| 1 |
// SPDX-License-Identifier: GPL-2.0 |
| 2 |
/* |
| 3 |
* Copyright (c) 2003-2006, Cluster File Systems, Inc, [email protected] |
| 4 |
* Written by Alex Tomas <[email protected]> |
| 5 |
*/ |
| 6 |
|
| 7 |
|
| 8 |
/* |
| 9 |
* mballoc.c contains the multiblocks allocation routines |
| 10 |
*/ |
| 11 |
|
| 12 |
#include "ext4_jbd2.h" |
| 13 |
#include "mballoc.h" |
| 14 |
#include <linux/log2.h> |
| 15 |
#include <linux/module.h> |
| 16 |
#include <linux/slab.h> |
| 17 |
#include <linux/nospec.h> |
| 18 |
#include <linux/backing-dev.h> |
| 19 |
#include <linux/freezer.h> |
| 20 |
#include <trace/events/ext4.h> |
| 21 |
#include <kunit/static_stub.h> |
| 22 |
|
| 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 |
* |
| 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. |
| 56 |
* |
| 57 |
* The main motivation for having small file use group preallocation is to |
| 58 |
* ensure that we have small files closer together on the disk. |
| 59 |
* |
| 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: |
| 64 |
* |
| 65 |
* pa_lstart -> the logical start block for this prealloc space |
| 66 |
* pa_pstart -> the physical start block for this prealloc space |
| 67 |
* pa_len -> length for this prealloc space (in clusters) |
| 68 |
* pa_free -> free space available in this prealloc space (in clusters) |
| 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 |
| 72 |
* space we will consume the particular prealloc space. This makes sure that |
| 73 |
* we have contiguous physical blocks representing the file blocks |
| 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 |
| 81 |
* prealloc space. These are per CPU prealloc list represented as |
| 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 |
| 89 |
* enough free space (pa_free) within the prealloc space. |
| 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 |
* { folio } |
| 102 |
* [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
| 103 |
* |
| 104 |
* |
| 105 |
* one block each for bitmap and buddy information. So for each group we |
| 106 |
* take up 2 blocks. A folio can contain blocks_per_folio (folio_size / |
| 107 |
* blocksize) blocks. So it can have information regarding groups_per_folio |
| 108 |
* which is blocks_per_folio/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 |
| 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 |
| 125 |
* 512 blocks. This can be tuned via |
| 126 |
* /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
| 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 |
| 129 |
* smallest multiple of the stripe value (sbi->s_stripe) which is |
| 130 |
* greater than the default mb_group_prealloc. |
| 131 |
* |
| 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 xarrays (sbi->s_mb_largest_free_orders) |
| 136 |
* |
| 137 |
* Locking: Writers use xa_lock, readers use rcu_read_lock. |
| 138 |
* |
| 139 |
* This is an array of xarrays where the index in the array represents the |
| 140 |
* largest free order in the buddy bitmap of the participating group infos of |
| 141 |
* that xarray. So, there are exactly MB_NUM_ORDERS(sb) (which means total |
| 142 |
* number of buddy bitmap orders possible) number of xarrays. Group-infos are |
| 143 |
* placed in appropriate xarrays. |
| 144 |
* |
| 145 |
* 2) Average fragment size xarrays (sbi->s_mb_avg_fragment_size) |
| 146 |
* |
| 147 |
* Locking: Writers use xa_lock, readers use rcu_read_lock. |
| 148 |
* |
| 149 |
* This is an array of xarrays where in the i-th xarray 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 xarray for completely empty |
| 153 |
* groups so we only have MB_NUM_ORDERS(sb) xarrays. Group-infos are placed |
| 154 |
* in appropriate xarrays. |
