlock(); |
3313 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo); | |
28623c2f | 3314 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
df3da4ea SJS |
3315 | if (old_groupinfo) |
3316 | ext4_kvfree_array_rcu(old_groupinfo); | |
666245d9 | 3317 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
28623c2f TT |
3318 | sbi->s_group_info_size); |
3319 | return 0; | |
3320 | } | |
3321 | ||
5f21b0e6 | 3322 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 3323 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
3324 | struct ext4_group_desc *desc) |
3325 | { | |
fb1813f4 | 3326 | int i; |
5f21b0e6 | 3327 | int metalen = 0; |
df3da4ea | 3328 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
5f21b0e6 FB |
3329 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
3330 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 3331 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
3332 | |
3333 | /* | |
3334 | * First check if this group is the first of a reserved block. | |
3335 | * If it's true, we have to allocate a new table of pointers | |
3336 | * to ext4_group_info structures | |
3337 | */ | |
3338 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
3339 | metalen = sizeof(*meta_group_info) << | |
3340 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
4fdb5543 | 3341 | meta_group_info = kmalloc(metalen, GFP_NOFS); |
5f21b0e6 | 3342 | if (meta_group_info == NULL) { |
7f6a11e7 | 3343 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 3344 | "for a buddy group"); |
df119095 | 3345 | return -ENOMEM; |
5f21b0e6 | 3346 | } |
df3da4ea SJS |
3347 | rcu_read_lock(); |
3348 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; | |
3349 | rcu_read_unlock(); | |
5f21b0e6 FB |
3350 | } |
3351 | ||
df3da4ea | 3352 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
5f21b0e6 FB |
3353 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
3354 | ||
4fdb5543 | 3355 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
5f21b0e6 | 3356 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 3357 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
3358 | goto exit_group_info; |
3359 | } | |
3360 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
3361 | &(meta_group_info[i]->bb_state)); | |
3362 | ||
3363 | /* | |
3364 | * initialize bb_free to be able to skip | |
3365 | * empty groups without initialization | |
3366 | */ | |
8844618d TT |
3367 | if (ext4_has_group_desc_csum(sb) && |
3368 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
5f21b0e6 | 3369 | meta_group_info[i]->bb_free = |
cff1dfd7 | 3370 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
3371 | } else { |
3372 | meta_group_info[i]->bb_free = | |
021b65bb | 3373 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
3374 | } |
3375 | ||
3376 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 3377 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 3378 | meta_group_info[i]->bb_free_root = RB_ROOT; |
196e402a | 3379 | INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); |
83e80a6e | 3380 | INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); |
8a57d9d6 | 3381 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
83e80a6e | 3382 | meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ |
196e402a | 3383 | meta_group_info[i]->bb_group = group; |
5f21b0e6 | 3384 | |
a3450215 | 3385 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
5f21b0e6 FB |
3386 | return 0; |
3387 | ||
3388 | exit_group_info: | |
3389 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 3390 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
df3da4ea SJS |
3391 | struct ext4_group_info ***group_info; |
3392 | ||
3393 | rcu_read_lock(); | |
3394 | group_info = rcu_dereference(sbi->s_group_info); | |
3395 | kfree(group_info[idx]); | |
3396 | group_info[idx] = NULL; | |
3397 | rcu_read_unlock(); | |
caaf7a29 | 3398 | } |
5f21b0e6 FB |
3399 | return -ENOMEM; |
3400 | } /* ext4_mb_add_groupinfo */ | |
3401 | ||
c9de560d AT |
3402 | static int ext4_mb_init_backend(struct super_block *sb) |
3403 | { | |
8df9675f | 3404 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 3405 | ext4_group_t i; |
c9de560d | 3406 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 3407 | int err; |
5f21b0e6 | 3408 | struct ext4_group_desc *desc; |
df3da4ea | 3409 | struct ext4_group_info ***group_info; |
fb1813f4 | 3410 | struct kmem_cache *cachep; |
5f21b0e6 | 3411 | |
28623c2f TT |
3412 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
3413 | if (err) | |
3414 | return err; | |
c9de560d | 3415 | |
c9de560d AT |
3416 | sbi->s_buddy_cache = new_inode(sb); |
3417 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 3418 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
3419 | goto err_freesgi; |
3420 | } | |
48e6061b YJ |
3421 | /* To avoid potentially colliding with an valid on-disk inode number, |
3422 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
3423 | * not in the inode hash, so it should never be found by iget(), but | |
3424 | * this will avoid confusion if it ever shows up during debugging. */ | |
3425 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 3426 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 3427 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3428 | cond_resched(); |
c9de560d AT |
3429 | desc = ext4_get_group_desc(sb, i, NULL); |
3430 | if (desc == NULL) { | |
9d8b9ec4 | 3431 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
3432 | goto err_freebuddy; |
3433 | } | |
5f21b0e6 FB |
3434 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
3435 | goto err_freebuddy; | |
c9de560d AT |
3436 | } |
3437 | ||
cfd73237 | 3438 | if (ext4_has_feature_flex_bg(sb)) { |
f91436d5 ST |
3439 | /* a single flex group is supposed to be read by a single IO. |
3440 | * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is | |
3441 | * unsigned integer, so the maximum shift is 32. | |
3442 | */ | |
3443 | if (sbi->s_es->s_log_groups_per_flex >= 32) { | |
3444 | ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); | |
a8867f4e | 3445 | goto err_freebuddy; |
f91436d5 ST |
3446 | } |
3447 | sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, | |
82ef1370 | 3448 | BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); |
cfd73237 AZ |
3449 | sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ |
3450 | } else { | |
3451 | sbi->s_mb_prefetch = 32; | |
3452 | } | |
3453 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) | |
3454 | sbi->s_mb_prefetch = ext4_get_groups_count(sb); | |
2caffb6a KS |
3455 | /* |
3456 | * now many real IOs to prefetch within a single allocation at | |
3457 | * CR_POWER2_ALIGNED. Given CR_POWER2_ALIGNED is an CPU-related | |
3458 | * optimization we shouldn't try to load too many groups, at some point | |
3459 | * we should start to use what we've got in memory. | |
cfd73237 AZ |
3460 | * with an average random access time 5ms, it'd take a second to get |
3461 | * 200 groups (* N with flex_bg), so let's make this limit 4 | |
3462 | */ | |
3463 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; | |
3464 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) | |
3465 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); | |
3466 | ||
c9de560d AT |
3467 | return 0; |
3468 | ||
3469 | err_freebuddy: | |
fb1813f4 | 3470 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5354b2af TT |
3471 | while (i-- > 0) { |
3472 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); | |
3473 | ||
3474 | if (grp) | |
3475 | kmem_cache_free(cachep, grp); | |
3476 | } | |
28623c2f | 3477 | i = sbi->s_group_info_size; |
df3da4ea SJS |
3478 | rcu_read_lock(); |
3479 | group_info = rcu_dereference(sbi->s_group_info); | |
f1fa3342 | 3480 | while (i-- > 0) |
df3da4ea SJS |
3481 | kfree(group_info[i]); |
3482 | rcu_read_unlock(); | |
c9de560d AT |
3483 | iput(sbi->s_buddy_cache); |
3484 | err_freesgi: | |
df3da4ea SJS |
3485 | rcu_read_lock(); |
3486 | kvfree(rcu_dereference(sbi->s_group_info)); | |
3487 | rcu_read_unlock(); | |
c9de560d AT |
3488 | return -ENOMEM; |
3489 | } | |
3490 | ||
2892c15d ES |
3491 | static void ext4_groupinfo_destroy_slabs(void) |
3492 | { | |
3493 | int i; | |
3494 | ||
3495 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
21c580d8 | 3496 | kmem_cache_destroy(ext4_groupinfo_caches[i]); |
2892c15d ES |
3497 | ext4_groupinfo_caches[i] = NULL; |
3498 | } | |
3499 | } | |
3500 | ||
3501 | static int ext4_groupinfo_create_slab(size_t size) | |
3502 | { | |
3503 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
3504 | int slab_size; | |
3505 | int blocksize_bits = order_base_2(size); | |
3506 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
3507 | struct kmem_cache *cachep; | |
3508 | ||
3509 | if (cache_index >= NR_GRPINFO_CACHES) | |
3510 | return -EINVAL; | |
3511 | ||
3512 | if (unlikely(cache_index < 0)) | |
3513 | cache_index = 0; | |
3514 | ||
3515 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
3516 | if (ext4_groupinfo_caches[cache_index]) { | |
3517 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
3518 | return 0; /* Already created */ | |
3519 | } | |
3520 | ||
3521 | slab_size = offsetof(struct ext4_group_info, | |
3522 | bb_counters[blocksize_bits + 2]); | |
3523 | ||
3524 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
3525 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
3526 | NULL); | |
3527 | ||
823ba01f TM |
3528 | ext4_groupinfo_caches[cache_index] = cachep; |
3529 | ||
2892c15d ES |
3530 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
3531 | if (!cachep) { | |
9d8b9ec4 TT |
3532 | printk(KERN_EMERG |
3533 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
3534 | return -ENOMEM; |
3535 | } | |
3536 | ||
2892c15d ES |
3537 | return 0; |
3538 | } | |
3539 | ||
55cdd0af WJ |
3540 | static void ext4_discard_work(struct work_struct *work) |
3541 | { | |
3542 | struct ext4_sb_info *sbi = container_of(work, | |
3543 | struct ext4_sb_info, s_discard_work); | |
3544 | struct super_block *sb = sbi->s_sb; | |
3545 | struct ext4_free_data *fd, *nfd; | |
3546 | struct ext4_buddy e4b; | |
0f6bc579 | 3547 | LIST_HEAD(discard_list); |
55cdd0af WJ |
3548 | ext4_group_t grp, load_grp; |
3549 | int err = 0; | |
3550 | ||
55cdd0af WJ |
3551 | spin_lock(&sbi->s_md_lock); |
3552 | list_splice_init(&sbi->s_discard_list, &discard_list); | |
3553 | spin_unlock(&sbi->s_md_lock); | |
3554 | ||
3555 | load_grp = UINT_MAX; | |
3556 | list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { | |
3557 | /* | |
5036ab8d WJ |
3558 | * If filesystem is umounting or no memory or suffering |
3559 | * from no space, give up the discard | |
55cdd0af | 3560 | */ |
5036ab8d WJ |
3561 | if ((sb->s_flags & SB_ACTIVE) && !err && |
3562 | !atomic_read(&sbi->s_retry_alloc_pending)) { | |
55cdd0af WJ |
3563 | grp = fd->efd_group; |
3564 | if (grp != load_grp) { | |
3565 | if (load_grp != UINT_MAX) | |
3566 | ext4_mb_unload_buddy(&e4b); | |
3567 | ||
3568 | err = ext4_mb_load_buddy(sb, grp, &e4b); | |
3569 | if (err) { | |
3570 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3571 | load_grp = UINT_MAX; | |
3572 | continue; | |
3573 | } else { | |
3574 | load_grp = grp; | |
3575 | } | |
3576 | } | |
3577 | ||
3578 | ext4_lock_group(sb, grp); | |
3579 | ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, | |
3580 | fd->efd_start_cluster + fd->efd_count - 1, 1); | |
3581 | ext4_unlock_group(sb, grp); | |
3582 | } | |
3583 | kmem_cache_free(ext4_free_data_cachep, fd); | |
3584 | } | |
3585 | ||
3586 | if (load_grp != UINT_MAX) | |
3587 | ext4_mb_unload_buddy(&e4b); | |
3588 | } | |
3589 | ||
9d99012f | 3590 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
3591 | { |
3592 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 3593 | unsigned i, j; |
935244cd | 3594 | unsigned offset, offset_incr; |
c9de560d | 3595 | unsigned max; |
74767c5a | 3596 | int ret; |
c9de560d | 3597 | |
4b68f6df | 3598 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
3599 | |
3600 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
3601 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
3602 | ret = -ENOMEM; |
3603 | goto out; | |
c9de560d | 3604 | } |
ff7ef329 | 3605 | |
4b68f6df | 3606 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
3607 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
3608 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
3609 | ret = -ENOMEM; |
3610 | goto out; | |
3611 | } | |
3612 | ||
2892c15d ES |
3613 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
3614 | if (ret < 0) | |
3615 | goto out; | |
c9de560d AT |
3616 | |
3617 | /* order 0 is regular bitmap */ | |
3618 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
3619 | sbi->s_mb_offsets[0] = 0; | |
3620 | ||
3621 | i = 1; | |
3622 | offset = 0; | |
935244cd | 3623 | offset_incr = 1 << (sb->s_blocksize_bits - 1); |
c9de560d AT |
3624 | max = sb->s_blocksize << 2; |
3625 | do { | |
3626 | sbi->s_mb_offsets[i] = offset; | |
3627 | sbi->s_mb_maxs[i] = max; | |
935244cd NS |
3628 | offset += offset_incr; |
3629 | offset_incr = offset_incr >> 1; | |
c9de560d AT |
3630 | max = max >> 1; |
3631 | i++; | |
4b68f6df HS |
3632 | } while (i < MB_NUM_ORDERS(sb)); |
3633 | ||
83e80a6e JK |
3634 | sbi->s_mb_avg_fragment_size = |
3635 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3636 | GFP_KERNEL); | |
3637 | if (!sbi->s_mb_avg_fragment_size) { | |
3638 | ret = -ENOMEM; | |
3639 | goto out; | |
3640 | } | |
3641 | sbi->s_mb_avg_fragment_size_locks = | |
3642 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3643 | GFP_KERNEL); | |
3644 | if (!sbi->s_mb_avg_fragment_size_locks) { | |
3645 | ret = -ENOMEM; | |
3646 | goto out; | |
3647 | } | |
3648 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3649 | INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); | |
3650 | rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); | |
3651 | } | |
196e402a HS |
3652 | sbi->s_mb_largest_free_orders = |
3653 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), | |
3654 | GFP_KERNEL); | |
3655 | if (!sbi->s_mb_largest_free_orders) { | |
3656 | ret = -ENOMEM; | |
3657 | goto out; | |
3658 | } | |
3659 | sbi->s_mb_largest_free_orders_locks = | |
3660 | kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), | |
3661 | GFP_KERNEL); | |
3662 | if (!sbi->s_mb_largest_free_orders_locks) { | |
3663 | ret = -ENOMEM; | |
3664 | goto out; | |
3665 | } | |
3666 | for (i = 0; i < MB_NUM_ORDERS(sb); i++) { | |
3667 | INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); | |
3668 | rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); | |
3669 | } | |
c9de560d | 3670 | |
c9de560d | 3671 | spin_lock_init(&sbi->s_md_lock); |
d08854f5 | 3672 | sbi->s_mb_free_pending = 0; |
ce774e53 JH |
3673 | INIT_LIST_HEAD(&sbi->s_freed_data_list[0]); |
3674 | INIT_LIST_HEAD(&sbi->s_freed_data_list[1]); | |
55cdd0af WJ |
3675 | INIT_LIST_HEAD(&sbi->s_discard_list); |
3676 | INIT_WORK(&sbi->s_discard_work, ext4_discard_work); | |
5036ab8d | 3677 | atomic_set(&sbi->s_retry_alloc_pending, 0); |
c9de560d AT |
3678 | |
3679 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
3680 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
3681 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
3682 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
3683 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
f52f3d2b | 3684 | sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER; |
7e170922 | 3685 | |
27baebb8 TT |
3686 | /* |
3687 | * The default group preallocation is 512, which for 4k block | |
3688 | * sizes translates to 2 megabytes. However for bigalloc file | |
3689 | * systems, this is probably too big (i.e, if the cluster size | |
3690 | * is 1 megabyte, then group preallocation size becomes half a | |
3691 | * gigabyte!). As a default, we will keep a two megabyte | |
3692 | * group pralloc size for cluster sizes up to 64k, and after | |
3693 | * that, we will force a minimum group preallocation size of | |
3694 | * 32 clusters. This translates to 8 megs when the cluster | |
3695 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
3696 | * which seems reasonable as a default. | |
3697 | */ | |
3698 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
3699 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
3700 | /* |
3701 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
3702 | * to the lowest multiple of s_stripe which is bigger than | |
3703 | * the s_mb_group_prealloc as determined above. We want | |
3704 | * the preallocation size to be an exact multiple of the | |
3705 | * RAID stripe size so that preallocations don't fragment | |
3706 | * the stripes. | |
3707 | */ | |
3708 | if (sbi->s_stripe > 1) { | |
3709 | sbi->s_mb_group_prealloc = roundup( | |
c3defd99 | 3710 | sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe)); |
d7a1fee1 | 3711 | } |
c9de560d | 3712 | |
730c213c | 3713 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 3714 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 3715 | ret = -ENOMEM; |
029b10c5 | 3716 | goto out; |
c9de560d | 3717 | } |
730c213c | 3718 | for_each_possible_cpu(i) { |
c9de560d | 3719 | struct ext4_locality_group *lg; |
730c213c | 3720 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 3721 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
3722 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
3723 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
3724 | spin_lock_init(&lg->lg_prealloc_lock); |
3725 | } | |
3726 | ||
10f0d2a5 | 3727 | if (bdev_nonrot(sb->s_bdev)) |
196e402a HS |
3728 | sbi->s_mb_max_linear_groups = 0; |
3729 | else | |
3730 | sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; | |
79a77c5a YJ |
3731 | /* init file for buddy data */ |
3732 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
3733 | if (ret != 0) |
3734 | goto out_free_locality_groups; | |
79a77c5a | 3735 | |
7aa0baea TM |
3736 | return 0; |
3737 | ||
3738 | out_free_locality_groups: | |
3739 | free_percpu(sbi->s_locality_groups); | |
3740 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 3741 | out: |
83e80a6e JK |
3742 | kfree(sbi->s_mb_avg_fragment_size); |
3743 | kfree(sbi->s_mb_avg_fragment_size_locks); | |
196e402a HS |
3744 | kfree(sbi->s_mb_largest_free_orders); |
3745 | kfree(sbi->s_mb_largest_free_orders_locks); | |
7aa0baea TM |
3746 | kfree(sbi->s_mb_offsets); |
3747 | sbi->s_mb_offsets = NULL; | |
3748 | kfree(sbi->s_mb_maxs); | |
3749 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 3750 | return ret; |
c9de560d AT |
3751 | } |
3752 | ||
955ce5f5 | 3753 | /* need to called with the ext4 group lock held */ |
d3df1453 | 3754 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
c9de560d AT |
3755 | { |
3756 | struct ext4_prealloc_space *pa; | |
3757 | struct list_head *cur, *tmp; | |
3758 | int count = 0; | |
3759 | ||
3760 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
3761 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3762 | list_del(&pa->pa_group_list); | |
3763 | count++; | |
688f05a0 | 3764 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d | 3765 | } |
d3df1453 | 3766 | return count; |
c9de560d AT |
3767 | } |
3768 | ||
90817717 | 3769 | void ext4_mb_release(struct super_block *sb) |
c9de560d | 3770 | { |
8df9675f | 3771 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
3772 | ext4_group_t i; |
3773 | int num_meta_group_infos; | |
df3da4ea | 3774 | struct ext4_group_info *grinfo, ***group_info; |
c9de560d | 3775 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
fb1813f4 | 3776 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
d3df1453 | 3777 | int count; |
c9de560d | 3778 | |
55cdd0af WJ |
3779 | if (test_opt(sb, DISCARD)) { |
3780 | /* | |
3781 | * wait the discard work to drain all of ext4_free_data | |
3782 | */ | |
3783 | flush_work(&sbi->s_discard_work); | |
3784 | WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); | |
3785 | } | |
3786 | ||
c9de560d | 3787 | if (sbi->s_group_info) { |
8df9675f | 3788 | for (i = 0; i < ngroups; i++) { |
4b99faa2 | 3789 | cond_resched(); |
c9de560d | 3790 | grinfo = ext4_get_group_info(sb, i); |
5354b2af TT |
3791 | if (!grinfo) |
3792 | continue; | |
a3450215 | 3793 | mb_group_bb_bitmap_free(grinfo); |
c9de560d | 3794 | ext4_lock_group(sb, i); |
d3df1453 RH |
3795 | count = ext4_mb_cleanup_pa(grinfo); |
3796 | if (count) | |
3797 | mb_debug(sb, "mballoc: %d PAs left\n", | |
3798 | count); | |
c9de560d | 3799 | ext4_unlock_group(sb, i); |
fb1813f4 | 3800 | kmem_cache_free(cachep, grinfo); |
c9de560d | 3801 | } |
8df9675f | 3802 | num_meta_group_infos = (ngroups + |
c9de560d AT |
3803 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
3804 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
df3da4ea SJS |
3805 | rcu_read_lock(); |
3806 | group_info = rcu_dereference(sbi->s_group_info); | |
c9de560d | 3807 | for (i = 0; i < num_meta_group_infos; i++) |
df3da4ea SJS |
3808 | kfree(group_info[i]); |
3809 | kvfree(group_info); | |
3810 | rcu_read_unlock(); | |
c9de560d | 3811 | } |
83e80a6e JK |
3812 | kfree(sbi->s_mb_avg_fragment_size); |
3813 | kfree(sbi->s_mb_avg_fragment_size_locks); | |
196e402a HS |
3814 | kfree(sbi->s_mb_largest_free_orders); |
3815 | kfree(sbi->s_mb_largest_free_orders_locks); | |
c9de560d AT |
3816 | kfree(sbi->s_mb_offsets); |
3817 | kfree(sbi->s_mb_maxs); | |
bfcba2d0 | 3818 | iput(sbi->s_buddy_cache); |
c9de560d | 3819 | if (sbi->s_mb_stats) { |
9d8b9ec4 TT |
3820 | ext4_msg(sb, KERN_INFO, |
3821 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
3822 | atomic_read(&sbi->s_bal_allocated), |
3823 | atomic_read(&sbi->s_bal_reqs), | |
3824 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 | 3825 | ext4_msg(sb, KERN_INFO, |
a6c75eaf | 3826 | "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " |
9d8b9ec4 | 3827 | "%u 2^N hits, %u breaks, %u lost", |
c9de560d | 3828 | atomic_read(&sbi->s_bal_ex_scanned), |
a6c75eaf | 3829 | atomic_read(&sbi->s_bal_groups_scanned), |
c9de560d AT |
3830 | atomic_read(&sbi->s_bal_goals), |
3831 | atomic_read(&sbi->s_bal_2orders), | |
3832 | atomic_read(&sbi->s_bal_breaks), | |
3833 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 | 3834 | ext4_msg(sb, KERN_INFO, |
67d25186 HS |
3835 | "mballoc: %u generated and it took %llu", |
3836 | atomic_read(&sbi->s_mb_buddies_generated), | |
3837 | atomic64_read(&sbi->s_mb_generation_time)); | |
9d8b9ec4 TT |
3838 | ext4_msg(sb, KERN_INFO, |
3839 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
3840 | atomic_read(&sbi->s_mb_preallocated), |
3841 | atomic_read(&sbi->s_mb_discarded)); | |
3842 | } | |
3843 | ||
730c213c | 3844 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
3845 | } |
3846 | ||
77ca6cdf | 3847 | static inline int ext4_issue_discard(struct super_block *sb, |
0efcd739 | 3848 | ext4_group_t block_group, ext4_grpblk_t cluster, int count) |
5c521830 | 3849 | { |
5c521830 JZ |
3850 | ext4_fsblk_t discard_block; |
3851 | ||
84130193 TT |
3852 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
3853 | ext4_group_first_block_no(sb, block_group)); | |
3854 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
3855 | trace_ext4_discard_blocks(sb, |
3856 | (unsigned long long) discard_block, count); | |
0efcd739 WH |
3857 | |
3858 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); | |
5c521830 JZ |
3859 | } |
3860 | ||
a0154344 DJ |
3861 | static void ext4_free_data_in_buddy(struct super_block *sb, |
3862 | struct ext4_free_data *entry) | |
c9de560d | 3863 | { |
c9de560d | 3864 | struct ext4_buddy e4b; |
c894058d | 3865 | struct ext4_group_info *db; |
c7f2bafa | 3866 | int err, count = 0; |
c9de560d | 3867 | |
d3df1453 | 3868 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
18aadd47 | 3869 | entry->efd_count, entry->efd_group, entry); |
c9de560d | 3870 | |
18aadd47 BJ |
3871 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
3872 | /* we expect to find existing buddy because it's pinned */ | |
3873 | BUG_ON(err != 0); | |
b90f6870 | 3874 | |
d08854f5 TT |
3875 | spin_lock(&EXT4_SB(sb)->s_md_lock); |
3876 | EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; | |
3877 | spin_unlock(&EXT4_SB(sb)->s_md_lock); | |
c9de560d | 3878 | |
18aadd47 BJ |
3879 | db = e4b.bd_info; |
3880 | /* there are blocks to put in buddy to make them really free */ | |
3881 | count += entry->efd_count; | |
18aadd47 BJ |
3882 | ext4_lock_group(sb, entry->efd_group); |
3883 | /* Take it out of per group rb tree */ | |
3884 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
3885 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 3886 | |
18aadd47 BJ |
3887 | /* |
3888 | * Clear the trimmed flag for the group so that the next | |
3889 | * ext4_trim_fs can trim it. | |
3890 | * If the volume is mounted with -o discard, online discard | |
3891 | * is supported and the free blocks will be trimmed online. | |
3892 | */ | |
3893 | if (!test_opt(sb, DISCARD)) | |
3894 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 3895 | |
18aadd47 BJ |
3896 | if (!db->bb_free_root.rb_node) { |
3897 | /* No more items in the per group rb tree | |
3898 | * balance refcounts from ext4_mb_free_metadata() | |
3899 | */ | |
5eea586b | 3900 | folio_put(e4b.bd_buddy_folio); |
99b150d8 | 3901 | folio_put(e4b.bd_bitmap_folio); |
3e624fc7 | 3902 | } |
18aadd47 | 3903 | ext4_unlock_group(sb, entry->efd_group); |
18aadd47 | 3904 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 3905 | |
c7f2bafa | 3906 | mb_debug(sb, "freed %d blocks in 1 structures\n", count); |
c9de560d AT |
3907 | } |
3908 | ||
a0154344 DJ |
3909 | /* |
3910 | * This function is called by the jbd2 layer once the commit has finished, | |
3911 | * so we know we can free the blocks that were released with that commit. | |
3912 | */ | |
3913 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) | |
3914 | { | |
3915 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3916 | struct ext4_free_data *entry, *tmp; | |
0f6bc579 | 3917 | LIST_HEAD(freed_data_list); |
ce774e53 | 3918 | struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1]; |
55cdd0af | 3919 | bool wake; |
a0154344 | 3920 | |
ce774e53 | 3921 | list_replace_init(s_freed_head, &freed_data_list); |
a0154344 | 3922 | |
55cdd0af WJ |
3923 | list_for_each_entry(entry, &freed_data_list, efd_list) |
3924 | ext4_free_data_in_buddy(sb, entry); | |
a0154344 | 3925 | |
55cdd0af WJ |
3926 | if (test_opt(sb, DISCARD)) { |
3927 | spin_lock(&sbi->s_md_lock); | |
3928 | wake = list_empty(&sbi->s_discard_list); | |
3929 | list_splice_tail(&freed_data_list, &sbi->s_discard_list); | |
3930 | spin_unlock(&sbi->s_md_lock); | |
3931 | if (wake) | |
3932 | queue_work(system_unbound_wq, &sbi->s_discard_work); | |
3933 | } else { | |
3934 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) | |
3935 | kmem_cache_free(ext4_free_data_cachep, entry); | |
a0154344 | 3936 | } |
a0154344 DJ |
3937 | } |
3938 | ||
5dabfc78 | 3939 | int __init ext4_init_mballoc(void) |
c9de560d | 3940 | { |
16828088 TT |
3941 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
3942 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d | 3943 | if (ext4_pspace_cachep == NULL) |
f283529a | 3944 | goto out; |
c9de560d | 3945 | |
16828088 TT |
3946 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
3947 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3948 | if (ext4_ac_cachep == NULL) |
3949 | goto out_pa_free; | |
c894058d | 3950 | |
18aadd47 BJ |
3951 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
3952 | SLAB_RECLAIM_ACCOUNT); | |
f283529a RH |
3953 | if (ext4_free_data_cachep == NULL) |
3954 | goto out_ac_free; | |
3955 | ||
c9de560d | 3956 | return 0; |
f283529a RH |
3957 | |
3958 | out_ac_free: | |
3959 | kmem_cache_destroy(ext4_ac_cachep); | |
3960 | out_pa_free: | |
3961 | kmem_cache_destroy(ext4_pspace_cachep); | |
3962 | out: | |
3963 | return -ENOMEM; | |
c9de560d AT |
3964 | } |
3965 | ||
5dabfc78 | 3966 | void ext4_exit_mballoc(void) |
c9de560d | 3967 | { |
60e6679e | 3968 | /* |
3e03f9ca JDB |
3969 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
3970 | * before destroying the slab cache. | |
3971 | */ | |
3972 | rcu_barrier(); | |
c9de560d | 3973 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 3974 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 3975 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 3976 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
3977 | } |
3978 | ||
c431d386 KS |
3979 | #define EXT4_MB_BITMAP_MARKED_CHECK 0x0001 |
3980 | #define EXT4_MB_SYNC_UPDATE 0x0002 | |
f9e2d95a | 3981 | static int |
c431d386 KS |
3982 | ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state, |
3983 | ext4_group_t group, ext4_grpblk_t blkoff, | |
3984 | ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed) | |
f9e2d95a KS |
3985 | { |
3986 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
3987 | struct buffer_head *bitmap_bh = NULL; | |
3988 | struct ext4_group_desc *gdp; | |
3989 | struct buffer_head *gdp_bh; | |
3990 | int err; | |
c431d386 | 3991 | unsigned int i, already, changed = len; |
bdefd689 KS |
3992 | |
3993 | KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context, | |
3994 | handle, sb, state, group, blkoff, len, | |
3995 | flags, ret_changed); | |
f9e2d95a | 3996 | |
c431d386 KS |
3997 | if (ret_changed) |
3998 | *ret_changed = 0; | |
f9e2d95a KS |
3999 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
4000 | if (IS_ERR(bitmap_bh)) | |
4001 | return PTR_ERR(bitmap_bh); | |
4002 | ||
c431d386 KS |
4003 | if (handle) { |
4004 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4005 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh, | |
4006 | EXT4_JTR_NONE); | |
4007 | if (err) | |
4008 | goto out_err; | |
4009 | } | |
4010 | ||
f9e2d95a KS |
4011 | err = -EIO; |
4012 | gdp = ext4_get_group_desc(sb, group, &gdp_bh); | |
4013 | if (!gdp) | |
4014 | goto out_err; | |
4015 | ||
c431d386 KS |
4016 | if (handle) { |
4017 | BUFFER_TRACE(gdp_bh, "get_write_access"); | |
4018 | err = ext4_journal_get_write_access(handle, sb, gdp_bh, | |
4019 | EXT4_JTR_NONE); | |
4020 | if (err) | |
4021 | goto out_err; | |
4022 | } | |
4023 | ||
f9e2d95a KS |
4024 | ext4_lock_group(sb, group); |
4025 | if (ext4_has_group_desc_csum(sb) && | |
4026 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { | |
4027 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
4028 | ext4_free_group_clusters_set(sb, gdp, | |
4029 | ext4_free_clusters_after_init(sb, group, gdp)); | |
4030 | } | |
4031 | ||
c431d386 KS |
4032 | if (flags & EXT4_MB_BITMAP_MARKED_CHECK) { |
4033 | already = 0; | |
4034 | for (i = 0; i < len; i++) | |
4035 | if (mb_test_bit(blkoff + i, bitmap_bh->b_data) == | |
4036 | state) | |
4037 | already++; | |
4038 | changed = len - already; | |
4039 | } | |
f9e2d95a KS |
4040 | |
4041 | if (state) { | |
4042 | mb_set_bits(bitmap_bh->b_data, blkoff, len); | |
4043 | ext4_free_group_clusters_set(sb, gdp, | |
4044 | ext4_free_group_clusters(sb, gdp) - changed); | |
4045 | } else { | |
4046 | mb_clear_bits(bitmap_bh->b_data, blkoff, len); | |
4047 | ext4_free_group_clusters_set(sb, gdp, | |
4048 | ext4_free_group_clusters(sb, gdp) + changed); | |
4049 | } | |
4050 | ||
4051 | ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); | |
4052 | ext4_group_desc_csum_set(sb, group, gdp); | |
4053 | ext4_unlock_group(sb, group); | |
c431d386 KS |
4054 | if (ret_changed) |
4055 | *ret_changed = changed; | |
f9e2d95a KS |
4056 | |
4057 | if (sbi->s_log_groups_per_flex) { | |
4058 | ext4_group_t flex_group = ext4_flex_group(sbi, group); | |
4059 | struct flex_groups *fg = sbi_array_rcu_deref(sbi, | |
4060 | s_flex_groups, flex_group); | |
4061 | ||
4062 | if (state) | |
4063 | atomic64_sub(changed, &fg->free_clusters); | |
4064 | else | |
4065 | atomic64_add(changed, &fg->free_clusters); | |
4066 | } | |
4067 | ||
c431d386 | 4068 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
f9e2d95a KS |
4069 | if (err) |
4070 | goto out_err; | |
c431d386 | 4071 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
f9e2d95a KS |
4072 | if (err) |
4073 | goto out_err; | |
4074 | ||
c431d386 KS |
4075 | if (flags & EXT4_MB_SYNC_UPDATE) { |
4076 | sync_dirty_buffer(bitmap_bh); | |
4077 | sync_dirty_buffer(gdp_bh); | |
4078 | } | |
f9e2d95a KS |
4079 | |
4080 | out_err: | |
4081 | brelse(bitmap_bh); | |
4082 | return err; | |
4083 | } | |
c9de560d AT |
4084 | |
4085 | /* | |
73b2c716 | 4086 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
4087 | * Returns 0 if success or error code |
4088 | */ | |
4ddfef7b ES |
4089 | static noinline_for_stack int |
4090 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 4091 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d | 4092 | { |
c9de560d | 4093 | struct ext4_group_desc *gdp; |
c9de560d AT |
4094 | struct ext4_sb_info *sbi; |
4095 | struct super_block *sb; | |
4096 | ext4_fsblk_t block; | |
519deca0 | 4097 | int err, len; |
2f94711b KS |
4098 | int flags = 0; |
4099 | ext4_grpblk_t changed; | |
c9de560d AT |
4100 | |
4101 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
4102 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
4103 | ||
4104 | sb = ac->ac_sb; | |
4105 | sbi = EXT4_SB(sb); | |
c9de560d | 4106 | |
2f94711b | 4107 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL); |
c9de560d | 4108 | if (!gdp) |
2f94711b | 4109 | return -EIO; |
a9df9a49 | 4110 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 4111 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 4112 | |
bda00de7 | 4113 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
53accfa9 | 4114 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
ce9f24cc | 4115 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
12062ddd | 4116 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 4117 | "fs metadata", block, block+len); |
519deca0 | 4118 | /* File system mounted not to panic on error |
554a5ccc | 4119 | * Fix the bitmap and return EFSCORRUPTED |
519deca0 AK |
4120 | * We leak some of the blocks here. |
4121 | */ | |
2f94711b KS |
4122 | err = ext4_mb_mark_context(handle, sb, true, |
4123 | ac->ac_b_ex.fe_group, | |
4124 | ac->ac_b_ex.fe_start, | |
4125 | ac->ac_b_ex.fe_len, | |
4126 | 0, NULL); | |
519deca0 | 4127 | if (!err) |
554a5ccc | 4128 | err = -EFSCORRUPTED; |
2f94711b | 4129 | return err; |
c9de560d | 4130 | } |
955ce5f5 | 4131 | |
c9de560d | 4132 | #ifdef AGGRESSIVE_CHECK |
2f94711b | 4133 | flags |= EXT4_MB_BITMAP_MARKED_CHECK; |
c9de560d | 4134 | #endif |
2f94711b KS |
4135 | err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group, |
4136 | ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len, | |
4137 | flags, &changed); | |
4138 | ||
4139 | if (err && changed == 0) | |
4140 | return err; | |
955ce5f5 | 4141 | |
2f94711b KS |
4142 | #ifdef AGGRESSIVE_CHECK |
4143 | BUG_ON(changed != ac->ac_b_ex.fe_len); | |
4144 | #endif | |
57042651 | 4145 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 4146 | /* |
6bc6e63f | 4147 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 4148 | */ |
6bc6e63f AK |
4149 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
4150 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
4151 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
4152 | reserv_clstrs); | |
c9de560d | 4153 | |
c9de560d AT |
4154 | return err; |
4155 | } | |
4156 | ||
8016e29f HS |
4157 | /* |
4158 | * Idempotent helper for Ext4 fast commit replay path to set the state of | |
4159 | * blocks in bitmaps and update counters. | |
4160 | */ | |
4161 | void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, | |
d2f7cf40 | 4162 | int len, bool state) |
8016e29f | 4163 | { |
8016e29f HS |
4164 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
4165 | ext4_group_t group; | |
4166 | ext4_grpblk_t blkoff; | |
f9e2d95a KS |
4167 | int err = 0; |
4168 | unsigned int clen, thisgrp_len; | |
8016e29f | 4169 | |
bfdc502a RH |
4170 | while (len > 0) { |
4171 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
8016e29f | 4172 | |
bfdc502a RH |
4173 | /* |
4174 | * Check to see if we are freeing blocks across a group | |
4175 | * boundary. | |
4176 | * In case of flex_bg, this can happen that (block, len) may | |
4177 | * span across more than one group. In that case we need to | |
4178 | * get the corresponding group metadata to work with. | |
4179 | * For this we have goto again loop. | |
4180 | */ | |
4181 | thisgrp_len = min_t(unsigned int, (unsigned int)len, | |
4182 | EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); | |
4183 | clen = EXT4_NUM_B2C(sbi, thisgrp_len); | |
4184 | ||
8c91c579 RH |
4185 | if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { |
4186 | ext4_error(sb, "Marking blocks in system zone - " | |
4187 | "Block = %llu, len = %u", | |
4188 | block, thisgrp_len); | |
bfdc502a RH |
4189 | break; |
4190 | } | |
8016e29f | 4191 | |
c431d386 KS |
4192 | err = ext4_mb_mark_context(NULL, sb, state, |
4193 | group, blkoff, clen, | |
4194 | EXT4_MB_BITMAP_MARKED_CHECK | | |
4195 | EXT4_MB_SYNC_UPDATE, | |
4196 | NULL); | |
bfdc502a RH |
4197 | if (err) |
4198 | break; | |
4199 | ||
4200 | block += thisgrp_len; | |
4201 | len -= thisgrp_len; | |
bfdc502a | 4202 | BUG_ON(len < 0); |
8016e29f | 4203 | } |
8016e29f HS |
4204 | } |
4205 | ||
c9de560d AT |
4206 | /* |
4207 | * here we normalize request for locality group | |
d7a1fee1 DE |
4208 | * Group request are normalized to s_mb_group_prealloc, which goes to |
4209 | * s_strip if we set the same via mount option. | |
4210 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 4211 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
4212 | * |
4213 | * XXX: should we try to preallocate more than the group has now? | |
4214 | */ | |
4215 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
4216 | { | |
4217 | struct super_block *sb = ac->ac_sb; | |
4218 | struct ext4_locality_group *lg = ac->ac_lg; | |
4219 | ||
4220 | BUG_ON(lg == NULL); | |
d7a1fee1 | 4221 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
d3df1453 | 4222 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
c9de560d AT |
4223 | } |
4224 | ||
38727786 OM |
4225 | /* |
4226 | * This function returns the next element to look at during inode | |