| 155 |
* |
| 156 |
* In xarray, the index is the block group number, the value is the block group |
| 157 |
* information, and a non-empty value indicates the block group is present in |
| 158 |
* the current xarray. |
| 159 |
* |
| 160 |
* When "mb_optimize_scan" mount option is set, mballoc consults the above data |
| 161 |
* structures to decide the order in which groups are to be traversed for |
| 162 |
* fulfilling an allocation request. |
| 163 |
* |
| 164 |
* At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order |
| 165 |
* >= the order of the request. We directly look at the largest free order list |
| 166 |
* in the data structure (1) above where largest_free_order = order of the |
| 167 |
* request. If that list is empty, we look at remaining list in the increasing |
| 168 |
* order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED |
| 169 |
* lookup in O(1) time. |
| 170 |
* |
| 171 |
* At CR_GOAL_LEN_FAST, we only consider groups where |
| 172 |
* average fragment size > request size. So, we lookup a group which has average |
| 173 |
* fragment size just above or equal to request size using our average fragment |
| 174 |
* size group lists (data structure 2) in O(1) time. |
| 175 |
* |
| 176 |
* At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied |
| 177 |
* in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in |
| 178 |
* CR_GOAL_LEN_FAST suggests that there is no BG that has avg |
| 179 |
* fragment size > goal length. So before falling to the slower |
| 180 |
* CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and |
| 181 |
* then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big |
| 182 |
* enough average fragment size. This increases the chances of finding a |
| 183 |
* suitable block group in O(1) time and results in faster allocation at the |
| 184 |
* cost of reduced size of allocation. |
| 185 |
* |
| 186 |
* If "mb_optimize_scan" mount option is not set, mballoc traverses groups in |
| 187 |
* linear order which requires O(N) search time for each CR_POWER2_ALIGNED and |
| 188 |
* CR_GOAL_LEN_FAST phase. |
| 189 |
* |
| 190 |
* The regular allocator (using the buddy cache) supports a few tunables. |
| 191 |
* |
| 192 |
* /sys/fs/ext4/<partition>/mb_min_to_scan |
| 193 |
* /sys/fs/ext4/<partition>/mb_max_to_scan |
| 194 |
* /sys/fs/ext4/<partition>/mb_order2_req |
| 195 |
* /sys/fs/ext4/<partition>/mb_max_linear_groups |
| 196 |
* |
| 197 |
* The regular allocator uses buddy scan only if the request len is power of |
| 198 |
* 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
| 199 |
* value of s_mb_order2_reqs can be tuned via |
| 200 |
* /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
| 201 |
* stripe size (sbi->s_stripe), we try to search for contiguous block in |
| 202 |
* stripe size. This should result in better allocation on RAID setups. If |
| 203 |
* not, we search in the specific group using bitmap for best extents. The |
| 204 |
* tunable min_to_scan and max_to_scan control the behaviour here. |
| 205 |
* min_to_scan indicate how long the mballoc __must__ look for a best |
| 206 |
* extent and max_to_scan indicates how long the mballoc __can__ look for a |
| 207 |
* best extent in the found extents. Searching for the blocks starts with |
| 208 |
* the group specified as the goal value in allocation context via |
| 209 |
* ac_g_ex. Each group is first checked based on the criteria whether it |
| 210 |
* can be used for allocation. ext4_mb_good_group explains how the groups are |
| 211 |
* checked. |
| 212 |
* |
| 213 |
* When "mb_optimize_scan" is turned on, as mentioned above, the groups may not |
| 214 |