4227 | * PA rbtree walk. We assume that we have held the inode PA rbtree lock | |
4228 | * (ei->i_prealloc_lock) | |
4229 | * | |
4230 | * new_start The start of the range we want to compare | |
4231 | * cur_start The existing start that we are comparing against | |
4232 | * node The node of the rb_tree | |
4233 | */ | |
4234 | static inline struct rb_node* | |
4235 | ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node) | |
4236 | { | |
4237 | if (new_start < cur_start) | |
4238 | return node->rb_left; | |
4239 | else | |
4240 | return node->rb_right; | |
4241 | } | |
4242 | ||
7692094a OM |
4243 | static inline void |
4244 | ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac, | |
bedc5d34 | 4245 | ext4_lblk_t start, loff_t end) |
7692094a OM |
4246 | { |
4247 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4248 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
4249 | struct ext4_prealloc_space *tmp_pa; | |
bedc5d34 BL |
4250 | ext4_lblk_t tmp_pa_start; |
4251 | loff_t tmp_pa_end; | |
38727786 | 4252 | struct rb_node *iter; |
7692094a | 4253 | |
38727786 OM |
4254 | read_lock(&ei->i_prealloc_lock); |
4255 | for (iter = ei->i_prealloc_node.rb_node; iter; | |
4256 | iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) { | |
4257 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4258 | pa_node.inode_node); | |
4259 | tmp_pa_start = tmp_pa->pa_lstart; | |
bedc5d34 | 4260 | tmp_pa_end = pa_logical_end(sbi, tmp_pa); |
7692094a | 4261 | |
38727786 OM |
4262 | spin_lock(&tmp_pa->pa_lock); |
4263 | if (tmp_pa->pa_deleted == 0) | |
7692094a | 4264 | BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start)); |
7692094a OM |
4265 | spin_unlock(&tmp_pa->pa_lock); |
4266 | } | |
38727786 | 4267 | read_unlock(&ei->i_prealloc_lock); |
7692094a OM |
4268 | } |
4269 | ||
0830344c OM |
4270 | /* |
4271 | * Given an allocation context "ac" and a range "start", "end", check | |
4272 | * and adjust boundaries if the range overlaps with any of the existing | |
4273 | * preallocatoins stored in the corresponding inode of the allocation context. | |
4274 | * | |
38727786 | 4275 | * Parameters: |
0830344c OM |
4276 | * ac allocation context |
4277 | * start start of the new range | |
4278 | * end end of the new range | |
4279 | */ | |
4280 | static inline void | |
4281 | ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac, | |
bedc5d34 | 4282 | ext4_lblk_t *start, loff_t *end) |
0830344c OM |
4283 | { |
4284 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); | |
4285 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
38727786 OM |
4286 | struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL; |
4287 | struct rb_node *iter; | |
bedc5d34 BL |
4288 | ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1; |
4289 | loff_t new_end, tmp_pa_end, left_pa_end = -1; | |
0830344c OM |
4290 | |
4291 | new_start = *start; | |
4292 | new_end = *end; | |
4293 | ||
38727786 OM |
4294 | /* |
4295 | * Adjust the normalized range so that it doesn't overlap with any | |
4296 | * existing preallocated blocks(PAs). Make sure to hold the rbtree lock | |
4297 | * so it doesn't change underneath us. | |
4298 | */ | |
4299 | read_lock(&ei->i_prealloc_lock); | |
4300 | ||
4301 | /* Step 1: find any one immediate neighboring PA of the normalized range */ | |
4302 | for (iter = ei->i_prealloc_node.rb_node; iter; | |
4303 | iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, | |
4304 | tmp_pa_start, iter)) { | |
4305 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4306 | pa_node.inode_node); | |
0830344c | 4307 | tmp_pa_start = tmp_pa->pa_lstart; |
bedc5d34 | 4308 | tmp_pa_end = pa_logical_end(sbi, tmp_pa); |
0830344c OM |
4309 | |
4310 | /* PA must not overlap original request */ | |
38727786 OM |
4311 | spin_lock(&tmp_pa->pa_lock); |
4312 | if (tmp_pa->pa_deleted == 0) | |
4313 | BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end || | |
4314 | ac->ac_o_ex.fe_logical < tmp_pa_start)); | |
4315 | spin_unlock(&tmp_pa->pa_lock); | |
4316 | } | |
4317 | ||
4318 | /* | |
4319 | * Step 2: check if the found PA is left or right neighbor and | |
4320 | * get the other neighbor | |
4321 | */ | |
4322 | if (tmp_pa) { | |
4323 | if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) { | |
4324 | struct rb_node *tmp; | |
4325 | ||
4326 | left_pa = tmp_pa; | |
4327 | tmp = rb_next(&left_pa->pa_node.inode_node); | |
4328 | if (tmp) { | |
4329 | right_pa = rb_entry(tmp, | |
4330 | struct ext4_prealloc_space, | |
4331 | pa_node.inode_node); | |
4332 | } | |
4333 | } else { | |
4334 | struct rb_node *tmp; | |
4335 | ||
4336 | right_pa = tmp_pa; | |
4337 | tmp = rb_prev(&right_pa->pa_node.inode_node); | |
4338 | if (tmp) { | |
4339 | left_pa = rb_entry(tmp, | |
4340 | struct ext4_prealloc_space, | |
4341 | pa_node.inode_node); | |
4342 | } | |
4343 | } | |
4344 | } | |
4345 | ||
4346 | /* Step 3: get the non deleted neighbors */ | |
4347 | if (left_pa) { | |
4348 | for (iter = &left_pa->pa_node.inode_node;; | |
4349 | iter = rb_prev(iter)) { | |
4350 | if (!iter) { | |
4351 | left_pa = NULL; | |
4352 | break; | |
4353 | } | |
0830344c | 4354 | |
38727786 OM |
4355 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
4356 | pa_node.inode_node); | |
4357 | left_pa = tmp_pa; | |
4358 | spin_lock(&tmp_pa->pa_lock); | |
4359 | if (tmp_pa->pa_deleted == 0) { | |
4360 | spin_unlock(&tmp_pa->pa_lock); | |
4361 | break; | |
4362 | } | |
0830344c | 4363 | spin_unlock(&tmp_pa->pa_lock); |
0830344c | 4364 | } |
38727786 OM |
4365 | } |
4366 | ||
4367 | if (right_pa) { | |
4368 | for (iter = &right_pa->pa_node.inode_node;; | |
4369 | iter = rb_next(iter)) { | |
4370 | if (!iter) { | |
4371 | right_pa = NULL; | |
4372 | break; | |
4373 | } | |
4374 | ||
4375 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4376 | pa_node.inode_node); | |
4377 | right_pa = tmp_pa; | |
4378 | spin_lock(&tmp_pa->pa_lock); | |
4379 | if (tmp_pa->pa_deleted == 0) { | |
4380 | spin_unlock(&tmp_pa->pa_lock); | |
4381 | break; | |
4382 | } | |
4383 | spin_unlock(&tmp_pa->pa_lock); | |
0830344c | 4384 | } |
0830344c | 4385 | } |
38727786 OM |
4386 | |
4387 | if (left_pa) { | |
bedc5d34 | 4388 | left_pa_end = pa_logical_end(sbi, left_pa); |
38727786 OM |
4389 | BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical); |
4390 | } | |
4391 | ||
4392 | if (right_pa) { | |
4393 | right_pa_start = right_pa->pa_lstart; | |
4394 | BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical); | |
4395 | } | |
4396 | ||
4397 | /* Step 4: trim our normalized range to not overlap with the neighbors */ | |
4398 | if (left_pa) { | |
4399 | if (left_pa_end > new_start) | |
4400 | new_start = left_pa_end; | |
4401 | } | |
4402 | ||
4403 | if (right_pa) { | |
4404 | if (right_pa_start < new_end) | |
4405 | new_end = right_pa_start; | |
4406 | } | |
4407 | read_unlock(&ei->i_prealloc_lock); | |
0830344c OM |
4408 | |
4409 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
4410 | ext4_mb_pa_assert_overlap(ac, new_start, new_end); | |
4411 | ||
4412 | *start = new_start; | |
4413 | *end = new_end; | |
4414 | } | |
4415 | ||
c9de560d AT |
4416 | /* |
4417 | * Normalization means making request better in terms of | |
4418 | * size and alignment | |
4419 | */ | |
4ddfef7b ES |
4420 | static noinline_for_stack void |
4421 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
4422 | struct ext4_allocation_request *ar) |
4423 | { | |
53accfa9 | 4424 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
b07ffe69 | 4425 | struct ext4_super_block *es = sbi->s_es; |
c9de560d | 4426 | int bsbits, max; |
bedc5d34 | 4427 | loff_t size, start_off, end; |
1592d2c5 | 4428 | loff_t orig_size __maybe_unused; |
5a0790c2 | 4429 | ext4_lblk_t start; |
c9de560d AT |
4430 | |
4431 | /* do normalize only data requests, metadata requests | |
4432 | do not need preallocation */ | |
4433 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4434 | return; | |
4435 | ||
4436 | /* sometime caller may want exact blocks */ | |
4437 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
4438 | return; | |
4439 | ||
4440 | /* caller may indicate that preallocation isn't | |
4441 | * required (it's a tail, for example) */ | |
4442 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
4443 | return; | |
4444 | ||
4445 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
4446 | ext4_mb_normalize_group_request(ac); | |
4447 | return ; | |
4448 | } | |
4449 | ||
4450 | bsbits = ac->ac_sb->s_blocksize_bits; | |
4451 | ||
4452 | /* first, let's learn actual file size | |
4453 | * given current request is allocated */ | |
43bbddc0 | 4454 | size = extent_logical_end(sbi, &ac->ac_o_ex); |
c9de560d AT |
4455 | size = size << bsbits; |
4456 | if (size < i_size_read(ac->ac_inode)) | |
4457 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 4458 | orig_size = size; |
c9de560d | 4459 | |
1930479c VC |
4460 | /* max size of free chunks */ |
4461 | max = 2 << bsbits; | |
c9de560d | 4462 | |
1930479c VC |
4463 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
4464 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
4465 | |
4466 | /* first, try to predict filesize */ | |
4467 | /* XXX: should this table be tunable? */ | |
4468 | start_off = 0; | |
4469 | if (size <= 16 * 1024) { | |
4470 | size = 16 * 1024; | |
4471 | } else if (size <= 32 * 1024) { | |
4472 | size = 32 * 1024; | |
4473 | } else if (size <= 64 * 1024) { | |
4474 | size = 64 * 1024; | |
4475 | } else if (size <= 128 * 1024) { | |
4476 | size = 128 * 1024; | |
4477 | } else if (size <= 256 * 1024) { | |
4478 | size = 256 * 1024; | |
4479 | } else if (size <= 512 * 1024) { | |
4480 | size = 512 * 1024; | |
4481 | } else if (size <= 1024 * 1024) { | |
4482 | size = 1024 * 1024; | |
1930479c | 4483 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 4484 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
4485 | (21 - bsbits)) << 21; |
4486 | size = 2 * 1024 * 1024; | |
4487 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
4488 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4489 | (22 - bsbits)) << 22; | |
4490 | size = 4 * 1024 * 1024; | |
b3916da0 | 4491 | } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len), |
1930479c | 4492 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
4493 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
4494 | (23 - bsbits)) << 23; | |
4495 | size = 8 * 1024 * 1024; | |
4496 | } else { | |
b27b1535 | 4497 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
91a48aaf | 4498 | size = (loff_t) EXT4_C2B(sbi, |
b27b1535 | 4499 | ac->ac_o_ex.fe_len) << bsbits; |
c9de560d | 4500 | } |
5a0790c2 AK |
4501 | size = size >> bsbits; |
4502 | start = start_off >> bsbits; | |
c9de560d | 4503 | |
a08f789d BL |
4504 | /* |
4505 | * For tiny groups (smaller than 8MB) the chosen allocation | |
4506 | * alignment may be larger than group size. Make sure the | |
4507 | * alignment does not move allocation to a different group which | |
4508 | * makes mballoc fail assertions later. | |
4509 | */ | |
4510 | start = max(start, rounddown(ac->ac_o_ex.fe_logical, | |
4511 | (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); | |
4512 | ||
2dcf5fde BL |
4513 | /* avoid unnecessary preallocation that may trigger assertions */ |
4514 | if (start + size > EXT_MAX_BLOCKS) | |
4515 | size = EXT_MAX_BLOCKS - start; | |
4516 | ||
c9de560d AT |
4517 | /* don't cover already allocated blocks in selected range */ |
4518 | if (ar->pleft && start <= ar->lleft) { | |
4519 | size -= ar->lleft + 1 - start; | |
4520 | start = ar->lleft + 1; | |
4521 | } | |
4522 | if (ar->pright && start + size - 1 >= ar->lright) | |
4523 | size -= start + size - ar->lright; | |
4524 | ||
cd648b8a JK |
4525 | /* |
4526 | * Trim allocation request for filesystems with artificially small | |
4527 | * groups. | |
4528 | */ | |
4529 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) | |
4530 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); | |
4531 | ||
c9de560d AT |
4532 | end = start + size; |
4533 | ||
0830344c | 4534 | ext4_mb_pa_adjust_overlap(ac, &start, &end); |
c9de560d | 4535 | |
c9de560d AT |
4536 | size = end - start; |
4537 | ||
cf4ff938 BL |
4538 | /* |
4539 | * In this function "start" and "size" are normalized for better | |
4540 | * alignment and length such that we could preallocate more blocks. | |
4541 | * This normalization is done such that original request of | |
4542 | * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and | |
4543 | * "size" boundaries. | |
4544 | * (Note fe_len can be relaxed since FS block allocation API does not | |
4545 | * provide gurantee on number of contiguous blocks allocation since that | |
4546 | * depends upon free space left, etc). | |
4547 | * In case of inode pa, later we use the allocated blocks | |
1221b235 | 4548 | * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated |
cf4ff938 BL |
4549 | * range of goal/best blocks [start, size] to put it at the |
4550 | * ac_o_ex.fe_logical extent of this inode. | |
4551 | * (See ext4_mb_use_inode_pa() for more details) | |
4552 | */ | |
4553 | if (start + size <= ac->ac_o_ex.fe_logical || | |
c9de560d | 4554 | start > ac->ac_o_ex.fe_logical) { |
9d8b9ec4 TT |
4555 | ext4_msg(ac->ac_sb, KERN_ERR, |
4556 | "start %lu, size %lu, fe_logical %lu", | |
4557 | (unsigned long) start, (unsigned long) size, | |
4558 | (unsigned long) ac->ac_o_ex.fe_logical); | |
dfe076c1 | 4559 | BUG(); |
c9de560d | 4560 | } |
b5b60778 | 4561 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
4562 | |
4563 | /* now prepare goal request */ | |
4564 | ||
4565 | /* XXX: is it better to align blocks WRT to logical | |
4566 | * placement or satisfy big request as is */ | |
4567 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 4568 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
7e170922 | 4569 | ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; |
c9de560d AT |
4570 | |
4571 | /* define goal start in order to merge */ | |
b07ffe69 KS |
4572 | if (ar->pright && (ar->lright == (start + size)) && |
4573 | ar->pright >= size && | |
4574 | ar->pright - size >= le32_to_cpu(es->s_first_data_block)) { | |
c9de560d AT |
4575 | /* merge to the right */ |
4576 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
b07ffe69 KS |
4577 | &ac->ac_g_ex.fe_group, |
4578 | &ac->ac_g_ex.fe_start); | |
c9de560d AT |
4579 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; |
4580 | } | |
b07ffe69 KS |
4581 | if (ar->pleft && (ar->lleft + 1 == start) && |
4582 | ar->pleft + 1 < ext4_blocks_count(es)) { | |
c9de560d AT |
4583 | /* merge to the left */ |
4584 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
b07ffe69 KS |
4585 | &ac->ac_g_ex.fe_group, |
4586 | &ac->ac_g_ex.fe_start); | |
c9de560d AT |
4587 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; |
4588 | } | |
4589 | ||
d3df1453 RH |
4590 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
4591 | orig_size, start); | |
c9de560d AT |
4592 | } |
4593 | ||
4594 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
4595 | { | |
4596 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4597 | ||
a6c75eaf | 4598 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { |
c9de560d AT |
4599 | atomic_inc(&sbi->s_bal_reqs); |
4600 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 4601 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d | 4602 | atomic_inc(&sbi->s_bal_success); |
fdd9a009 | 4603 | |
c9de560d | 4604 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); |
fdd9a009 OM |
4605 | for (int i=0; i<EXT4_MB_NUM_CRS; i++) { |
4606 | atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]); | |
4607 | } | |
4608 | ||
a6c75eaf | 4609 | atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); |
c9de560d AT |
4610 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && |
4611 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
4612 | atomic_inc(&sbi->s_bal_goals); | |
7e170922 OM |
4613 | /* did we allocate as much as normalizer originally wanted? */ |
4614 | if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len) | |
3ef5d263 | 4615 | atomic_inc(&sbi->s_bal_len_goals); |
7e170922 | 4616 | |
c9de560d AT |
4617 | if (ac->ac_found > sbi->s_mb_max_to_scan) |
4618 | atomic_inc(&sbi->s_bal_breaks); | |
4619 | } | |
4620 | ||
296c355c TT |
4621 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
4622 | trace_ext4_mballoc_alloc(ac); | |
4623 | else | |
4624 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
4625 | } |
4626 | ||
b844167e CW |
4627 | /* |
4628 | * Called on failure; free up any blocks from the inode PA for this | |
4629 | * context. We don't need this for MB_GROUP_PA because we only change | |
4630 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
4631 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
4632 | */ | |
4633 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
4634 | { | |
4635 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
4636 | struct ext4_buddy e4b; |
4637 | int err; | |
b844167e | 4638 | |
86f0afd4 | 4639 | if (pa == NULL) { |
c99d1e6e TT |
4640 | if (ac->ac_f_ex.fe_len == 0) |
4641 | return; | |
86f0afd4 | 4642 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
19b8b035 TT |
4643 | if (WARN_RATELIMIT(err, |
4644 | "ext4: mb_load_buddy failed (%d)", err)) | |
86f0afd4 TT |
4645 | /* |
4646 | * This should never happen since we pin the | |
4647 | * pages in the ext4_allocation_context so | |
4648 | * ext4_mb_load_buddy() should never fail. | |
4649 | */ | |
86f0afd4 | 4650 | return; |
86f0afd4 TT |
4651 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); |
4652 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
4653 | ac->ac_f_ex.fe_len); | |
4654 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
c99d1e6e | 4655 | ext4_mb_unload_buddy(&e4b); |
86f0afd4 TT |
4656 | return; |
4657 | } | |
36cb0f52 KS |
4658 | if (pa->pa_type == MB_INODE_PA) { |
4659 | spin_lock(&pa->pa_lock); | |
400db9d3 | 4660 | pa->pa_free += ac->ac_b_ex.fe_len; |
36cb0f52 KS |
4661 | spin_unlock(&pa->pa_lock); |
4662 | } | |
b844167e CW |
4663 | } |
4664 | ||
c9de560d AT |
4665 | /* |
4666 | * use blocks preallocated to inode | |
4667 | */ | |
4668 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
4669 | struct ext4_prealloc_space *pa) | |
4670 | { | |
53accfa9 | 4671 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
4672 | ext4_fsblk_t start; |
4673 | ext4_fsblk_t end; | |
4674 | int len; | |
4675 | ||
4676 | /* found preallocated blocks, use them */ | |
4677 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
4678 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
4679 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
4680 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
4681 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
4682 | &ac->ac_b_ex.fe_start); | |
4683 | ac->ac_b_ex.fe_len = len; | |
4684 | ac->ac_status = AC_STATUS_FOUND; | |
4685 | ac->ac_pa = pa; | |
4686 | ||
4687 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 4688 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d | 4689 | BUG_ON(pa->pa_free < len); |