* get traversed linearly. That may result in subsequent allocations being not |
| 215 |
* close to each other. And so, the underlying device may get filled up in a |
| 216 |
* non-linear fashion. While that may not matter on non-rotational devices, for |
| 217 |
* rotational devices that may result in higher seek times. "mb_max_linear_groups" |
| 218 |
* tells mballoc how many groups mballoc should search linearly before |
| 219 |
* performing consulting above data structures for more efficient lookups. For |
| 220 |
* non rotational devices, this value defaults to 0 and for rotational devices |
| 221 |
* this is set to MB_DEFAULT_LINEAR_LIMIT. |
| 222 |
* |
| 223 |
* Both the prealloc space are getting populated as above. So for the first |
| 224 |
* request we will hit the buddy cache which will result in this prealloc |
| 225 |
* space getting filled. The prealloc space is then later used for the |
| 226 |
* subsequent request. |
| 227 |
*/ |
| 228 |
|
| 229 |
/* |
| 230 |
* mballoc operates on the following data: |
| 231 |
* - on-disk bitmap |
| 232 |
* - in-core buddy (actually includes buddy and bitmap) |
| 233 |
* - preallocation descriptors (PAs) |
| 234 |
* |
| 235 |
* there are two types of preallocations: |
| 236 |
* - inode |
| 237 |
* assiged to specific inode and can be used for this inode only. |
| 238 |
* it describes part of inode's space preallocated to specific |
| 239 |
* physical blocks. any block from that preallocated can be used |
| 240 |
* independent. the descriptor just tracks number of blocks left |
| 241 |
* unused. so, before taking some block from descriptor, one must |
| 242 |
* make sure corresponded logical block isn't allocated yet. this |
| 243 |
* also means that freeing any block within descriptor's range |
| 244 |
* must discard all preallocated blocks. |
| 245 |
* - locality group |
| 246 |
* assigned to specific locality group which does not translate to |
| 247 |
* permanent set of inodes: inode can join and leave group. space |
| 248 |
* from this type of preallocation can be used for any inode. thus |
| 249 |
* it's consumed from the beginning to the end. |
| 250 |
* |
| 251 |
* relation between them can be expressed as: |
| 252 |
* in-core buddy = on-disk bitmap + preallocation descriptors |
| 253 |
* |
| 254 |
* this mean blocks mballoc considers used are: |
| 255 |
* - allocated blocks (persistent) |
| 256 |
* - preallocated blocks (non-persistent) |
| 257 |
* |
| 258 |
* consistency in mballoc world means that at any time a block is either |
| 259 |
* free or used in ALL structures. notice: "any time" should not be read |
| 260 |
* literally -- time is discrete and delimited by locks. |
| 261 |
* |
| 262 |
* to keep it simple, we don't use block numbers, instead we count number of |
| 263 |
* blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. |
| 264 |
* |
| 265 |
* all operations can be expressed as: |
| 266 |
* - init buddy: buddy = on-disk + PAs |
| 267 |
* - new PA: buddy += N; PA = N |
| 268 |
* - use inode PA: on-disk += N; PA -= N |
| 269 |
* - discard inode PA buddy -= on-disk - PA; PA = 0 |
| 270 |
* - use locality group PA on-disk += N; PA -= N |
| 271 |
* - discard locality group PA buddy -= PA; PA = 0 |
| 272 |
* note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap |
| 273 |
* is used in real operation because we can't know actual used |
| 274 |
* bits from PA, only from on-disk bitmap |
| 275 |
* |
| 276 |
* if we follow this strict logic, then all operations above should be atomic. |
| 277 |
* given some of them can block, we'd have to use something like semaphores |
| 278 |
* killing performance on high-end SMP hardware. let's try to relax it using |
| 279 |
* the following knowledge: |
| 280 |
* 1) if buddy is referenced, it's already initialized |
| 281 |
* 2) while block is used in buddy and the buddy is referenced, |
| 282 |
* nobody can re-allocate that block |
| 283 |
* 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has |
| 284 |
* bit set and PA claims same block, it's OK. IOW, one can set bit in |
| 285 |
* on-disk bitmap if buddy has same bit set or/and PA covers corresponded |
| 286 |
* block |
| 287 |
* |
| 288 |
* so, now we're building a concurrency table: |
| 289 |
* - init buddy vs. |
| 290 |
* - new PA |
| 291 |
* blocks for PA are allocated in the buddy, buddy must be referenced |
| 292 |
* until PA is linked to allocation group to avoid concurrent buddy init |
| 293 |
* - use inode PA |
| 294 |
* we need to make sure that either on-disk bitmap or PA has uptodate data |
| 295 |
* given (3) we care that PA-=N operation doesn't interfere with init |
| 296 |
* - discard inode PA |
| 297 |
* the simplest way would be to have buddy initialized by the discard |
| 298 |
* - use locality group PA |
| 299 |
* again PA-=N must be serialized with init |
| 300 |
* - discard locality group PA |
| 301 |
* the simplest way would be to have buddy initialized by the discard |
| 302 |
* - new PA vs. |
| 303 |
* - use inode PA |
| 304 |
* i_data_sem serializes them |
| 305 |
* - discard inode PA |
| 306 |
* discard process must wait until PA isn't used by another process |
| 307 |
* - use locality group PA |
| 308 |
* some mutex should serialize them |
| 309 |
* - discard locality group PA |
| 310 |
* discard process must wait until PA isn't used by another process |
| 311 |
* - use inode PA |
| 312 |
* - use inode PA |
| 313 |
* i_data_sem or another mutex should serializes them |
| 314 |
* - discard inode PA |
| 315 |
* discard process must wait until PA isn't used by another process |
| 316 |
* - use locality group PA |
| 317 |
* nothing wrong here -- they're different PAs covering different blocks |
| 318 |
* - discard locality group PA |
| 319 |
* discard process must wait until PA isn't used by another process |
| 320 |
* |
| 321 |
* now we're ready to make few consequences: |
| 322 |
* - PA is referenced and while it is no discard is possible |
| 323 |
* - PA is referenced until block isn't marked in on-disk bitmap |
| 324 |
* - PA changes only after on-disk bitmap |
| 325 |
* - discard must not compete with init. either init is done before |
| 326 |
* any discard or they're serialized somehow |
| 327 |
* - buddy init as sum of on-disk bitmap and PAs is done atomically |
| 328 |
* |
| 329 |
* a special case when we've used PA to emptiness. no need to modify buddy |
| 330 |
* in this case, but we should care about concurrent init |
| 331 |
* |
| 332 |
*/ |
| 333 |
|
| 334 |
/* |
| 335 |
* Logic in few words: |
| 336 |
* |
| 337 |
* - allocation: |
| 338 |
* load group |
| 339 |
* find blocks |
| 340 |
* mark bits in on-disk bitmap |
| 341 |
* release group |
| 342 |
* |
| 343 |
* - use preallocation: |
| 344 |
* find proper PA (per-inode or group) |
| 345 |
* load group |
| 346 |
* mark bits in on-disk bitmap |
| 347 |
* release group |
| 348 |
* release PA |
| 349 |
* |
| 350 |
* - free: |
| 351 |
* load group |
| 352 |
* mark bits in on-disk bitmap |
| 353 |
* release group |
| 354 |
* |
| 355 |
* - discard preallocations in group: |
| 356 |
* mark PAs deleted |
| 357 |
* move them onto local list |
| 358 |
* load on-disk bitmap |
| 359 |
* load group |
| 360 |
* remove PA from object (inode or locality group) |
| 361 |
* mark free blocks in-core |
| 362 |
* |
| 363 |
* - discard inode's preallocations: |
| 364 |
*/ |
| 365 |
|
| 366 |
/* |
| 367 |
* Locking rules |
| 368 |
* |
| 369 |