93cdf49f | 4690 | BUG_ON(ac->ac_b_ex.fe_len <= 0); |
c9de560d AT |
4691 | pa->pa_free -= len; |
4692 | ||
d3df1453 | 4693 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
c9de560d AT |
4694 | } |
4695 | ||
4696 | /* | |
4697 | * use blocks preallocated to locality group | |
4698 | */ | |
4699 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
4700 | struct ext4_prealloc_space *pa) | |
4701 | { | |
03cddb80 | 4702 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 4703 | |
c9de560d AT |
4704 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
4705 | &ac->ac_b_ex.fe_group, | |
4706 | &ac->ac_b_ex.fe_start); | |
4707 | ac->ac_b_ex.fe_len = len; | |
4708 | ac->ac_status = AC_STATUS_FOUND; | |
4709 | ac->ac_pa = pa; | |
4710 | ||
1221b235 | 4711 | /* we don't correct pa_pstart or pa_len here to avoid |
26346ff6 | 4712 | * possible race when the group is being loaded concurrently |
c9de560d | 4713 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
4714 | * in on-disk bitmap -- see ext4_mb_release_context() |
4715 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 4716 | */ |
d3df1453 | 4717 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
1afdc588 | 4718 | pa->pa_lstart, len, pa); |
c9de560d AT |
4719 | } |
4720 | ||
5e745b04 AK |
4721 | /* |
4722 | * Return the prealloc space that have minimal distance | |
4723 | * from the goal block. @cpa is the prealloc | |
4724 | * space that is having currently known minimal distance | |
4725 | * from the goal block. | |
4726 | */ | |
4727 | static struct ext4_prealloc_space * | |
4728 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
4729 | struct ext4_prealloc_space *pa, | |
4730 | struct ext4_prealloc_space *cpa) | |
4731 | { | |
4732 | ext4_fsblk_t cur_distance, new_distance; | |
4733 | ||
4734 | if (cpa == NULL) { | |
4735 | atomic_inc(&pa->pa_count); | |
4736 | return pa; | |
4737 | } | |
79211c8e AM |
4738 | cur_distance = abs(goal_block - cpa->pa_pstart); |
4739 | new_distance = abs(goal_block - pa->pa_pstart); | |
5e745b04 | 4740 | |
5a54b2f1 | 4741 | if (cur_distance <= new_distance) |
5e745b04 AK |
4742 | return cpa; |
4743 | ||
4744 | /* drop the previous reference */ | |
4745 | atomic_dec(&cpa->pa_count); | |
4746 | atomic_inc(&pa->pa_count); | |
4747 | return pa; | |
4748 | } | |
4749 | ||
1eff5904 KS |
4750 | /* |
4751 | * check if found pa meets EXT4_MB_HINT_GOAL_ONLY | |
4752 | */ | |
4753 | static bool | |
4754 | ext4_mb_pa_goal_check(struct ext4_allocation_context *ac, | |
4755 | struct ext4_prealloc_space *pa) | |
4756 | { | |
4757 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4758 | ext4_fsblk_t start; | |
4759 | ||
4760 | if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))) | |
4761 | return true; | |
4762 | ||
4763 | /* | |
4764 | * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted | |
4765 | * in ext4_mb_normalize_request and will keep same with ac_o_ex | |
4766 | * from ext4_mb_initialize_context. Choose ac_g_ex here to keep | |
4767 | * consistent with ext4_mb_find_by_goal. | |
4768 | */ | |
4769 | start = pa->pa_pstart + | |
4770 | (ac->ac_g_ex.fe_logical - pa->pa_lstart); | |
4771 | if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start) | |
4772 | return false; | |
4773 | ||
4774 | if (ac->ac_g_ex.fe_len > pa->pa_len - | |
4775 | EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart)) | |
4776 | return false; | |
4777 | ||
4778 | return true; | |
4779 | } | |
4780 | ||
c9de560d AT |
4781 | /* |
4782 | * search goal blocks in preallocated space | |
4783 | */ | |
4fca8f07 | 4784 | static noinline_for_stack bool |
4ddfef7b | 4785 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
c9de560d | 4786 | { |
53accfa9 | 4787 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 4788 | int order, i; |
c9de560d AT |
4789 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
4790 | struct ext4_locality_group *lg; | |
9d3de7ee | 4791 | struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL; |
38727786 | 4792 | struct rb_node *iter; |
5e745b04 | 4793 | ext4_fsblk_t goal_block; |
c9de560d AT |
4794 | |
4795 | /* only data can be preallocated */ | |
4796 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4fca8f07 | 4797 | return false; |
c9de560d | 4798 | |
9d3de7ee OM |
4799 | /* |
4800 | * first, try per-file preallocation by searching the inode pa rbtree. | |
4801 | * | |
4802 | * Here, we can't do a direct traversal of the tree because | |
4803 | * ext4_mb_discard_group_preallocation() can paralelly mark the pa | |
4804 | * deleted and that can cause direct traversal to skip some entries. | |
4805 | */ | |
38727786 | 4806 | read_lock(&ei->i_prealloc_lock); |
9d3de7ee OM |
4807 | |
4808 | if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) { | |
4809 | goto try_group_pa; | |
4810 | } | |
4811 | ||
4812 | /* | |
4813 | * Step 1: Find a pa with logical start immediately adjacent to the | |
4814 | * original logical start. This could be on the left or right. | |
4815 | * | |
4816 | * (tmp_pa->pa_lstart never changes so we can skip locking for it). | |
4817 | */ | |
38727786 OM |
4818 | for (iter = ei->i_prealloc_node.rb_node; iter; |
4819 | iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, | |
9d3de7ee | 4820 | tmp_pa->pa_lstart, iter)) { |
38727786 OM |
4821 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
4822 | pa_node.inode_node); | |
9d3de7ee | 4823 | } |
c9de560d | 4824 | |
9d3de7ee OM |
4825 | /* |
4826 | * Step 2: The adjacent pa might be to the right of logical start, find | |
4827 | * the left adjacent pa. After this step we'd have a valid tmp_pa whose | |
4828 | * logical start is towards the left of original request's logical start | |
4829 | */ | |
4830 | if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) { | |
4831 | struct rb_node *tmp; | |
4832 | tmp = rb_prev(&tmp_pa->pa_node.inode_node); | |
c9de560d | 4833 | |
9d3de7ee OM |
4834 | if (tmp) { |
4835 | tmp_pa = rb_entry(tmp, struct ext4_prealloc_space, | |
4836 | pa_node.inode_node); | |
4837 | } else { | |
e86a7182 | 4838 | /* |
9d3de7ee OM |
4839 | * If there is no adjacent pa to the left then finding |
4840 | * an overlapping pa is not possible hence stop searching | |
4841 | * inode pa tree | |
e86a7182 | 4842 | */ |
9d3de7ee | 4843 | goto try_group_pa; |
e86a7182 | 4844 | } |
9d3de7ee OM |
4845 | } |
4846 | ||
4847 | BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); | |
fb0a387d | 4848 | |
9d3de7ee OM |
4849 | /* |
4850 | * Step 3: If the left adjacent pa is deleted, keep moving left to find | |
4851 | * the first non deleted adjacent pa. After this step we should have a | |
4852 | * valid tmp_pa which is guaranteed to be non deleted. | |
4853 | */ | |
4854 | for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) { | |
4855 | if (!iter) { | |
4856 | /* | |
4857 | * no non deleted left adjacent pa, so stop searching | |
4858 | * inode pa tree | |
4859 | */ | |
4860 | goto try_group_pa; | |
4861 | } | |
4862 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space, | |
4863 | pa_node.inode_node); | |
bcf43499 | 4864 | spin_lock(&tmp_pa->pa_lock); |
9d3de7ee OM |
4865 | if (tmp_pa->pa_deleted == 0) { |
4866 | /* | |
4867 | * We will keep holding the pa_lock from | |
4868 | * this point on because we don't want group discard | |
4869 | * to delete this pa underneath us. Since group | |
4870 | * discard is anyways an ENOSPC operation it | |
4871 | * should be okay for it to wait a few more cycles. | |
4872 | */ | |
4873 | break; | |
4874 | } else { | |
bcf43499 | 4875 | spin_unlock(&tmp_pa->pa_lock); |
c9de560d | 4876 | } |
9d3de7ee OM |
4877 | } |
4878 | ||
4879 | BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); | |
4880 | BUG_ON(tmp_pa->pa_deleted == 1); | |
4881 | ||
4882 | /* | |
4883 | * Step 4: We now have the non deleted left adjacent pa. Only this | |
4884 | * pa can possibly satisfy the request hence check if it overlaps | |
4885 | * original logical start and stop searching if it doesn't. | |
4886 | */ | |
43bbddc0 | 4887 | if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) { |
bcf43499 | 4888 | spin_unlock(&tmp_pa->pa_lock); |
9d3de7ee OM |
4889 | goto try_group_pa; |
4890 | } | |
4891 | ||
4892 | /* non-extent files can't have physical blocks past 2^32 */ | |
4893 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && | |
4894 | (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) > | |
4895 | EXT4_MAX_BLOCK_FILE_PHYS)) { | |
4896 | /* | |
4897 | * Since PAs don't overlap, we won't find any other PA to | |
4898 | * satisfy this. | |
4899 | */ | |
4900 | spin_unlock(&tmp_pa->pa_lock); | |
4901 | goto try_group_pa; | |
4902 | } | |
4903 | ||
4904 | if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) { | |
4905 | atomic_inc(&tmp_pa->pa_count); | |
4906 | ext4_mb_use_inode_pa(ac, tmp_pa); | |
4907 | spin_unlock(&tmp_pa->pa_lock); | |
4908 | read_unlock(&ei->i_prealloc_lock); | |
4909 | return true; | |
4910 | } else { | |
4911 | /* | |
4912 | * We found a valid overlapping pa but couldn't use it because | |
4913 | * it had no free blocks. This should ideally never happen | |
4914 | * because: | |
4915 | * | |
4916 | * 1. When a new inode pa is added to rbtree it must have | |
4917 | * pa_free > 0 since otherwise we won't actually need | |
4918 | * preallocation. | |
4919 | * | |
4920 | * 2. An inode pa that is in the rbtree can only have it's | |
4921 | * pa_free become zero when another thread calls: | |
4922 | * ext4_mb_new_blocks | |
4923 | * ext4_mb_use_preallocated | |
4924 | * ext4_mb_use_inode_pa | |
4925 | * | |
4926 | * 3. Further, after the above calls make pa_free == 0, we will | |
4927 | * immediately remove it from the rbtree in: | |
4928 | * ext4_mb_new_blocks | |
4929 | * ext4_mb_release_context | |
4930 | * ext4_mb_put_pa | |
4931 | * | |
4932 | * 4. Since the pa_free becoming 0 and pa_free getting removed | |
4933 | * from tree both happen in ext4_mb_new_blocks, which is always | |
4934 | * called with i_data_sem held for data allocations, we can be | |
4935 | * sure that another process will never see a pa in rbtree with | |
4936 | * pa_free == 0. | |
4937 | */ | |
4938 | WARN_ON_ONCE(tmp_pa->pa_free == 0); | |
c9de560d | 4939 | } |
9d3de7ee OM |
4940 | spin_unlock(&tmp_pa->pa_lock); |
4941 | try_group_pa: | |
38727786 | 4942 | read_unlock(&ei->i_prealloc_lock); |
c9de560d AT |
4943 | |
4944 | /* can we use group allocation? */ | |
4945 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
4fca8f07 | 4946 | return false; |
c9de560d AT |
4947 | |
4948 | /* inode may have no locality group for some reason */ | |
4949 | lg = ac->ac_lg; | |
4950 | if (lg == NULL) | |
4fca8f07 | 4951 | return false; |
6be2ded1 AK |
4952 | order = fls(ac->ac_o_ex.fe_len) - 1; |
4953 | if (order > PREALLOC_TB_SIZE - 1) | |
4954 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4955 | order = PREALLOC_TB_SIZE - 1; | |
4956 | ||
bda00de7 | 4957 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
4958 | /* |
4959 | * search for the prealloc space that is having | |
4960 | * minimal distance from the goal block. | |
4961 | */ | |
6be2ded1 AK |
4962 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
4963 | rcu_read_lock(); | |
bcf43499 | 4964 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i], |
a8e38fd3 | 4965 | pa_node.lg_list) { |
bcf43499 OM |
4966 | spin_lock(&tmp_pa->pa_lock); |
4967 | if (tmp_pa->pa_deleted == 0 && | |
4968 | tmp_pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
4969 | |
4970 | cpa = ext4_mb_check_group_pa(goal_block, | |
bcf43499 | 4971 | tmp_pa, cpa); |
6be2ded1 | 4972 | } |
bcf43499 | 4973 | spin_unlock(&tmp_pa->pa_lock); |
c9de560d | 4974 | } |
6be2ded1 | 4975 | rcu_read_unlock(); |
c9de560d | 4976 | } |
5e745b04 AK |
4977 | if (cpa) { |
4978 | ext4_mb_use_group_pa(ac, cpa); | |
4fca8f07 | 4979 | return true; |
5e745b04 | 4980 | } |
4fca8f07 | 4981 | return false; |
c9de560d AT |
4982 | } |
4983 | ||
4984 | /* | |
4985 | * the function goes through all preallocation in this group and marks them | |
4986 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 4987 | * Need to be called with ext4 group lock held |
c9de560d | 4988 | */ |
089ceecc ES |
4989 | static noinline_for_stack |
4990 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
4991 | ext4_group_t group) |
4992 | { | |
4993 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
4994 | struct ext4_prealloc_space *pa; | |
4995 | struct list_head *cur; | |
4996 | ext4_group_t groupnr; | |
4997 | ext4_grpblk_t start; | |
4998 | int preallocated = 0; | |
c9de560d AT |
4999 | int len; |
5000 | ||
5354b2af TT |
5001 | if (!grp) |
5002 | return; | |
5003 | ||
c9de560d AT |
5004 | /* all form of preallocation discards first load group, |
5005 | * so the only competing code is preallocation use. | |
5006 | * we don't need any locking here | |
5007 | * notice we do NOT ignore preallocations with pa_deleted | |
5008 | * otherwise we could leave used blocks available for | |
5009 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
5010 | * is dropping preallocation | |
5011 | */ | |
5012 | list_for_each(cur, &grp->bb_prealloc_list) { | |
5013 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
5014 | spin_lock(&pa->pa_lock); | |
5015 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
5016 | &groupnr, &start); | |
5017 | len = pa->pa_len; | |
5018 | spin_unlock(&pa->pa_lock); | |
5019 | if (unlikely(len == 0)) | |
5020 | continue; | |
5021 | BUG_ON(groupnr != group); | |
123e3016 | 5022 | mb_set_bits(bitmap, start, len); |
c9de560d | 5023 | preallocated += len; |
c9de560d | 5024 | } |
d3df1453 | 5025 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
c9de560d AT |
5026 | } |
5027 | ||
27bc446e | 5028 | static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
5029 | struct ext4_prealloc_space *pa) | |
5030 | { | |
5031 | struct ext4_inode_info *ei; | |
5032 | ||
5033 | if (pa->pa_deleted) { | |
5034 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", | |
5035 | pa->pa_type, pa->pa_pstart, pa->pa_lstart, | |
5036 | pa->pa_len); | |
5037 | return; | |
5038 | } | |
5039 | ||
5040 | pa->pa_deleted = 1; | |
5041 | ||
5042 | if (pa->pa_type == MB_INODE_PA) { | |
5043 | ei = EXT4_I(pa->pa_inode); | |
5044 | atomic_dec(&ei->i_prealloc_active); | |
5045 | } | |
5046 | } | |
5047 | ||
82089725 | 5048 | static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa) |
c9de560d | 5049 | { |
82089725 | 5050 | BUG_ON(!pa); |
4e8d2139 JR |
5051 | BUG_ON(atomic_read(&pa->pa_count)); |
5052 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
5053 | kmem_cache_free(ext4_pspace_cachep, pa); |
5054 | } | |
5055 | ||
82089725 OM |
5056 | static void ext4_mb_pa_callback(struct rcu_head *head) |
5057 | { | |
5058 | struct ext4_prealloc_space *pa; | |
5059 | ||
5060 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
5061 | ext4_mb_pa_free(pa); | |
5062 | } | |
5063 | ||
c9de560d AT |
5064 | /* |
5065 | * drops a reference to preallocated space descriptor | |
5066 | * if this was the last reference and the space is consumed | |
5067 | */ | |
5068 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
5069 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
5070 | { | |
a9df9a49 | 5071 | ext4_group_t grp; |
d33a1976 | 5072 | ext4_fsblk_t grp_blk; |
38727786 | 5073 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
c9de560d | 5074 | |
c9de560d AT |
5075 | /* in this short window concurrent discard can set pa_deleted */ |
5076 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
5077 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
5078 | spin_unlock(&pa->pa_lock); | |
5079 | return; | |
5080 | } | |
5081 | ||
c9de560d AT |
5082 | if (pa->pa_deleted == 1) { |
5083 | spin_unlock(&pa->pa_lock); | |
5084 | return; | |
5085 | } | |
5086 | ||
27bc446e | 5087 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d AT |
5088 | spin_unlock(&pa->pa_lock); |
5089 | ||
d33a1976 | 5090 | grp_blk = pa->pa_pstart; |
60e6679e | 5091 | /* |
cc0fb9ad AK |
5092 | * If doing group-based preallocation, pa_pstart may be in the |
5093 | * next group when pa is used up | |
5094 | */ | |
5095 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
5096 | grp_blk--; |
5097 | ||
bd86298e | 5098 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
5099 | |
5100 | /* | |
5101 | * possible race: | |
5102 | * | |
5103 | * P1 (buddy init) P2 (regular allocation) | |
5104 | * find block B in PA | |
5105 | * copy on-disk bitmap to buddy | |
5106 | * mark B in on-disk bitmap | |
5107 | * drop PA from group | |
5108 | * mark all PAs in buddy | |
5109 | * | |
5110 | * thus, P1 initializes buddy with B available. to prevent this | |
5111 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
5112 | * against that pair | |
5113 | */ | |
5114 | ext4_lock_group(sb, grp); | |
5115 | list_del(&pa->pa_group_list); | |
5116 | ext4_unlock_group(sb, grp); | |
5117 | ||
a8e38fd3 | 5118 | if (pa->pa_type == MB_INODE_PA) { |
38727786 OM |
5119 | write_lock(pa->pa_node_lock.inode_lock); |
5120 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); | |
5121 | write_unlock(pa->pa_node_lock.inode_lock); | |
5122 | ext4_mb_pa_free(pa); | |
a8e38fd3 OM |
5123 | } else { |
5124 | spin_lock(pa->pa_node_lock.lg_lock); | |
5125 | list_del_rcu(&pa->pa_node.lg_list); | |
5126 | spin_unlock(pa->pa_node_lock.lg_lock); | |
38727786 OM |
5127 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
5128 | } | |
5129 | } | |
5130 | ||
5131 | static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new) | |
5132 | { | |
5133 | struct rb_node **iter = &root->rb_node, *parent = NULL; | |
5134 | struct ext4_prealloc_space *iter_pa, *new_pa; | |
5135 | ext4_lblk_t iter_start, new_start; | |
5136 | ||
5137 | while (*iter) { | |
5138 | iter_pa = rb_entry(*iter, struct ext4_prealloc_space, | |
5139 | pa_node.inode_node); | |
5140 | new_pa = rb_entry(new, struct ext4_prealloc_space, | |
5141 | pa_node.inode_node); | |
5142 | iter_start = iter_pa->pa_lstart; | |
5143 | new_start = new_pa->pa_lstart; | |
5144 | ||
5145 | parent = *iter; | |
5146 | if (new_start < iter_start) | |
5147 | iter = &((*iter)->rb_left); | |
5148 | else | |
5149 | iter = &((*iter)->rb_right); | |
a8e38fd3 | 5150 | } |
c9de560d | 5151 | |
38727786 OM |
5152 | rb_link_node(new, parent, iter); |
5153 | rb_insert_color(new, root); | |
c9de560d AT |
5154 | } |
5155 | ||
5156 | /* | |
5157 | * creates new preallocated space for given inode | |
5158 | */ | |
53f86b17 | 5159 | static noinline_for_stack void |
4ddfef7b | 5160 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
5161 | { |
5162 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 5163 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
5164 | struct ext4_prealloc_space *pa; |
5165 | struct ext4_group_info *grp; | |
5166 | struct ext4_inode_info *ei; | |
5167 | ||
5168 | /* preallocate only when found space is larger then requested */ | |
5169 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
5170 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