* Locks: |
| 370 |
* - bitlock on a group (group) |
| 371 |
* - object (inode/locality) (object) |
| 372 |
* - per-pa lock (pa) |
| 373 |
* - cr_power2_aligned lists lock (cr_power2_aligned) |
| 374 |
* - cr_goal_len_fast lists lock (cr_goal_len_fast) |
| 375 |
* |
| 376 |
* Paths: |
| 377 |
* - new pa |
| 378 |
* object |
| 379 |
* group |
| 380 |
* |
| 381 |
* - find and use pa: |
| 382 |
* pa |
| 383 |
* |
| 384 |
* - release consumed pa: |
| 385 |
* pa |
| 386 |
* group |
| 387 |
* object |
| 388 |
* |
| 389 |
* - generate in-core bitmap: |
| 390 |
* group |
| 391 |
* pa |
| 392 |
* |
| 393 |
* - discard all for given object (inode, locality group): |
| 394 |
* object |
| 395 |
* pa |
| 396 |
* group |
| 397 |
* |
| 398 |
* - discard all for given group: |
| 399 |
* group |
| 400 |
* pa |
| 401 |
* group |
| 402 |
* object |
| 403 |
* |
| 404 |
* - allocation path (ext4_mb_regular_allocator) |
| 405 |
* group |
| 406 |
* cr_power2_aligned/cr_goal_len_fast |
| 407 |
*/ |
| 408 |
static struct kmem_cache *ext4_pspace_cachep; |
| 409 |
static struct kmem_cache *ext4_ac_cachep; |
| 410 |
static struct kmem_cache *ext4_free_data_cachep; |
| 411 |
|
| 412 |
/* We create slab caches for groupinfo data structures based on the |
| 413 |
* superblock block size. There will be one per mounted filesystem for |
| 414 |
* each unique s_blocksize_bits */ |
| 415 |
#define NR_GRPINFO_CACHES 8 |
| 416 |
static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
| 417 |
|
| 418 |
static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
| 419 |
"ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", |
| 420 |
"ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", |
| 421 |
"ext4_groupinfo_64k", "ext4_groupinfo_128k" |
| 422 |
}; |
| 423 |
|
| 424 |
static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
| 425 |
ext4_group_t group); |
| 426 |
static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); |
| 427 |
|
| 428 |
static int ext4_mb_scan_group(struct ext4_allocation_context *ac, |
| 429 |
ext4_group_t group); |
| 430 |
|
| 431 |
static int ext4_try_to_trim_range(struct super_block *sb, |
| 432 |
struct ext4_buddy *e4b, ext4_grpblk_t start, |
| 433 |
ext4_grpblk_t max, ext4_grpblk_t minblocks); |
| 434 |
|
| 435 |
/* |
| 436 |
* The algorithm using this percpu seq counter goes below: |
| 437 |
* 1. We sample the percpu discard_pa_seq counter before trying for block |
| 438 |
* allocation in ext4_mb_new_blocks(). |
| 439 |
* 2. We increment this percpu discard_pa_seq counter when we either allocate |
| 440 |
* or free these blocks i.e. while marking those blocks as used/free in |
| 441 |
* mb_mark_used()/mb_free_blocks(). |
| 442 |
* 3. We also increment this percpu seq counter when we successfully identify |
| 443 |
* that the bb_prealloc_list is not empty and hence proceed for discarding |
| 444 |
* of those PAs inside ext4_mb_discard_group_preallocations(). |
| 445 |
* |
| 446 |
* Now to make sure that the regular fast path of block allocation is not |
| 447 |
* affected, as a small optimization we only sample the percpu seq counter |
| 448 |
* on that cpu. Only when the block allocation fails and when freed blocks |
| 449 |
* found were 0, that is when we sample percpu seq counter for all cpus using |
| 450 |
* below function ext4_get_discard_pa_seq_sum(). This happens after making |
| 451 |
* sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. |
| 452 |
*/ |
| 453 |
static DEFINE_PER_CPU(u64, discard_pa_seq); |
| 454 |
static inline u64 ext4_get_discard_pa_seq_sum(void) |
| 455 |
{ |
| 456 |
int __cpu; |
| 457 |
u64 __seq = 0; |
| 458 |
|
| 459 |
for_each_possible_cpu(__cpu) |
| 460 |
__seq += per_cpu(discard_pa_seq, __cpu); |
| 461 |
return __seq; |
| 462 |
} |
| 463 |
|
| 464 |
static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
| 465 |
{ |
| 466 |
#if BITS_PER_LONG == 64 |
| 467 |
*bit += ((unsigned long) addr & 7UL) << 3; |
| 468 |
addr = (void *) ((unsigned long) addr & ~7UL); |
| 469 |
#elif BITS_PER_LONG == 32 |
| 470 |
*bit += ((unsigned long) addr & 3UL) << 3; |
| 471 |
addr = (void *) ((unsigned long) addr & ~3UL); |
| 472 |
#else |
| 473 |
#error "how many bits you are?!" |
| 474 |
#endif |
| 475 |
return addr; |
| 476 |
} |
| 477 |
|
| 478 |
static inline int mb_test_bit(int bit, void *addr) |
| 479 |
{ |
| 480 |
/* |
| 481 |
* ext4_test_bit on architecture like powerpc |
| 482 |
* needs unsigned long aligned address |
| 483 |
*/ |
| 484 |
addr = mb_correct_addr_and_bit(&bit, addr); |
| 485 |
return ext4_test_bit(bit, addr); |
| 486 |
} |
| 487 |
|
| 488 |
static inline void mb_set_bit(int bit, void *addr) |
| 489 |
{ |
| 490 |
addr = mb_correct_addr_and_bit(&bit, addr); |
| 491 |
ext4_set_bit(bit, addr); |
| 492 |
} |
| 493 |
|
| 494 |
static inline void mb_clear_bit(int bit, void *addr) |
| 495 |
{ |
| 496 |
addr = mb_correct_addr_and_bit(&bit, addr); |
| 497 |
ext4_clear_bit(bit, addr); |
| 498 |
} |
| 499 |
|
| 500 |
static inline int mb_test_and_clear_bit(int bit, void *addr) |
| 501 |
{ |
| 502 |
addr = mb_correct_addr_and_bit(&bit, addr); |
| 503 |
return ext4_test_and_clear_bit(bit, addr); |
| 504 |
} |
| 505 |
|
| 506 |
static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
| 507 |
{ |
| 508 |
int fix = 0, ret, tmpmax; |
| 509 |
addr = mb_correct_addr_and_bit(&fix, addr); |
| 510 |
tmpmax = max + fix; |
| 511 |
start += fix; |
| 512 |
|
| 513 |
ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
| 514 |
if (ret > max) |
| 515 |
return max; |
| 516 |
return ret; |
| 517 |
} |
| 518 |
|
| 519 |
static inline int mb_find_next_bit(void *addr, int max, int start) |
| 520 |
{ |
| 521 |
int fix = 0, ret, tmpmax; |
| 522 |
addr = mb_correct_addr_and_bit(&fix, addr); |
| 523 |
tmpmax = max + fix; |
| 524 |
start += fix; |
| 525 |
|
| 526 |
ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
| 527 |
if (ret > max) |
| 528 |
return max; |
| 529 |
return ret; |
| 530 |
} |
| 531 |
|
| 532 |
static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
| 533 |
{ |
| 534 |
char *bb; |
| 535 |
|
| 536 |
BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
| 537 |
BUG_ON(max == NULL); |
| 538 |
|
| 539 |
if (order > e4b->bd_blkbits + 1) { |
| 540 |
*max = 0; |
| 541 |
return NULL; |
| 542 |
} |
| 543 |
|
| 544 |
/* at order 0 we see each particular block */ |
| 545 |
if (order == 0) { |
| 546 |
*max = 1 << (e4b->bd_blkbits + 3); |
| 547 |
return e4b->bd_bitmap; |
| 548 |
} |
| 549 |
|
| 550 |
bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
| 551 |
*max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
| 552 |
|
| 553 |
return bb; |
| 554 |
} |
| 555 |
|
| 556 |
#ifdef DOUBLE_CHECK |
| 557 |
static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, |
| 558 |
int first, int count) |
| 559 |
{ |
| 560 |
int i; |
| 561 |
struct super_block *sb = e4b->bd_sb; |
| 562 |
|
| 563 |
if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
| 564 |
return; |
| 565 |
assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
| 566 |
for (i = 0; i < count; i++) { |
| 567 |
if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { |
| 568 |
ext4_fsblk_t blocknr; |
| 569 |
|
| 570 |
blocknr = ext4_group_first_block_no(sb, e4b->bd_group); |
| 571 |
blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
| 572 |
ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
| 573 |
EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