5171 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 5172 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 5173 | |
53f86b17 | 5174 | pa = ac->ac_pa; |
c9de560d | 5175 | |
7e170922 | 5176 | if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) { |
bc056e71 BL |
5177 | struct ext4_free_extent ex = { |
5178 | .fe_logical = ac->ac_g_ex.fe_logical, | |
5179 | .fe_len = ac->ac_orig_goal_len, | |
5180 | }; | |
5181 | loff_t orig_goal_end = extent_logical_end(sbi, &ex); | |
4fbf8bc7 | 5182 | loff_t o_ex_end = extent_logical_end(sbi, &ac->ac_o_ex); |
c9de560d | 5183 | |
4fbf8bc7 BL |
5184 | /* |
5185 | * We can't allocate as much as normalizer wants, so we try | |
5186 | * to get proper lstart to cover the original request, except | |
5187 | * when the goal doesn't cover the original request as below: | |
5188 | * | |
5189 | * orig_ex:2045/2055(10), isize:8417280 -> normalized:0/2048 | |
5190 | * best_ex:0/200(200) -> adjusted: 1848/2048(200) | |
5191 | */ | |
c9de560d AT |
5192 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); |
5193 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
5194 | ||
93cdf49f OM |
5195 | /* |
5196 | * Use the below logic for adjusting best extent as it keeps | |
5197 | * fragmentation in check while ensuring logical range of best | |
5198 | * extent doesn't overflow out of goal extent: | |
5199 | * | |
7e170922 OM |
5200 | * 1. Check if best ex can be kept at end of goal (before |
5201 | * cr_best_avail trimmed it) and still cover original start | |
93cdf49f | 5202 | * 2. Else, check if best ex can be kept at start of goal and |
4fbf8bc7 | 5203 | * still cover original end |
93cdf49f OM |
5204 | * 3. Else, keep the best ex at start of original request. |
5205 | */ | |
bc056e71 | 5206 | ex.fe_len = ac->ac_b_ex.fe_len; |
c9de560d | 5207 | |
bc056e71 BL |
5208 | ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len); |
5209 | if (ac->ac_o_ex.fe_logical >= ex.fe_logical) | |
93cdf49f | 5210 | goto adjust_bex; |
c9de560d | 5211 | |
bc056e71 | 5212 | ex.fe_logical = ac->ac_g_ex.fe_logical; |
4fbf8bc7 | 5213 | if (o_ex_end <= extent_logical_end(sbi, &ex)) |
bc056e71 | 5214 | goto adjust_bex; |
c9de560d | 5215 | |
bc056e71 | 5216 | ex.fe_logical = ac->ac_o_ex.fe_logical; |
93cdf49f | 5217 | adjust_bex: |
bc056e71 | 5218 | ac->ac_b_ex.fe_logical = ex.fe_logical; |
c9de560d | 5219 | |
c9de560d | 5220 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
bc056e71 | 5221 | BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end); |
c9de560d AT |
5222 | } |
5223 | ||
c9de560d AT |
5224 | pa->pa_lstart = ac->ac_b_ex.fe_logical; |
5225 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
5226 | pa->pa_len = ac->ac_b_ex.fe_len; | |
5227 | pa->pa_free = pa->pa_len; | |
c9de560d | 5228 | spin_lock_init(&pa->pa_lock); |
d794bf8e | 5229 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 5230 | pa->pa_deleted = 0; |
cc0fb9ad | 5231 | pa->pa_type = MB_INODE_PA; |
c9de560d | 5232 | |
d3df1453 RH |
5233 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
5234 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 5235 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d | 5236 | |
53accfa9 | 5237 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
abc075d4 | 5238 | ext4_mb_use_inode_pa(ac, pa); |
c9de560d AT |
5239 | |
5240 | ei = EXT4_I(ac->ac_inode); | |
5241 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
5354b2af TT |
5242 | if (!grp) |
5243 | return; | |
c9de560d | 5244 | |
a8e38fd3 | 5245 | pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock; |
c9de560d AT |
5246 | pa->pa_inode = ac->ac_inode; |
5247 | ||
c9de560d | 5248 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 5249 | |
38727786 OM |
5250 | write_lock(pa->pa_node_lock.inode_lock); |
5251 | ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node); | |
5252 | write_unlock(pa->pa_node_lock.inode_lock); | |
27bc446e | 5253 | atomic_inc(&ei->i_prealloc_active); |
c9de560d AT |
5254 | } |
5255 | ||
5256 | /* | |
5257 | * creates new preallocated space for locality group inodes belongs to | |
5258 | */ | |
53f86b17 | 5259 | static noinline_for_stack void |
4ddfef7b | 5260 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
c9de560d AT |
5261 | { |
5262 | struct super_block *sb = ac->ac_sb; | |
5263 | struct ext4_locality_group *lg; | |
5264 | struct ext4_prealloc_space *pa; | |
5265 | struct ext4_group_info *grp; | |
5266 | ||
5267 | /* preallocate only when found space is larger then requested */ | |
5268 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
5269 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
5270 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
53f86b17 | 5271 | BUG_ON(ac->ac_pa == NULL); |
c9de560d | 5272 | |
53f86b17 | 5273 | pa = ac->ac_pa; |
c9de560d | 5274 | |
c9de560d AT |
5275 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
5276 | pa->pa_lstart = pa->pa_pstart; | |
5277 | pa->pa_len = ac->ac_b_ex.fe_len; | |
5278 | pa->pa_free = pa->pa_len; | |
c9de560d | 5279 | spin_lock_init(&pa->pa_lock); |
a8e38fd3 | 5280 | INIT_LIST_HEAD(&pa->pa_node.lg_list); |
d794bf8e | 5281 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 5282 | pa->pa_deleted = 0; |
cc0fb9ad | 5283 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 5284 | |
d3df1453 RH |
5285 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
5286 | pa->pa_len, pa->pa_lstart); | |
9bffad1e | 5287 | trace_ext4_mb_new_group_pa(ac, pa); |
c9de560d AT |
5288 | |
5289 | ext4_mb_use_group_pa(ac, pa); | |
5290 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
5291 | ||
5292 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
5354b2af TT |
5293 | if (!grp) |
5294 | return; | |
c9de560d AT |
5295 | lg = ac->ac_lg; |
5296 | BUG_ON(lg == NULL); | |
5297 | ||
a8e38fd3 | 5298 | pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock; |
c9de560d AT |
5299 | pa->pa_inode = NULL; |
5300 | ||
c9de560d | 5301 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
c9de560d | 5302 | |
6be2ded1 AK |
5303 | /* |
5304 | * We will later add the new pa to the right bucket | |
5305 | * after updating the pa_free in ext4_mb_release_context | |
5306 | */ | |
c9de560d AT |
5307 | } |
5308 | ||
53f86b17 | 5309 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
c9de560d | 5310 | { |
c9de560d | 5311 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
53f86b17 | 5312 | ext4_mb_new_group_pa(ac); |
c9de560d | 5313 | else |
53f86b17 | 5314 | ext4_mb_new_inode_pa(ac); |
c9de560d AT |
5315 | } |
5316 | ||
5317 | /* | |
5318 | * finds all unused blocks in on-disk bitmap, frees them in | |
5319 | * in-core bitmap and buddy. | |
5320 | * @pa must be unlinked from inode and group lists, so that | |
5321 | * nobody else can find/use it. | |
5322 | * the caller MUST hold group/inode locks. | |
5323 | * TODO: optimize the case when there are no in-core structures yet | |
5324 | */ | |
820c2808 | 5325 | static noinline_for_stack void |
4ddfef7b | 5326 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, |
3e1e5f50 | 5327 | struct ext4_prealloc_space *pa) |
c9de560d | 5328 | { |
c9de560d AT |
5329 | struct super_block *sb = e4b->bd_sb; |
5330 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
5331 | unsigned int end; |
5332 | unsigned int next; | |
c9de560d AT |
5333 | ext4_group_t group; |
5334 | ext4_grpblk_t bit; | |
ba80b101 | 5335 | unsigned long long grp_blk_start; |
c9de560d AT |
5336 | int free = 0; |
5337 | ||
5338 | BUG_ON(pa->pa_deleted == 0); | |
5339 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 5340 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
5341 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
5342 | end = bit + pa->pa_len; | |
5343 | ||
c9de560d | 5344 | while (bit < end) { |
ffad0a44 | 5345 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
5346 | if (bit >= end) |
5347 | break; | |
ffad0a44 | 5348 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
d3df1453 | 5349 | mb_debug(sb, "free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
5350 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
5351 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
5352 | free += next - bit; |
5353 | ||
3e1e5f50 | 5354 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
5355 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
5356 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 5357 | next - bit); |
c9de560d AT |
5358 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
5359 | bit = next + 1; | |
5360 | } | |
5361 | if (free != pa->pa_free) { | |
9d8b9ec4 | 5362 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
36bad423 | 5363 | "pa %p: logic %lu, phys. %lu, len %d", |
9d8b9ec4 TT |
5364 | pa, (unsigned long) pa->pa_lstart, |
5365 | (unsigned long) pa->pa_pstart, | |
36bad423 | 5366 | pa->pa_len); |
e29136f8 | 5367 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 5368 | free, pa->pa_free); |
e56eb659 AK |
5369 | /* |
5370 | * pa is already deleted so we use the value obtained | |
5371 | * from the bitmap and continue. | |
5372 | */ | |
c9de560d | 5373 | } |
c9de560d | 5374 | atomic_add(free, &sbi->s_mb_discarded); |
c9de560d AT |
5375 | } |
5376 | ||
20427949 | 5377 | static noinline_for_stack void |
4ddfef7b | 5378 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, |
3e1e5f50 | 5379 | struct ext4_prealloc_space *pa) |
c9de560d | 5380 | { |
c9de560d AT |
5381 | struct super_block *sb = e4b->bd_sb; |
5382 | ext4_group_t group; | |
5383 | ext4_grpblk_t bit; | |
5384 | ||
60e07cf5 | 5385 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
5386 | BUG_ON(pa->pa_deleted == 0); |
5387 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
463808f2 TT |
5388 | if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) { |
5389 | ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu", | |
5390 | e4b->bd_group, group, pa->pa_pstart); | |
20427949 | 5391 | return; |
463808f2 | 5392 | } |
c9de560d AT |
5393 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); |
5394 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 5395 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
5396 | } |
5397 | ||
5398 | /* | |
5399 | * releases all preallocations in given group | |
5400 | * | |
5401 | * first, we need to decide discard policy: | |
5402 | * - when do we discard | |
5403 | * 1) ENOSPC | |
5404 | * - how many do we discard | |
5405 | * 1) how many requested | |
5406 | */ | |
4ddfef7b ES |
5407 | static noinline_for_stack int |
5408 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
8c80fb31 | 5409 | ext4_group_t group, int *busy) |
c9de560d AT |
5410 | { |
5411 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
5412 | struct buffer_head *bitmap_bh = NULL; | |
5413 | struct ext4_prealloc_space *pa, *tmp; | |
0f6bc579 | 5414 | LIST_HEAD(list); |
c9de560d | 5415 | struct ext4_buddy e4b; |
38727786 | 5416 | struct ext4_inode_info *ei; |
c9de560d | 5417 | int err; |
8c80fb31 | 5418 | int free = 0; |
c9de560d | 5419 | |
5354b2af TT |
5420 | if (!grp) |
5421 | return 0; | |
d3df1453 | 5422 | mb_debug(sb, "discard preallocation for group %u\n", group); |
c9de560d | 5423 | if (list_empty(&grp->bb_prealloc_list)) |
bbc4ec77 | 5424 | goto out_dbg; |
c9de560d | 5425 | |
574ca174 | 5426 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5427 | if (IS_ERR(bitmap_bh)) { |
5428 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5429 | ext4_error_err(sb, -err, |
5430 | "Error %d reading block bitmap for %u", | |
5431 | err, group); | |
bbc4ec77 | 5432 | goto out_dbg; |
c9de560d AT |
5433 | } |
5434 | ||
5435 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 5436 | if (err) { |
9651e6b2 KK |
5437 | ext4_warning(sb, "Error %d loading buddy information for %u", |
5438 | err, group); | |
ce89f46c | 5439 | put_bh(bitmap_bh); |
bbc4ec77 | 5440 | goto out_dbg; |
ce89f46c | 5441 | } |
c9de560d | 5442 | |
c9de560d AT |
5443 | ext4_lock_group(sb, group); |
5444 | list_for_each_entry_safe(pa, tmp, | |
5445 | &grp->bb_prealloc_list, pa_group_list) { | |
5446 | spin_lock(&pa->pa_lock); | |
5447 | if (atomic_read(&pa->pa_count)) { | |
5448 | spin_unlock(&pa->pa_lock); | |
8c80fb31 | 5449 | *busy = 1; |
c9de560d AT |
5450 | continue; |
5451 | } | |
5452 | if (pa->pa_deleted) { | |
5453 | spin_unlock(&pa->pa_lock); | |
5454 | continue; | |
5455 | } | |
5456 | ||
5457 | /* seems this one can be freed ... */ | |
27bc446e | 5458 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 5459 | |
70022da8 YB |
5460 | if (!free) |
5461 | this_cpu_inc(discard_pa_seq); | |
5462 | ||
c9de560d AT |
5463 | /* we can trust pa_free ... */ |
5464 | free += pa->pa_free; | |
5465 | ||
5466 | spin_unlock(&pa->pa_lock); | |
5467 | ||
5468 | list_del(&pa->pa_group_list); | |
5469 | list_add(&pa->u.pa_tmp_list, &list); | |
5470 | } | |
5471 | ||
c9de560d AT |
5472 | /* now free all selected PAs */ |
5473 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
5474 | ||
5475 | /* remove from object (inode or locality group) */ | |
a8e38fd3 OM |
5476 | if (pa->pa_type == MB_GROUP_PA) { |
5477 | spin_lock(pa->pa_node_lock.lg_lock); | |
5478 | list_del_rcu(&pa->pa_node.lg_list); | |
5479 | spin_unlock(pa->pa_node_lock.lg_lock); | |
5480 | } else { | |
38727786 OM |
5481 | write_lock(pa->pa_node_lock.inode_lock); |
5482 | ei = EXT4_I(pa->pa_inode); | |
5483 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); | |
5484 | write_unlock(pa->pa_node_lock.inode_lock); | |
a8e38fd3 | 5485 | } |
c9de560d | 5486 | |
38727786 OM |
5487 | list_del(&pa->u.pa_tmp_list); |
5488 | ||
5489 | if (pa->pa_type == MB_GROUP_PA) { | |
3e1e5f50 | 5490 | ext4_mb_release_group_pa(&e4b, pa); |
38727786 OM |
5491 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
5492 | } else { | |
3e1e5f50 | 5493 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
38727786 OM |
5494 | ext4_mb_pa_free(pa); |
5495 | } | |
c9de560d AT |
5496 | } |
5497 | ||
c9de560d | 5498 | ext4_unlock_group(sb, group); |
e39e07fd | 5499 | ext4_mb_unload_buddy(&e4b); |
c9de560d | 5500 | put_bh(bitmap_bh); |
bbc4ec77 | 5501 | out_dbg: |
d3df1453 | 5502 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
8c80fb31 CX |
5503 | free, group, grp->bb_free); |
5504 | return free; | |
c9de560d AT |
5505 | } |
5506 | ||
5507 | /* | |
5508 | * releases all non-used preallocated blocks for given inode | |
5509 | * | |
5510 | * It's important to discard preallocations under i_data_sem | |
5511 | * We don't want another block to be served from the prealloc | |
5512 | * space when we are discarding the inode prealloc space. | |
5513 | * | |
5514 | * FIXME!! Make sure it is valid at all the call sites | |
5515 | */ | |
2ffd2a6a | 5516 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
5517 | { |
5518 | struct ext4_inode_info *ei = EXT4_I(inode); | |
5519 | struct super_block *sb = inode->i_sb; | |
5520 | struct buffer_head *bitmap_bh = NULL; | |
5521 | struct ext4_prealloc_space *pa, *tmp; | |
5522 | ext4_group_t group = 0; | |
0f6bc579 | 5523 | LIST_HEAD(list); |
c9de560d | 5524 | struct ext4_buddy e4b; |
38727786 | 5525 | struct rb_node *iter; |
c9de560d AT |
5526 | int err; |
5527 | ||
f0e54b60 | 5528 | if (!S_ISREG(inode->i_mode)) |
c9de560d | 5529 | return; |
c9de560d | 5530 | |
8016e29f HS |
5531 | if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) |
5532 | return; | |
5533 | ||
d3df1453 RH |
5534 | mb_debug(sb, "discard preallocation for inode %lu\n", |
5535 | inode->i_ino); | |
27bc446e | 5536 | trace_ext4_discard_preallocations(inode, |
f0e54b60 | 5537 | atomic_read(&ei->i_prealloc_active)); |
27bc446e | 5538 | |
c9de560d AT |
5539 | repeat: |
5540 | /* first, collect all pa's in the inode */ | |
38727786 | 5541 | write_lock(&ei->i_prealloc_lock); |
2ffd2a6a | 5542 | for (iter = rb_first(&ei->i_prealloc_node); iter; |
38727786 OM |
5543 | iter = rb_next(iter)) { |
5544 | pa = rb_entry(iter, struct ext4_prealloc_space, | |
5545 | pa_node.inode_node); | |
a8e38fd3 | 5546 | BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock); |
38727786 | 5547 | |
c9de560d AT |
5548 | spin_lock(&pa->pa_lock); |
5549 | if (atomic_read(&pa->pa_count)) { | |
5550 | /* this shouldn't happen often - nobody should | |
5551 | * use preallocation while we're discarding it */ | |
5552 | spin_unlock(&pa->pa_lock); | |
38727786 | 5553 | write_unlock(&ei->i_prealloc_lock); |
9d8b9ec4 TT |
5554 | ext4_msg(sb, KERN_ERR, |
5555 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
5556 | WARN_ON(1); |
5557 | schedule_timeout_uninterruptible(HZ); | |
5558 | goto repeat; | |
5559 | ||
5560 | } | |
5561 | if (pa->pa_deleted == 0) { | |
27bc446e | 5562 | ext4_mb_mark_pa_deleted(sb, pa); |
c9de560d | 5563 | spin_unlock(&pa->pa_lock); |
38727786 | 5564 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); |
c9de560d AT |
5565 | list_add(&pa->u.pa_tmp_list, &list); |
5566 | continue; | |
5567 | } | |
5568 | ||
5569 | /* someone is deleting pa right now */ | |
5570 | spin_unlock(&pa->pa_lock); | |
38727786 | 5571 | write_unlock(&ei->i_prealloc_lock); |
c9de560d AT |
5572 | |
5573 | /* we have to wait here because pa_deleted | |
5574 | * doesn't mean pa is already unlinked from | |
5575 | * the list. as we might be called from | |
5576 | * ->clear_inode() the inode will get freed | |
5577 | * and concurrent thread which is unlinking | |
5578 | * pa from inode's list may access already | |
5579 | * freed memory, bad-bad-bad */ | |
5580 | ||
5581 | /* XXX: if this happens too often, we can | |
5582 | * add a flag to force wait only in case | |
5583 | * of ->clear_inode(), but not in case of | |
5584 | * regular truncate */ | |
5585 | schedule_timeout_uninterruptible(HZ); | |
5586 | goto repeat; | |
5587 | } | |
38727786 | 5588 | write_unlock(&ei->i_prealloc_lock); |
c9de560d AT |
5589 | |
5590 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 5591 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 5592 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d | 5593 | |
9651e6b2 KK |
5594 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5595 | GFP_NOFS|__GFP_NOFAIL); | |
ce89f46c | 5596 | if (err) { |
54d3adbc TT |
5597 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5598 | err, group); | |
ce89f46c AK |
5599 | continue; |
5600 | } | |
c9de560d | 5601 | |
574ca174 | 5602 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
9008a58e DW |
5603 | if (IS_ERR(bitmap_bh)) { |
5604 | err = PTR_ERR(bitmap_bh); | |
54d3adbc TT |
5605 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
5606 | err, group); | |
e39e07fd | 5607 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 5608 | continue; |
c9de560d AT |
5609 | } |
5610 | ||
5611 | ext4_lock_group(sb, group); | |
5612 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5613 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
5614 | ext4_unlock_group(sb, group); |
5615 | ||
e39e07fd | 5616 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
5617 | put_bh(bitmap_bh); |
5618 | ||
5619 | list_del(&pa->u.pa_tmp_list); | |
38727786 | 5620 | ext4_mb_pa_free(pa); |
c9de560d AT |
5621 | } |
5622 | } | |
5623 | ||
53f86b17 RH |
5624 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
5625 | { | |
5626 | struct ext4_prealloc_space *pa; | |
5627 | ||
5628 | BUG_ON(ext4_pspace_cachep == NULL); | |
5629 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); | |
5630 | if (!pa) | |
5631 | return -ENOMEM; | |
5632 | atomic_set(&pa->pa_count, 1); | |
5633 | ac->ac_pa = pa; | |
5634 | return 0; | |
5635 | } | |
5636 | ||
82089725 | 5637 | static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac) |
53f86b17 RH |
5638 | { |
5639 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
5640 | ||
5641 | BUG_ON(!pa); | |
5642 | ac->ac_pa = NULL; | |
5643 | WARN_ON(!atomic_dec_and_test(&pa->pa_count)); | |
82089725 OM |
5644 | /* |
5645 | * current function is only called due to an error or due to | |
5646 | * len of found blocks < len of requested blocks hence the PA has not | |
5647 | * been added to grp->bb_prealloc_list. So we don't need to lock it | |
5648 | */ | |
5649 | pa->pa_deleted = 1; | |
5650 | ext4_mb_pa_free(pa); | |
53f86b17 RH |
5651 | } |
5652 | ||
6ba495e9 | 5653 | #ifdef CONFIG_EXT4_DEBUG |
e68cf40c | 5654 | static inline void ext4_mb_show_pa(struct super_block *sb) |
c9de560d | 5655 | { |
e68cf40c | 5656 | ext4_group_t i, ngroups; |
c9de560d | 5657 | |
95257987 | 5658 | if (ext4_forced_shutdown(sb)) |
e3570639 ES |
5659 | return; |
5660 | ||
8df9675f | 5661 | ngroups = ext4_get_groups_count(sb); |
d3df1453 | 5662 | mb_debug(sb, "groups: "); |
8df9675f | 5663 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
5664 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
5665 | struct ext4_prealloc_space *pa; | |
5666 | ext4_grpblk_t start; | |
5667 | struct list_head *cur; | |
5354b2af TT |
5668 | |
5669 | if (!grp) | |
5670 | continue; | |
c9de560d AT |
5671 | ext4_lock_group(sb, i); |
5672 | list_for_each(cur, &grp->bb_prealloc_list) { | |
5673 | pa = list_entry(cur, struct ext4_prealloc_space, | |
5674 | pa_group_list); | |
5675 | spin_lock(&pa->pa_lock); | |
5676 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
5677 | NULL, &start); | |
5678 | spin_unlock(&pa->pa_lock); | |
d3df1453 RH |
5679 | mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
5680 | pa->pa_len); | |
c9de560d | 5681 | } |
60bd63d1 | 5682 | ext4_unlock_group(sb, i); |
d3df1453 RH |
5683 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
5684 | grp->bb_fragments); | |
c9de560d | 5685 | } |
c9de560d | 5686 | } |
e68cf40c RH |
5687 | |
5688 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
5689 | { | |
5690 | struct super_block *sb = ac->ac_sb; | |
5691 | ||
95257987 | 5692 | if (ext4_forced_shutdown(sb)) |
e68cf40c RH |
5693 | return; |
5694 | ||
d3df1453 | 5695 | mb_debug(sb, "Can't allocate:" |
e68cf40c | 5696 | " Allocation context details:"); |
d3df1453 | 5697 | mb_debug(sb, "status %u flags 0x%x", |
e68cf40c | 5698 | ac->ac_status, ac->ac_flags); |
d3df1453 | 5699 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
e68cf40c RH |
5700 | "goal %lu/%lu/%lu@%lu, " |
5701 | "best %lu/%lu/%lu@%lu cr %d", | |
5702 | (unsigned long)ac->ac_o_ex.fe_group, | |
5703 | (unsigned long)ac->ac_o_ex.fe_start, | |
5704 | (unsigned long)ac->ac_o_ex.fe_len, | |
5705 | (unsigned long)ac->ac_o_ex.fe_logical, | |
5706 | (unsigned long)ac->ac_g_ex.fe_group, | |
5707 | (unsigned long)ac->ac_g_ex.fe_start, | |
5708 | (unsigned long)ac->ac_g_ex.fe_len, | |
5709 | (unsigned long)ac->ac_g_ex.fe_logical, | |
5710 | (unsigned long)ac->ac_b_ex.fe_group, | |
5711 | (unsigned long)ac->ac_b_ex.fe_start, | |
5712 | (unsigned long)ac->ac_b_ex.fe_len, | |
5713 | (unsigned long)ac->ac_b_ex.fe_logical, | |
5714 | (int)ac->ac_criteria); | |
d3df1453 | 5715 | mb_debug(sb, "%u found", ac->ac_found); |
569f196f RH |
5716 | mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no"); |
5717 | if (ac->ac_pa) | |
5718 | mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ? | |
5719 | "group pa" : "inode pa"); | |
e68cf40c RH |
5720 | ext4_mb_show_pa(sb); |
5721 | } | |
c9de560d | 5722 | #else |
e68cf40c RH |
5723 | static inline void ext4_mb_show_pa(struct super_block *sb) |
5724 | { | |
e68cf40c | 5725 | } |
c9de560d AT |
5726 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
5727 | { | |
e68cf40c | 5728 | ext4_mb_show_pa(ac->ac_sb); |
c9de560d AT |
5729 | } |
5730 | #endif | |
5731 | ||
5732 | /* | |
5733 | * We use locality group preallocation for small size file. The size of the | |
5734 | * file is determined by the current size or the resulting size after | |
5735 | * allocation which ever is larger | |
5736 | * | |
b713a5ec | 5737 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
5738 | */ |
5739 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
5740 | { | |
5741 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
5742 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
5743 | loff_t size, isize; | |
a9f2a293 | 5744 | bool inode_pa_eligible, group_pa_eligible; |
c9de560d AT |
5745 | |
5746 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
5747 | return; | |
5748 | ||
4ba74d00 TT |
5749 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
5750 | return; | |
5751 | ||
a9f2a293 JK |
5752 | group_pa_eligible = sbi->s_mb_group_prealloc > 0; |
5753 | inode_pa_eligible = true; | |
43bbddc0 | 5754 | size = extent_logical_end(sbi, &ac->ac_o_ex); |
50797481 TT |
5755 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
5756 | >> bsbits; | |
c9de560d | 5757 | |
a9f2a293 | 5758 | /* No point in using inode preallocation for closed files */ |
82dd124c | 5759 | if ((size == isize) && !ext4_fs_is_busy(sbi) && |
a9f2a293 JK |
5760 | !inode_is_open_for_write(ac->ac_inode)) |
5761 | inode_pa_eligible = false; | |
50797481 | 5762 | |
71780577 | 5763 | size = max(size, isize); |
a9f2a293 JK |
5764 | /* Don't use group allocation for large files */ |
5765 | if (size > sbi->s_mb_stream_request) | |
5766 | group_pa_eligible = false; | |
5767 | ||
5768 | if (!group_pa_eligible) { | |
5769 | if (inode_pa_eligible) | |
5770 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
5771 | else | |
5772 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
c9de560d | 5773 | return; |
4ba74d00 | 5774 | } |
c9de560d AT |
5775 | |
5776 | BUG_ON(ac->ac_lg != NULL); | |
5777 | /* | |
5778 | * locality group prealloc space are per cpu. The reason for having | |
5779 | * per cpu locality group is to reduce the contention between block | |
5780 | * request from multiple CPUs. | |
5781 | */ | |
a0b6bc63 | 5782 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
5783 | |
5784 | /* we're going to use group allocation */ | |
5785 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
5786 | ||
5787 | /* serialize all allocations in the group */ | |
5788 | mutex_lock(&ac->ac_lg->lg_mutex); | |
5789 | } | |
5790 | ||
d73eff68 | 5791 | static noinline_for_stack void |
4ddfef7b | 5792 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, |
c9de560d AT |
5793 | struct ext4_allocation_request *ar) |
5794 | { | |
5795 | struct super_block *sb = ar->inode->i_sb; | |
5796 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
5797 | struct ext4_super_block *es = sbi->s_es; | |
5798 | ext4_group_t group; | |
498e5f24 TT |
5799 | unsigned int len; |
5800 | ext4_fsblk_t goal; | |
c9de560d AT |
5801 | ext4_grpblk_t block; |
5802 | ||
5803 | /* we can't allocate > group size */ | |
5804 | len = ar->len; | |
5805 | ||
5806 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
5807 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
5808 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
5809 | |
5810 | /* start searching from the goal */ | |
5811 | goal = ar->goal; | |
5812 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
5813 | goal >= ext4_blocks_count(es)) | |
5814 | goal = le32_to_cpu(es->s_first_data_block); | |
5815 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
5816 | ||
5817 | /* set up allocation goals */ | |
f5a44db5 | 5818 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 5819 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
5820 | ac->ac_sb = sb; |
5821 | ac->ac_inode = ar->inode; | |
53accfa9 | 5822 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
5823 | ac->ac_o_ex.fe_group = group; |
5824 | ac->ac_o_ex.fe_start = block; | |
5825 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 5826 | ac->ac_g_ex = ac->ac_o_ex; |
7e170922 | 5827 | ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; |
c9de560d | 5828 | ac->ac_flags = ar->flags; |
c9de560d | 5829 | |
3cb77bd2 | 5830 | /* we have to define context: we'll work with a file or |
c9de560d AT |
5831 | * locality group. this is a policy, actually */ |
5832 | ext4_mb_group_or_file(ac); | |
5833 | ||
d3df1453 | 5834 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
c9de560d AT |
5835 | "left: %u/%u, right %u/%u to %swritable\n", |
5836 | (unsigned) ar->len, (unsigned) ar->logical, | |
5837 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
5838 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
5839 | (unsigned) ar->lright, (unsigned) ar->pright, | |
82dd124c | 5840 | inode_is_open_for_write(ar->inode) ? "" : "non-"); |
c9de560d AT |
5841 | } |
5842 | ||
6be2ded1 AK |
5843 | static noinline_for_stack void |
5844 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
5845 | struct ext4_locality_group *lg, | |
5846 | int order, int total_entries) | |
5847 | { | |
5848 | ext4_group_t group = 0; | |
5849 | struct ext4_buddy e4b; | |
0f6bc579 | 5850 | LIST_HEAD(discard_list); |
6be2ded1 | 5851 | struct ext4_prealloc_space *pa, *tmp; |
6be2ded1 | 5852 | |
d3df1453 | 5853 | mb_debug(sb, "discard locality group preallocation\n"); |
6be2ded1 | 5854 | |
6be2ded1 AK |
5855 | spin_lock(&lg->lg_prealloc_lock); |
5856 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
a8e38fd3 | 5857 | pa_node.lg_list, |
92e9c58c | 5858 | lockdep_is_held(&lg->lg_prealloc_lock)) { |
6be2ded1 AK |
5859 | spin_lock(&pa->pa_lock); |
5860 | if (atomic_read(&pa->pa_count)) { | |
5861 | /* | |
5862 | * This is the pa that we just used | |
5863 | * for block allocation. So don't | |
5864 | * free that | |
5865 | */ | |
5866 | spin_unlock(&pa->pa_lock); | |
5867 | continue; | |
5868 | } | |
5869 | if (pa->pa_deleted) { | |
5870 | spin_unlock(&pa->pa_lock); | |
5871 | continue; | |
5872 | } | |
5873 | /* only lg prealloc space */ | |
cc0fb9ad | 5874 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
5875 | |
5876 | /* seems this one can be freed ... */ | |
27bc446e | 5877 | ext4_mb_mark_pa_deleted(sb, pa); |
6be2ded1 AK |
5878 | spin_unlock(&pa->pa_lock); |
5879 | ||
a8e38fd3 | 5880 | list_del_rcu(&pa->pa_node.lg_list); |
6be2ded1 AK |
5881 | list_add(&pa->u.pa_tmp_list, &discard_list); |
5882 | ||
5883 | total_entries--; | |
5884 | if (total_entries <= 5) { | |
5885 | /* | |
5886 | * we want to keep only 5 entries | |
5887 | * allowing it to grow to 8. This | |
5888 | * mak sure we don't call discard | |
5889 | * soon for this list. | |
5890 | */ | |
5891 | break; | |
5892 | } | |
5893 | } | |
5894 | spin_unlock(&lg->lg_prealloc_lock); | |
5895 | ||
5896 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
9651e6b2 | 5897 | int err; |
6be2ded1 | 5898 | |
bd86298e | 5899 | group = ext4_get_group_number(sb, pa->pa_pstart); |
9651e6b2 KK |
5900 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
5901 | GFP_NOFS|__GFP_NOFAIL); | |
5902 | if (err) { | |
54d3adbc TT |
5903 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
5904 | err, group); | |
6be2ded1 AK |
5905 | continue; |
5906 | } | |
5907 | ext4_lock_group(sb, group); | |
5908 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 5909 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
5910 | ext4_unlock_group(sb, group); |
5911 | ||
e39e07fd | 5912 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
5913 | list_del(&pa->u.pa_tmp_list); |
5914 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
5915 | } | |
6be2ded1 AK |
5916 | } |
5917 | ||
5918 | /* | |
5919 | * We have incremented pa_count. So it cannot be freed at this | |
5920 | * point. Also we hold lg_mutex. So no parallel allocation is | |
5921 | * possible from this lg. That means pa_free cannot be updated. | |
5922 | * | |
5923 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
5924 | * which can cause the lg_prealloc_list to be updated. | |
5925 | */ | |
5926 | ||
5927 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
5928 | { | |
5929 | int order, added = 0, lg_prealloc_count = 1; | |
5930 | struct super_block *sb = ac->ac_sb; | |
5931 | struct ext4_locality_group *lg = ac->ac_lg; | |
5932 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
5933 | ||
5934 | order = fls(pa->pa_free) - 1; | |
5935 | if (order > PREALLOC_TB_SIZE - 1) | |
5936 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
5937 | order = PREALLOC_TB_SIZE - 1; | |
5938 | /* Add the prealloc space to lg */ | |
f1167009 | 5939 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 | 5940 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
a8e38fd3 | 5941 | pa_node.lg_list, |
92e9c58c | 5942 | lockdep_is_held(&lg->lg_prealloc_lock)) { |
6be2ded1 AK |
5943 | spin_lock(&tmp_pa->pa_lock); |
5944 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 5945 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
5946 | continue; |
5947 | } | |
5948 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
5949 | /* Add to the tail of the previous entry */ | |
a8e38fd3 OM |
5950 | list_add_tail_rcu(&pa->pa_node.lg_list, |
5951 | &tmp_pa->pa_node.lg_list); | |
6be2ded1 AK |
5952 | added = 1; |
5953 | /* | |
5954 | * we want to count the total | |
5955 | * number of entries in the list | |
5956 | */ | |
5957 | } | |
5958 | spin_unlock(&tmp_pa->pa_lock); | |
5959 | lg_prealloc_count++; | |
5960 | } | |
5961 | if (!added) | |
a8e38fd3 | 5962 | list_add_tail_rcu(&pa->pa_node.lg_list, |
6be2ded1 | 5963 | &lg->lg_prealloc_list[order]); |
f1167009 | 5964 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
5965 | |
5966 | /* Now trim the list to be not more than 8 elements */ | |
ad635507 | 5967 | if (lg_prealloc_count > 8) |
6be2ded1 | 5968 | ext4_mb_discard_lg_preallocations(sb, lg, |
f1167009 | 5969 | order, lg_prealloc_count); |
6be2ded1 AK |
5970 | } |
5971 | ||
c9de560d AT |
5972 | /* |
5973 | * release all resource we used in allocation | |
5974 | */ | |
11fd1a5d | 5975 | static void ext4_mb_release_context(struct ext4_allocation_context *ac) |
c9de560d | 5976 | { |
53accfa9 | 5977 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
5978 | struct ext4_prealloc_space *pa = ac->ac_pa; |
5979 | if (pa) { | |
cc0fb9ad | 5980 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 5981 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 5982 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
5983 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
5984 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
5985 | pa->pa_free -= ac->ac_b_ex.fe_len; |
5986 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
5987 | spin_unlock(&pa->pa_lock); | |
66d5e027 | 5988 | |
5989 | /* | |
5990 | * We want to add the pa to the right bucket. | |
5991 | * Remove it from the list and while adding | |
5992 | * make sure the list to which we are adding | |
5993 | * doesn't grow big. | |
5994 | */ | |
5995 | if (likely(pa->pa_free)) { | |
a8e38fd3 OM |
5996 | spin_lock(pa->pa_node_lock.lg_lock); |
5997 | list_del_rcu(&pa->pa_node.lg_list); | |
5998 | spin_unlock(pa->pa_node_lock.lg_lock); | |
66d5e027 | 5999 | ext4_mb_add_n_trim(ac); |
6000 | } | |
ba443916 | 6001 | } |
27bc446e | 6002 | |
ba443916 AK |
6003 | ext4_mb_put_pa(ac, ac->ac_sb, pa); |
6004 | } | |
ccedf35b MWO |
6005 | if (ac->ac_bitmap_folio) |
6006 | folio_put(ac->ac_bitmap_folio); | |
c84f1510 MWO |
6007 | if (ac->ac_buddy_folio) |
6008 | folio_put(ac->ac_buddy_folio); | |
c9de560d AT |
6009 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
6010 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
6011 | ext4_mb_collect_stats(ac); | |
c9de560d AT |
6012 | } |
6013 | ||
6014 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
6015 | { | |
8df9675f | 6016 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d | 6017 | int ret; |
8c80fb31 CX |
6018 | int freed = 0, busy = 0; |
6019 | int retry = 0; | |
c9de560d | 6020 | |
9bffad1e | 6021 | trace_ext4_mb_discard_preallocations(sb, needed); |
8c80fb31 CX |
6022 | |
6023 | if (needed == 0) | |
6024 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; | |
6025 | repeat: | |
8df9675f | 6026 | for (i = 0; i < ngroups && needed > 0; i++) { |
8c80fb31 | 6027 | ret = ext4_mb_discard_group_preallocations(sb, i, &busy); |
c9de560d AT |
6028 | freed += ret; |
6029 | needed -= ret; | |
8c80fb31 CX |
6030 | cond_resched(); |
6031 | } | |
6032 | ||
6033 | if (needed > 0 && busy && ++retry < 3) { | |
6034 | busy = 0; | |
6035 | goto repeat; | |
c9de560d AT |
6036 | } |
6037 | ||
6038 | return freed; | |
6039 | } | |
6040 | ||
cf5e2ca6 | 6041 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
07b5b8e1 | 6042 | struct ext4_allocation_context *ac, u64 *seq) |
cf5e2ca6 RH |
6043 | { |
6044 | int freed; | |
07b5b8e1 RH |
6045 | u64 seq_retry = 0; |
6046 | bool ret = false; | |
cf5e2ca6 RH |
6047 | |
6048 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); | |
07b5b8e1 RH |
6049 | if (freed) { |
6050 | ret = true; | |
6051 | goto out_dbg; | |
6052 | } | |
6053 | seq_retry = ext4_get_discard_pa_seq_sum(); | |
99377830 RH |
6054 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
6055 | ac->ac_flags |= EXT4_MB_STRICT_CHECK; | |
07b5b8e1 RH |
6056 | *seq = seq_retry; |
6057 | ret = true; | |
6058 | } | |
6059 | ||
6060 | out_dbg: | |
6061 | mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); | |
6062 | return ret; | |
cf5e2ca6 RH |
6063 | } |
6064 | ||
ad78b5ef KS |
6065 | /* |
6066 | * Simple allocator for Ext4 fast commit replay path. It searches for blocks | |
6067 | * linearly starting at the goal block and also excludes the blocks which | |
6068 | * are going to be in use after fast commit replay. | |
6069 | */ | |
6070 | static ext4_fsblk_t | |
6071 | ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp) | |
6072 | { | |
6073 | struct buffer_head *bitmap_bh; | |
6074 | struct super_block *sb = ar->inode->i_sb; | |
6075 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6076 | ext4_group_t group, nr; | |
6077 | ext4_grpblk_t blkoff; | |
6078 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); | |
6079 | ext4_grpblk_t i = 0; | |
6080 | ext4_fsblk_t goal, block; | |
79ebf48c | 6081 | struct ext4_super_block *es = sbi->s_es; |
ad78b5ef KS |
6082 | |
6083 | goal = ar->goal; | |
6084 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
6085 | goal >= ext4_blocks_count(es)) | |
6086 | goal = le32_to_cpu(es->s_first_data_block); | |
6087 | ||
6088 | ar->len = 0; | |
6089 | ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); | |
6090 | for (nr = ext4_get_groups_count(sb); nr > 0; nr--) { | |
6091 | bitmap_bh = ext4_read_block_bitmap(sb, group); | |
6092 | if (IS_ERR(bitmap_bh)) { | |
6093 | *errp = PTR_ERR(bitmap_bh); | |
6094 | pr_warn("Failed to read block bitmap\n"); | |
6095 | return 0; | |
6096 | } | |
6097 | ||
6098 | while (1) { | |
6099 | i = mb_find_next_zero_bit(bitmap_bh->b_data, max, | |
6100 | blkoff); | |
6101 | if (i >= max) | |
6102 | break; | |
6103 | if (ext4_fc_replay_check_excluded(sb, | |
6104 | ext4_group_first_block_no(sb, group) + | |
6105 | EXT4_C2B(sbi, i))) { | |
6106 | blkoff = i + 1; | |
6107 | } else | |
6108 | break; | |
6109 | } | |
6110 | brelse(bitmap_bh); | |
6111 | if (i < max) | |
6112 | break; | |
6113 | ||
6114 | if (++group >= ext4_get_groups_count(sb)) | |
6115 | group = 0; | |
6116 | ||
6117 | blkoff = 0; | |
6118 | } | |
6119 | ||
6120 | if (i >= max) { | |
6121 | *errp = -ENOSPC; | |
6122 | return 0; | |
6123 | } | |
6124 | ||
6125 | block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i); | |
d2f7cf40 | 6126 | ext4_mb_mark_bb(sb, block, 1, true); |
ad78b5ef KS |
6127 | ar->len = 1; |
6128 | ||
3f4830ab | 6129 | *errp = 0; |
ad78b5ef KS |
6130 | return block; |
6131 | } | |
8016e29f | 6132 | |
c9de560d AT |
6133 | /* |
6134 | * Main entry point into mballoc to allocate blocks | |
6135 | * it tries to use preallocation first, then falls back | |
6136 | * to usual allocation | |
6137 | */ | |
6138 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 6139 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 6140 | { |
256bdb49 | 6141 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
6142 | struct ext4_sb_info *sbi; |
6143 | struct super_block *sb; | |
6144 | ext4_fsblk_t block = 0; | |
60e58e0f | 6145 | unsigned int inquota = 0; |
53accfa9 | 6146 | unsigned int reserv_clstrs = 0; |
80fa46d6 | 6147 | int retries = 0; |
07b5b8e1 | 6148 | u64 seq; |
c9de560d | 6149 | |
b10a44c3 | 6150 | might_sleep(); |
c9de560d AT |
6151 | sb = ar->inode->i_sb; |
6152 | sbi = EXT4_SB(sb); | |
6153 | ||
9bffad1e | 6154 | trace_ext4_request_blocks(ar); |
8016e29f | 6155 | if (sbi->s_mount_state & EXT4_FC_REPLAY) |
ad78b5ef | 6156 | return ext4_mb_new_blocks_simple(ar, errp); |
ba80b101 | 6157 | |
45dc63e7 | 6158 | /* Allow to use superuser reservation for quota file */ |
02749a4c | 6159 | if (ext4_is_quota_file(ar->inode)) |
45dc63e7 DM |
6160 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
6161 | ||
e3cf5d5d | 6162 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
60e58e0f MC |
6163 | /* Without delayed allocation we need to verify |
6164 | * there is enough free blocks to do block allocation | |
6165 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 6166 | */ |
55f020db | 6167 | while (ar->len && |
e7d5f315 | 6168 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 6169 | |
030ba6bc | 6170 | /* let others to free the space */ |
bb8b20ed | 6171 | cond_resched(); |
030ba6bc AK |
6172 | ar->len = ar->len >> 1; |
6173 | } | |
6174 | if (!ar->len) { | |
bbc4ec77 | 6175 | ext4_mb_show_pa(sb); |
a30d542a AK |
6176 | *errp = -ENOSPC; |
6177 | return 0; | |
6178 | } | |
53accfa9 | 6179 | reserv_clstrs = ar->len; |
55f020db | 6180 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
6181 | dquot_alloc_block_nofail(ar->inode, |
6182 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
6183 | } else { |
6184 | while (ar->len && | |
53accfa9 TT |
6185 | dquot_alloc_block(ar->inode, |
6186 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
6187 | |
6188 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
6189 | ar->len--; | |
6190 | } | |
60e58e0f MC |
6191 | } |
6192 | inquota = ar->len; | |
6193 | if (ar->len == 0) { | |
6194 | *errp = -EDQUOT; | |
6c7a120a | 6195 | goto out; |
60e58e0f | 6196 | } |
07031431 | 6197 | } |
d2a17637 | 6198 | |
85556c9a | 6199 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 6200 | if (!ac) { |
363d4251 | 6201 | ar->len = 0; |
256bdb49 | 6202 | *errp = -ENOMEM; |
6c7a120a | 6203 | goto out; |
256bdb49 ES |
6204 | } |
6205 | ||
d73eff68 | 6206 | ext4_mb_initialize_context(ac, ar); |
c9de560d | 6207 | |
256bdb49 | 6208 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
81198536 | 6209 | seq = this_cpu_read(discard_pa_seq); |
256bdb49 | 6210 | if (!ext4_mb_use_preallocated(ac)) { |
256bdb49 ES |
6211 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
6212 | ext4_mb_normalize_request(ac, ar); | |
53f86b17 RH |
6213 | |
6214 | *errp = ext4_mb_pa_alloc(ac); | |
6215 | if (*errp) | |
6216 | goto errout; | |
c9de560d AT |
6217 | repeat: |
6218 | /* allocate space in core */ | |
6c7a120a | 6219 | *errp = ext4_mb_regular_allocator(ac); |
53f86b17 RH |
6220 | /* |
6221 | * pa allocated above is added to grp->bb_prealloc_list only | |
6222 | * when we were able to allocate some block i.e. when | |
6223 | * ac->ac_status == AC_STATUS_FOUND. | |
6224 | * And error from above mean ac->ac_status != AC_STATUS_FOUND | |
6225 | * So we have to free this pa here itself. | |
6226 | */ | |
2c00ef3e | 6227 | if (*errp) { |
82089725 | 6228 | ext4_mb_pa_put_free(ac); |
2c00ef3e AK |
6229 | ext4_discard_allocated_blocks(ac); |
6230 | goto errout; | |
6231 | } | |
53f86b17 RH |
6232 | if (ac->ac_status == AC_STATUS_FOUND && |
6233 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) | |
82089725 | 6234 | ext4_mb_pa_put_free(ac); |
c9de560d | 6235 | } |
256bdb49 | 6236 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 6237 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
554a5ccc | 6238 | if (*errp) { |
b844167e | 6239 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
6240 | goto errout; |
6241 | } else { | |
519deca0 AK |
6242 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
6243 | ar->len = ac->ac_b_ex.fe_len; | |
6244 | } | |
c9de560d | 6245 | } else { |
80fa46d6 TT |
6246 | if (++retries < 3 && |
6247 | ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) | |
c9de560d | 6248 | goto repeat; |
53f86b17 RH |
6249 | /* |
6250 | * If block allocation fails then the pa allocated above | |
6251 | * needs to be freed here itself. | |
6252 | */ | |
82089725 | 6253 | ext4_mb_pa_put_free(ac); |
c9de560d | 6254 | *errp = -ENOSPC; |
6c7a120a AK |
6255 | } |
6256 | ||
6257 | if (*errp) { | |
aaae558d | 6258 | errout: |
256bdb49 | 6259 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 6260 | ar->len = 0; |
256bdb49 | 6261 | ext4_mb_show_ac(ac); |
c9de560d | 6262 | } |
256bdb49 | 6263 | ext4_mb_release_context(ac); |
aaae558d | 6264 | kmem_cache_free(ext4_ac_cachep, ac); |
6c7a120a | 6265 | out: |
60e58e0f | 6266 | if (inquota && ar->len < inquota) |
53accfa9 | 6267 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 6268 | if (!ar->len) { |
e3cf5d5d | 6269 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) |
0087d9fb | 6270 | /* release all the reserved blocks if non delalloc */ |
57042651 | 6271 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 6272 | reserv_clstrs); |
0087d9fb | 6273 | } |
c9de560d | 6274 | |
9bffad1e | 6275 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 6276 | |
c9de560d AT |
6277 | return block; |
6278 | } | |
c9de560d | 6279 | |
c894058d AK |
6280 | /* |
6281 | * We can merge two free data extents only if the physical blocks | |
6282 | * are contiguous, AND the extents were freed by the same transaction, | |
6283 | * AND the blocks are associated with the same group. | |
6284 | */ | |
a0154344 DJ |
6285 | static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, |
6286 | struct ext4_free_data *entry, | |
6287 | struct ext4_free_data *new_entry, | |
6288 | struct rb_root *entry_rb_root) | |
c894058d | 6289 | { |
a0154344 DJ |
6290 | if ((entry->efd_tid != new_entry->efd_tid) || |
6291 | (entry->efd_group != new_entry->efd_group)) | |
6292 | return; | |
6293 | if (entry->efd_start_cluster + entry->efd_count == | |
6294 | new_entry->efd_start_cluster) { | |
6295 | new_entry->efd_start_cluster = entry->efd_start_cluster; | |
6296 | new_entry->efd_count += entry->efd_count; | |
6297 | } else if (new_entry->efd_start_cluster + new_entry->efd_count == | |
6298 | entry->efd_start_cluster) { | |
6299 | new_entry->efd_count += entry->efd_count; | |
6300 | } else | |
6301 | return; | |
6302 | spin_lock(&sbi->s_md_lock); | |
6303 | list_del(&entry->efd_list); | |
6304 | spin_unlock(&sbi->s_md_lock); | |
6305 | rb_erase(&entry->efd_node, entry_rb_root); | |
6306 | kmem_cache_free(ext4_free_data_cachep, entry); | |
c894058d AK |
6307 | } |
6308 | ||
85b67ffb | 6309 | static noinline_for_stack void |
4ddfef7b | 6310 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, |
7a2fcbf7 | 6311 | struct ext4_free_data *new_entry) |
c9de560d | 6312 | { |
e29136f8 | 6313 | ext4_group_t group = e4b->bd_group; |
84130193 | 6314 | ext4_grpblk_t cluster; |
d08854f5 | 6315 | ext4_grpblk_t clusters = new_entry->efd_count; |
7a2fcbf7 | 6316 | struct ext4_free_data *entry; |
c9de560d AT |
6317 | struct ext4_group_info *db = e4b->bd_info; |
6318 | struct super_block *sb = e4b->bd_sb; | |
6319 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
6320 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
6321 | struct rb_node *parent = NULL, *new_node; | |
6322 | ||
0390131b | 6323 | BUG_ON(!ext4_handle_valid(handle)); |
99b150d8 | 6324 | BUG_ON(e4b->bd_bitmap_folio == NULL); |
5eea586b | 6325 | BUG_ON(e4b->bd_buddy_folio == NULL); |
c9de560d | 6326 | |
18aadd47 BJ |
6327 | new_node = &new_entry->efd_node; |
6328 | cluster = new_entry->efd_start_cluster; | |
c894058d | 6329 | |
c894058d AK |
6330 | if (!*n) { |
6331 | /* first free block exent. We need to | |
6332 | protect buddy cache from being freed, | |
6333 | * otherwise we'll refresh it from | |
6334 | * on-disk bitmap and lose not-yet-available | |
6335 | * blocks */ | |
5eea586b | 6336 | folio_get(e4b->bd_buddy_folio); |
99b150d8 | 6337 | folio_get(e4b->bd_bitmap_folio); |
c894058d AK |
6338 | } |
6339 | while (*n) { | |
6340 | parent = *n; | |
18aadd47 BJ |
6341 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
6342 | if (cluster < entry->efd_start_cluster) | |
c894058d | 6343 | n = &(*n)->rb_left; |
18aadd47 | 6344 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
6345 | n = &(*n)->rb_right; |
6346 | else { | |
e29136f8 | 6347 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
6348 | ext4_group_first_block_no(sb, group) + |
6349 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 6350 | "Block already on to-be-freed list"); |
cca41553 | 6351 | kmem_cache_free(ext4_free_data_cachep, new_entry); |
85b67ffb | 6352 | return; |
c9de560d | 6353 | } |
c894058d | 6354 | } |
c9de560d | 6355 | |
c894058d AK |
6356 | rb_link_node(new_node, parent, n); |
6357 | rb_insert_color(new_node, &db->bb_free_root); | |
6358 | ||
6359 | /* Now try to see the extent can be merged to left and right */ | |
6360 | node = rb_prev(new_node); | |
6361 | if (node) { | |
18aadd47 | 6362 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
6363 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
6364 | &(db->bb_free_root)); | |
c894058d | 6365 | } |
c9de560d | 6366 | |
c894058d AK |
6367 | node = rb_next(new_node); |
6368 | if (node) { | |
18aadd47 | 6369 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
a0154344 DJ |
6370 | ext4_try_merge_freed_extent(sbi, entry, new_entry, |
6371 | &(db->bb_free_root)); | |
c9de560d | 6372 | } |
a0154344 | 6373 | |
d08854f5 | 6374 | spin_lock(&sbi->s_md_lock); |
ce774e53 | 6375 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]); |
d08854f5 TT |
6376 | sbi->s_mb_free_pending += clusters; |
6377 | spin_unlock(&sbi->s_md_lock); | |
c9de560d AT |
6378 | } |
6379 | ||
8016e29f HS |
6380 | static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, |
6381 | unsigned long count) | |
6382 | { | |
8016e29f | 6383 | struct super_block *sb = inode->i_sb; |
8016e29f HS |
6384 | ext4_group_t group; |
6385 | ext4_grpblk_t blkoff; | |
8016e29f HS |
6386 | |
6387 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff); | |
c431d386 KS |
6388 | ext4_mb_mark_context(NULL, sb, false, group, blkoff, count, |
6389 | EXT4_MB_BITMAP_MARKED_CHECK | | |
6390 | EXT4_MB_SYNC_UPDATE, | |
6391 | NULL); | |
8016e29f HS |
6392 | } |
6393 | ||
44338711 | 6394 | /** |
8ac3939d RH |
6395 | * ext4_mb_clear_bb() -- helper function for freeing blocks. |
6396 | * Used by ext4_free_blocks() | |
44338711 TT |
6397 | * @handle: handle for this transaction |
6398 | * @inode: inode | |
c60990b3 TT |
6399 | * @block: starting physical block to be freed |
6400 | * @count: number of blocks to be freed | |
5def1360 | 6401 | * @flags: flags used by ext4_free_blocks |
c9de560d | 6402 | */ |
8ac3939d RH |
6403 | static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, |
6404 | ext4_fsblk_t block, unsigned long count, | |
6405 | int flags) | |
c9de560d | 6406 | { |
c9de560d | 6407 | struct super_block *sb = inode->i_sb; |
5354b2af | 6408 | struct ext4_group_info *grp; |
498e5f24 | 6409 | unsigned int overflow; |
c9de560d | 6410 | ext4_grpblk_t bit; |
c9de560d AT |
6411 | ext4_group_t block_group; |
6412 | struct ext4_sb_info *sbi; | |
6413 | struct ext4_buddy e4b; | |
84130193 | 6414 | unsigned int count_clusters; |
c9de560d | 6415 | int err = 0; |
38b8f70c KS |
6416 | int mark_flags = 0; |
6417 | ext4_grpblk_t changed; | |
c9de560d | 6418 | |
8016e29f HS |
6419 | sbi = EXT4_SB(sb); |
6420 | ||
1e1c2b86 LC |
6421 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
6422 | !ext4_inode_block_valid(inode, block, count)) { | |
6423 | ext4_error(sb, "Freeing blocks in system zone - " | |
6424 | "Block = %llu, count = %lu", block, count); | |
6425 | /* err = 0. ext4_std_error should be a no op */ | |
33e728c6 | 6426 | goto error_out; |
1e1c2b86 LC |
6427 | } |
6428 | flags |= EXT4_FREE_BLOCKS_VALIDATED; | |
6429 | ||
c9de560d AT |
6430 | do_more: |
6431 | overflow = 0; | |
6432 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
6433 | ||
5354b2af TT |
6434 | grp = ext4_get_group_info(sb, block_group); |
6435 | if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) | |
163a203d DW |
6436 | return; |
6437 | ||
c9de560d AT |
6438 | /* |
6439 | * Check to see if we are freeing blocks across a group | |
6440 | * boundary. | |
6441 | */ | |
84130193 TT |
6442 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
6443 | overflow = EXT4_C2B(sbi, bit) + count - | |
6444 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d | 6445 | count -= overflow; |
1e1c2b86 LC |
6446 | /* The range changed so it's no longer validated */ |
6447 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d | 6448 | } |
810da240 | 6449 | count_clusters = EXT4_NUM_B2C(sbi, count); |
33e728c6 KS |
6450 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
6451 | ||
6452 | /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ | |
6453 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, | |
6454 | GFP_NOFS|__GFP_NOFAIL); | |
6455 | if (err) | |
6456 | goto error_out; | |
6457 | ||
6458 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && | |
6459 | !ext4_inode_block_valid(inode, block, count)) { | |
6460 | ext4_error(sb, "Freeing blocks in system zone - " | |
6461 | "Block = %llu, count = %lu", block, count); | |
6462 | /* err = 0. ext4_std_error should be a no op */ | |
6463 | goto error_clean; | |
6464 | } | |
6465 | ||
c9de560d | 6466 | #ifdef AGGRESSIVE_CHECK |
38b8f70c | 6467 | mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK; |
c9de560d | 6468 | #endif |
38b8f70c KS |
6469 | err = ext4_mb_mark_context(handle, sb, false, block_group, bit, |
6470 | count_clusters, mark_flags, &changed); | |
c9de560d | 6471 | |
33e728c6 | 6472 | |
38b8f70c KS |
6473 | if (err && changed == 0) |
6474 | goto error_clean; | |
33e728c6 | 6475 | |
38b8f70c KS |
6476 | #ifdef AGGRESSIVE_CHECK |
6477 | BUG_ON(changed != count_clusters); | |
6478 | #endif | |
e6362609 | 6479 | |
f96c450d DJ |
6480 | /* |
6481 | * We need to make sure we don't reuse the freed block until after the | |
6482 | * transaction is committed. We make an exception if the inode is to be | |
6483 | * written in writeback mode since writeback mode has weak data | |
6484 | * consistency guarantees. | |
6485 | */ | |
6486 | if (ext4_handle_valid(handle) && | |
6487 | ((flags & EXT4_FREE_BLOCKS_METADATA) || | |
6488 | !ext4_should_writeback_data(inode))) { | |
7a2fcbf7 AK |
6489 | struct ext4_free_data *new_entry; |
6490 | /* | |
7444a072 MH |
6491 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
6492 | * to fail. | |
7a2fcbf7 | 6493 | */ |
7444a072 MH |
6494 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
6495 | GFP_NOFS|__GFP_NOFAIL); | |
18aadd47 BJ |
6496 | new_entry->efd_start_cluster = bit; |
6497 | new_entry->efd_group = block_group; | |
6498 | new_entry->efd_count = count_clusters; | |
6499 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 6500 | |
7a2fcbf7 | 6501 | ext4_lock_group(sb, block_group); |
7a2fcbf7 | 6502 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 6503 | } else { |
d71c1ae2 | 6504 | if (test_opt(sb, DISCARD)) { |
247c3d21 | 6505 | err = ext4_issue_discard(sb, block_group, bit, |
0efcd739 | 6506 | count_clusters); |
fa606293 JK |
6507 | /* |
6508 | * Ignore EOPNOTSUPP error. This is consistent with | |
6509 | * what happens when using journal. | |
6510 | */ | |
6511 | if (err == -EOPNOTSUPP) | |
6512 | err = 0; | |
6513 | if (err) | |
d71c1ae2 | 6514 | ext4_msg(sb, KERN_WARNING, "discard request in" |
a00b482b | 6515 | " group:%u block:%d count:%lu failed" |
d71c1ae2 LC |
6516 | " with %d", block_group, bit, count, |
6517 | err); | |
8f9ff189 LC |
6518 | } else |
6519 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 6520 | |
955ce5f5 | 6521 | ext4_lock_group(sb, block_group); |
84130193 | 6522 | mb_free_blocks(inode, &e4b, bit, count_clusters); |
c9de560d AT |
6523 | } |
6524 | ||
955ce5f5 | 6525 | ext4_unlock_group(sb, block_group); |
c9de560d | 6526 | |
9fe67149 EW |
6527 | /* |
6528 | * on a bigalloc file system, defer the s_freeclusters_counter | |
6529 | * update to the caller (ext4_remove_space and friends) so they | |
6530 | * can determine if a cluster freed here should be rereserved | |
6531 | */ | |
6532 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { | |
6533 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) | |
6534 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); | |
6535 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
6536 | count_clusters); | |
6537 | } | |
7d734532 | 6538 | |
c9de560d AT |
6539 | if (overflow && !err) { |
6540 | block += count; | |
6541 | count = overflow; | |
33e728c6 | 6542 | ext4_mb_unload_buddy(&e4b); |
1e1c2b86 LC |
6543 | /* The range changed so it's no longer validated */ |
6544 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
c9de560d AT |
6545 | goto do_more; |
6546 | } | |
33e728c6 KS |
6547 | |
6548 | error_clean: | |
6549 | ext4_mb_unload_buddy(&e4b); | |
33e728c6 | 6550 | error_out: |
c9de560d | 6551 | ext4_std_error(sb, err); |
c9de560d | 6552 | } |
7360d173 | 6553 | |
8ac3939d RH |
6554 | /** |
6555 | * ext4_free_blocks() -- Free given blocks and update quota | |
6556 | * @handle: handle for this transaction | |
6557 | * @inode: inode | |
6558 | * @bh: optional buffer of the block to be freed | |
6559 | * @block: starting physical block to be freed | |
6560 | * @count: number of blocks to be freed | |
6561 | * @flags: flags used by ext4_free_blocks | |
6562 | */ | |
6563 | void ext4_free_blocks(handle_t *handle, struct inode *inode, | |
6564 | struct buffer_head *bh, ext4_fsblk_t block, | |
6565 | unsigned long count, int flags) | |
6566 | { | |
6567 | struct super_block *sb = inode->i_sb; | |
6568 | unsigned int overflow; | |
6569 | struct ext4_sb_info *sbi; | |
6570 | ||
6571 | sbi = EXT4_SB(sb); | |
6572 | ||
8ac3939d RH |
6573 | if (bh) { |
6574 | if (block) | |
6575 | BUG_ON(block != bh->b_blocknr); | |
6576 | else | |
6577 | block = bh->b_blocknr; | |
6578 | } | |
6579 | ||
11b6890b | 6580 | if (sbi->s_mount_state & EXT4_FC_REPLAY) { |
2ec6d0a5 | 6581 | ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count)); |
11b6890b KS |
6582 | return; |
6583 | } | |
6584 | ||
6585 | might_sleep(); | |
6586 | ||
8ac3939d RH |
6587 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
6588 | !ext4_inode_block_valid(inode, block, count)) { | |
6589 | ext4_error(sb, "Freeing blocks not in datazone - " | |
6590 | "block = %llu, count = %lu", block, count); | |
6591 | return; | |
6592 | } | |
1e1c2b86 | 6593 | flags |= EXT4_FREE_BLOCKS_VALIDATED; |
8ac3939d RH |
6594 | |
6595 | ext4_debug("freeing block %llu\n", block); | |
6596 | trace_ext4_free_blocks(inode, block, count, flags); | |
6597 | ||
6598 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6599 | BUG_ON(count > 1); | |
6600 | ||
6601 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, | |
6602 | inode, bh, block); | |
6603 | } | |
6604 | ||
6605 | /* | |
6606 | * If the extent to be freed does not begin on a cluster | |
6607 | * boundary, we need to deal with partial clusters at the | |
6608 | * beginning and end of the extent. Normally we will free | |
6609 | * blocks at the beginning or the end unless we are explicitly | |
6610 | * requested to avoid doing so. | |
6611 | */ | |
6612 | overflow = EXT4_PBLK_COFF(sbi, block); | |
6613 | if (overflow) { | |
6614 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
6615 | overflow = sbi->s_cluster_ratio - overflow; | |
6616 | block += overflow; | |
6617 | if (count > overflow) | |
6618 | count -= overflow; | |
6619 | else | |
6620 | return; | |
6621 | } else { | |
6622 | block -= overflow; | |
6623 | count += overflow; | |
6624 | } | |
1e1c2b86 LC |
6625 | /* The range changed so it's no longer validated */ |
6626 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6627 | } |
6628 | overflow = EXT4_LBLK_COFF(sbi, count); | |
6629 | if (overflow) { | |
6630 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
6631 | if (count > overflow) | |
6632 | count -= overflow; | |
6633 | else | |
6634 | return; | |
6635 | } else | |
6636 | count += sbi->s_cluster_ratio - overflow; | |
1e1c2b86 LC |
6637 | /* The range changed so it's no longer validated */ |
6638 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED; | |
8ac3939d RH |
6639 | } |
6640 | ||
6641 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { | |
6642 | int i; | |
6643 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; | |
6644 | ||
6645 | for (i = 0; i < count; i++) { | |
6646 | cond_resched(); | |
6647 | if (is_metadata) | |
6648 | bh = sb_find_get_block(inode->i_sb, block + i); | |
6649 | ext4_forget(handle, is_metadata, inode, bh, block + i); | |
6650 | } | |
6651 | } | |
6652 | ||
6653 | ext4_mb_clear_bb(handle, inode, block, count, flags); | |
8ac3939d RH |
6654 | } |
6655 | ||
2846e820 | 6656 | /** |
0529155e | 6657 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
6658 | * @handle: handle to this transaction |
6659 | * @sb: super block | |
4907cb7b | 6660 | * @block: start physical block to add to the block group |
2846e820 AG |
6661 | * @count: number of blocks to free |
6662 | * | |
e73a347b | 6663 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 6664 | */ |
cc7365df | 6665 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
6666 | ext4_fsblk_t block, unsigned long count) |
6667 | { | |
2846e820 AG |
6668 | ext4_group_t block_group; |
6669 | ext4_grpblk_t bit; | |
2846e820 | 6670 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
e73a347b | 6671 | struct ext4_buddy e4b; |
5c657db4 | 6672 | int err = 0; |
d77147ff | 6673 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); |
6674 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); | |
6675 | unsigned long cluster_count = last_cluster - first_cluster + 1; | |
5c657db4 | 6676 | ext4_grpblk_t changed; |
2846e820 AG |
6677 | |
6678 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
6679 | ||
5c657db4 | 6680 | if (cluster_count == 0) |
4740b830 YY |
6681 | return 0; |
6682 | ||
2846e820 | 6683 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
6684 | /* |
6685 | * Check to see if we are freeing blocks across a group | |
6686 | * boundary. | |
6687 | */ | |
d77147ff | 6688 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
6689 | ext4_warning(sb, "too many blocks added to group %u", | |
cc7365df YY |
6690 | block_group); |
6691 | err = -EINVAL; | |
03c7fc39 KS |
6692 | goto error_out; |
6693 | } | |
6694 | ||
6695 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
6696 | if (err) | |
6697 | goto error_out; | |
6698 | ||
6699 | if (!ext4_sb_block_valid(sb, NULL, block, count)) { | |
6700 | ext4_error(sb, "Adding blocks in system zones - " | |
6701 | "Block = %llu, count = %lu", | |
6702 | block, count); | |
6703 | err = -EINVAL; | |
6704 | goto error_clean; | |
cc7365df | 6705 | } |
2cd05cc3 | 6706 | |
5c657db4 KS |
6707 | err = ext4_mb_mark_context(handle, sb, false, block_group, bit, |
6708 | cluster_count, EXT4_MB_BITMAP_MARKED_CHECK, | |
6709 | &changed); | |
6710 | if (err && changed == 0) | |
03c7fc39 | 6711 | goto error_clean; |
e73a347b | 6712 | |
5c657db4 KS |
6713 | if (changed != cluster_count) |
6714 | ext4_error(sb, "bit already cleared in group %u", block_group); | |
2846e820 | 6715 | |
03c7fc39 KS |
6716 | ext4_lock_group(sb, block_group); |
6717 | mb_free_blocks(NULL, &e4b, bit, cluster_count); | |
6718 | ext4_unlock_group(sb, block_group); | |
6719 | percpu_counter_add(&sbi->s_freeclusters_counter, | |
5c657db4 | 6720 | changed); |
03c7fc39 KS |
6721 | |
6722 | error_clean: | |
03c7fc39 KS |
6723 | ext4_mb_unload_buddy(&e4b); |
6724 | error_out: | |
2846e820 | 6725 | ext4_std_error(sb, err); |
cc7365df | 6726 | return err; |
2846e820 AG |
6727 | } |
6728 | ||
7360d173 LC |
6729 | /** |
6730 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
6731 | * @sb: super block for the file system | |
6732 | * @start: starting block of the free extent in the alloc. group | |
6733 | * @count: number of blocks to TRIM | |
7360d173 LC |
6734 | * @e4b: ext4 buddy for the group |
6735 | * | |
6736 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
6737 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
6738 | * be called with under the group lock. | |
6739 | */ | |
bd2eea8d WJ |
6740 | static int ext4_trim_extent(struct super_block *sb, |
6741 | int start, int count, struct ext4_buddy *e4b) | |
e2cbd587 | 6742 | __releases(bitlock) |
6743 | __acquires(bitlock) | |
7360d173 LC |
6744 | { |
6745 | struct ext4_free_extent ex; | |
bd2eea8d | 6746 | ext4_group_t group = e4b->bd_group; |
d71c1ae2 | 6747 | int ret = 0; |
7360d173 | 6748 | |
b3d4c2b1 TM |
6749 | trace_ext4_trim_extent(sb, group, start, count); |
6750 | ||
7360d173 LC |
6751 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
6752 | ||
6753 | ex.fe_start = start; | |
6754 | ex.fe_group = group; | |
6755 | ex.fe_len = count; | |
6756 | ||
6757 | /* | |
6758 | * Mark blocks used, so no one can reuse them while | |
6759 | * being trimmed. | |
6760 | */ | |
6761 | mb_mark_used(e4b, &ex); | |
6762 | ext4_unlock_group(sb, group); | |
0efcd739 | 6763 | ret = ext4_issue_discard(sb, group, start, count); |
7360d173 LC |
6764 | ext4_lock_group(sb, group); |
6765 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 6766 | return ret; |
7360d173 LC |
6767 | } |
6768 | ||
45e4ab32 JK |
6769 | static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb, |
6770 | ext4_group_t grp) | |
6771 | { | |
7c784d62 SJS |
6772 | unsigned long nr_clusters_in_group; |
6773 | ||
6774 | if (grp < (ext4_get_groups_count(sb) - 1)) | |
6775 | nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb); | |
6776 | else | |
6777 | nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) - | |
6778 | ext4_group_first_block_no(sb, grp)) | |
6779 | >> EXT4_CLUSTER_BITS(sb); | |
6780 | ||
6781 | return nr_clusters_in_group - 1; | |
45e4ab32 JK |
6782 | } |
6783 | ||
5229a658 JK |
6784 | static bool ext4_trim_interrupted(void) |
6785 | { | |
6786 | return fatal_signal_pending(current) || freezing(current); | |
6787 | } | |
6788 | ||
6920b391 WJ |
6789 | static int ext4_try_to_trim_range(struct super_block *sb, |
6790 | struct ext4_buddy *e4b, ext4_grpblk_t start, | |
6791 | ext4_grpblk_t max, ext4_grpblk_t minblocks) | |
a5fda113 TT |
6792 | __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) |
6793 | __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) | |
6920b391 | 6794 | { |
68da4c44 | 6795 | ext4_grpblk_t next, count, free_count, last, origin_start; |
45e4ab32 | 6796 | bool set_trimmed = false; |
6920b391 | 6797 | void *bitmap; |
6920b391 | 6798 | |
17220215 BL |
6799 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
6800 | return 0; | |
6801 | ||
68da4c44 | 6802 | last = ext4_last_grp_cluster(sb, e4b->bd_group); |
6920b391 | 6803 | bitmap = e4b->bd_bitmap; |
68da4c44 | 6804 | if (start == 0 && max >= last) |
45e4ab32 | 6805 | set_trimmed = true; |
68da4c44 | 6806 | origin_start = start; |
de8bf0e5 | 6807 | start = max(e4b->bd_info->bb_first_free, start); |
6920b391 WJ |
6808 | count = 0; |
6809 | free_count = 0; | |
6810 | ||
6811 | while (start <= max) { | |
6812 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
6813 | if (start > max) | |
6814 | break; | |
68da4c44 YB |
6815 | |
6816 | next = mb_find_next_bit(bitmap, last + 1, start); | |
6817 | if (origin_start == 0 && next >= last) | |
6818 | set_trimmed = true; | |
6920b391 WJ |
6819 | |
6820 | if ((next - start) >= minblocks) { | |
afcc4e32 LB |
6821 | int ret = ext4_trim_extent(sb, start, next - start, e4b); |
6822 | ||
6920b391 | 6823 | if (ret && ret != -EOPNOTSUPP) |
45e4ab32 | 6824 | return count; |
6920b391 WJ |
6825 | count += next - start; |
6826 | } | |
6827 | free_count += next - start; | |
6828 | start = next + 1; | |
6829 | ||
5229a658 JK |
6830 | if (ext4_trim_interrupted()) |
6831 | return count; | |
6920b391 WJ |
6832 | |
6833 | if (need_resched()) { | |
6834 | ext4_unlock_group(sb, e4b->bd_group); | |
6835 | cond_resched(); | |
6836 | ext4_lock_group(sb, e4b->bd_group); | |
6837 | } | |
6838 | ||
6839 | if ((e4b->bd_info->bb_free - free_count) < minblocks) | |
6840 | break; | |
6841 | } | |
6842 | ||
45e4ab32 JK |
6843 | if (set_trimmed) |
6844 | EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info); | |
6845 | ||
6920b391 WJ |
6846 | return count; |
6847 | } | |
6848 | ||
7360d173 LC |
6849 | /** |
6850 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
6851 | * @sb: super block for file system | |
22612283 | 6852 | * @group: group to be trimmed |
7360d173 LC |
6853 | * @start: first group block to examine |
6854 | * @max: last group block to examine | |
6855 | * @minblocks: minimum extent block count | |
6856 | * | |
7360d173 LC |
6857 | * ext4_trim_all_free walks through group's block bitmap searching for free |
6858 | * extents. When the free extent is found, mark it as used in group buddy | |
6859 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
b6f5558c | 6860 | * the group buddy bitmap. |
7360d173 | 6861 | */ |
0b75a840 | 6862 | static ext4_grpblk_t |
78944086 LC |
6863 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
6864 | ext4_grpblk_t start, ext4_grpblk_t max, | |
45e4ab32 | 6865 | ext4_grpblk_t minblocks) |
7360d173 | 6866 | { |
78944086 | 6867 | struct ext4_buddy e4b; |
6920b391 | 6868 | int ret; |
7360d173 | 6869 | |
b3d4c2b1 TM |
6870 | trace_ext4_trim_all_free(sb, group, start, max); |
6871 | ||
78944086 LC |
6872 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
6873 | if (ret) { | |
9651e6b2 KK |
6874 | ext4_warning(sb, "Error %d loading buddy information for %u", |
6875 | ret, group); | |
78944086 LC |
6876 | return ret; |
6877 | } | |
28739eea LC |
6878 | |
6879 | ext4_lock_group(sb, group); | |
7360d173 | 6880 | |
6920b391 | 6881 | if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || |
45e4ab32 | 6882 | minblocks < EXT4_SB(sb)->s_last_trim_minblks) |
6920b391 | 6883 | ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); |
45e4ab32 | 6884 | else |
6920b391 | 6885 | ret = 0; |
3d56b8d2 | 6886 | |
7360d173 | 6887 | ext4_unlock_group(sb, group); |
78944086 | 6888 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
6889 | |
6890 | ext4_debug("trimmed %d blocks in the group %d\n", | |
6920b391 | 6891 | ret, group); |
7360d173 | 6892 | |
d71c1ae2 | 6893 | return ret; |
7360d173 LC |
6894 | } |
6895 | ||
6896 | /** | |
6897 | * ext4_trim_fs() -- trim ioctl handle function | |
6898 | * @sb: superblock for filesystem | |
6899 | * @range: fstrim_range structure | |
6900 | * | |
6901 | * start: First Byte to trim | |
6902 | * len: number of Bytes to trim from start | |
6903 | * minlen: minimum extent length in Bytes | |
6904 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
6905 | * start to start+len. For each such a group ext4_trim_all_free function | |
6906 | * is invoked to trim all free space. | |
6907 | */ | |
6908 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
6909 | { | |
7b47ef52 | 6910 | unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); |
78944086 | 6911 | struct ext4_group_info *grp; |
913eed83 | 6912 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 6913 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 6914 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
6915 | ext4_fsblk_t first_data_blk = |
6916 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 6917 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
6918 | int ret = 0; |
6919 | ||
6920 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 6921 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
6922 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
6923 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 6924 | |
5de35e8d LC |
6925 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
6926 | start >= max_blks || | |
6927 | range->len < sb->s_blocksize) | |
7360d173 | 6928 | return -EINVAL; |
173b6e38 | 6929 | /* No point to try to trim less than discard granularity */ |
7b47ef52 | 6930 | if (range->minlen < discard_granularity) { |
173b6e38 | 6931 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
7b47ef52 | 6932 | discard_granularity >> sb->s_blocksize_bits); |
173b6e38 JK |
6933 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) |
6934 | goto out; | |
6935 | } | |
45e4ab32 | 6936 | if (end >= max_blks - 1) |
913eed83 LC |
6937 | end = max_blks - 1; |
6938 | if (end <= first_data_blk) | |
22f10457 | 6939 | goto out; |
913eed83 | 6940 | if (start < first_data_blk) |
0f0a25bf | 6941 | start = first_data_blk; |
7360d173 | 6942 | |
913eed83 | 6943 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 6944 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 6945 | &first_group, &first_cluster); |
913eed83 | 6946 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 6947 | &last_group, &last_cluster); |
7360d173 | 6948 | |
913eed83 LC |
6949 | /* end now represents the last cluster to discard in this group */ |
6950 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
6951 | |
6952 | for (group = first_group; group <= last_group; group++) { | |
5229a658 JK |
6953 | if (ext4_trim_interrupted()) |
6954 | break; | |
78944086 | 6955 | grp = ext4_get_group_info(sb, group); |
5354b2af TT |
6956 | if (!grp) |
6957 | continue; | |
78944086 LC |
6958 | /* We only do this if the grp has never been initialized */ |
6959 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
adb7ef60 | 6960 | ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
78944086 LC |
6961 | if (ret) |
6962 | break; | |
7360d173 LC |
6963 | } |
6964 | ||
0ba08517 | 6965 | /* |
913eed83 LC |
6966 | * For all the groups except the last one, last cluster will |
6967 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
6968 | * change it for the last group, note that last_cluster is | |
6969 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 6970 | */ |
45e4ab32 | 6971 | if (group == last_group) |
913eed83 | 6972 | end = last_cluster; |
78944086 | 6973 | if (grp->bb_free >= minlen) { |
7137d7a4 | 6974 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
45e4ab32 | 6975 | end, minlen); |
7360d173 LC |
6976 | if (cnt < 0) { |
6977 | ret = cnt; | |
7360d173 LC |
6978 | break; |
6979 | } | |
21e7fd22 | 6980 | trimmed += cnt; |
7360d173 | 6981 | } |
913eed83 LC |
6982 | |
6983 | /* | |
6984 | * For every group except the first one, we are sure | |
6985 | * that the first cluster to discard will be cluster #0. | |
6986 | */ | |
7137d7a4 | 6987 | first_cluster = 0; |
7360d173 | 6988 | } |
7360d173 | 6989 | |
3d56b8d2 | 6990 | if (!ret) |
2327fb2e | 6991 | EXT4_SB(sb)->s_last_trim_minblks = minlen; |
3d56b8d2 | 6992 | |
22f10457 | 6993 | out: |
aaf7d73e | 6994 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
6995 | return ret; |
6996 | } | |
0c9ec4be DW |
6997 | |
6998 | /* Iterate all the free extents in the group. */ | |
6999 | int | |
7000 | ext4_mballoc_query_range( | |
7001 | struct super_block *sb, | |
7002 | ext4_group_t group, | |
7003 | ext4_grpblk_t start, | |
7004 | ext4_grpblk_t end, | |
7005 | ext4_mballoc_query_range_fn formatter, | |
7006 | void *priv) | |
7007 | { | |
7008 | void *bitmap; | |
7009 | ext4_grpblk_t next; | |
7010 | struct ext4_buddy e4b; | |
7011 | int error; | |
7012 | ||
7013 | error = ext4_mb_load_buddy(sb, group, &e4b); | |
7014 | if (error) | |
7015 | return error; | |
7016 | bitmap = e4b.bd_bitmap; | |
7017 | ||
7018 | ext4_lock_group(sb, group); | |
7019 | ||
de8bf0e5 | 7020 | start = max(e4b.bd_info->bb_first_free, start); |
0c9ec4be DW |
7021 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) |
7022 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7023 | ||
7024 | while (start <= end) { | |
7025 | start = mb_find_next_zero_bit(bitmap, end + 1, start); | |
7026 | if (start > end) | |
7027 | break; | |
7028 | next = mb_find_next_bit(bitmap, end + 1, start); | |
7029 | ||
7030 | ext4_unlock_group(sb, group); | |
7031 | error = formatter(sb, group, start, next - start, priv); | |
7032 | if (error) | |
7033 | goto out_unload; | |
7034 | ext4_lock_group(sb, group); | |
7035 | ||
7036 | start = next + 1; | |
7037 | } | |
7038 | ||
7039 | ext4_unlock_group(sb, group); | |
7040 | out_unload: | |
7041 | ext4_mb_unload_buddy(&e4b); | |
7042 | ||
7043 | return error; | |
7044 | } | |
7c9fa399 KS |
7045 | |
7046 | #ifdef CONFIG_EXT4_KUNIT_TESTS | |
7047 | #include "mballoc-test.c" | |
7048 | #endif |