| 574 |
ext4_grp_locked_error(sb, e4b->bd_group, |
| 575 |
inode ? inode->i_ino : 0, |
| 576 |
blocknr, |
| 577 |
"freeing block already freed " |
| 578 |
"(bit %u)", |
| 579 |
first + i); |
| 580 |
} |
| 581 |
mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); |
| 582 |
} |
| 583 |
} |
| 584 |
|
| 585 |
static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) |
| 586 |
{ |
| 587 |
int i; |
| 588 |
|
| 589 |
if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
| 590 |
return; |
| 591 |
assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
| 592 |
for (i = 0; i < count; i++) { |
| 593 |
BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); |
| 594 |
mb_set_bit(first + i, e4b->bd_info->bb_bitmap); |
| 595 |
} |
| 596 |
} |
| 597 |
|
| 598 |
static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) |
| 599 |
{ |
| 600 |
if (unlikely(e4b->bd_info->bb_bitmap == NULL)) |
| 601 |
return; |
| 602 |
if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { |
| 603 |
unsigned char *b1, *b2; |
| 604 |
int i; |
| 605 |
b1 = (unsigned char *) e4b->bd_info->bb_bitmap; |
| 606 |
b2 = (unsigned char *) bitmap; |
| 607 |
for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { |
| 608 |
if (b1[i] != b2[i]) { |
| 609 |
ext4_msg(e4b->bd_sb, KERN_ERR, |
| 610 |
"corruption in group %u " |
| 611 |
"at byte %u(%u): %x in copy != %x " |
| 612 |
"on disk/prealloc", |
| 613 |
e4b->bd_group, i, i * 8, b1[i], b2[i]); |
| 614 |
BUG(); |
| 615 |
} |
| 616 |
} |
| 617 |
} |
| 618 |
} |
| 619 |
|
| 620 |
static void mb_group_bb_bitmap_alloc(struct super_block *sb, |
| 621 |
struct ext4_group_info *grp, ext4_group_t group) |
| 622 |
{ |
| 623 |
struct buffer_head *bh; |
| 624 |
|
| 625 |
grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); |
| 626 |
if (!grp->bb_bitmap) |
| 627 |
return; |
| 628 |
|
| 629 |
bh = ext4_read_block_bitmap(sb, group); |
| 630 |
if (IS_ERR_OR_NULL(bh)) { |
| 631 |
kfree(grp->bb_bitmap); |
| 632 |
grp->bb_bitmap = NULL; |
| 633 |
return; |
| 634 |
} |
| 635 |
|
| 636 |
memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); |
| 637 |
put_bh(bh); |
| 638 |
} |
| 639 |
|
| 640 |
static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) |
| 641 |
{ |
| 642 |
kfree(grp->bb_bitmap); |
| 643 |
} |
| 644 |
|
| 645 |
#else |
| 646 |
static inline void mb_free_blocks_double(struct inode *inode, |
| 647 |
struct ext4_buddy *e4b, int first, int count) |
| 648 |
{ |
| 649 |
return; |
| 650 |
} |
| 651 |
static inline void mb_mark_used_double(struct ext4_buddy *e4b, |
| 652 |
int first, int count) |
| 653 |
{ |
| 654 |
return; |
| 655 |
} |
| 656 |
static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) |
| 657 |
{ |
| 658 |
return; |
| 659 |
} |
| 660 |
|
| 661 |
static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, |
| 662 |
struct ext4_group_info *grp, ext4_group_t group) |
| 663 |
{ |
| 664 |
return; |
| 665 |
} |
| 666 |
|
| 667 |
static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) |
| 668 |
{ |
| 669 |
return; |
| 670 |
} |
| 671 |
#endif |
| 672 |
|
| 673 |
#ifdef AGGRESSIVE_CHECK |
| 674 |
|
| 675 |
#define MB_CHECK_ASSERT(assert) \ |
| 676 |
do { \ |
| 677 |
if (!(assert)) { \ |
| 678 |
printk(KERN_EMERG \ |
| 679 |
"Assertion failure in %s() at %s:%d: \"%s\"\n", \ |
| 680 |
function, file, line, # assert); \ |
| 681 |
BUG(); \ |
| 682 |
} \ |
| 683 |
} while (0) |
| 684 |
|
| 685 |
/* |
| 686 |
* Perform buddy integrity check with the following steps: |
| 687 |
* |
| 688 |
* 1. Top-down validation (from highest order down to order 1, excluding order-0 bitmap): |
| 689 |
* For each pair of adjacent orders, if a higher-order bit is set (indicating a free block), |
| 690 |
* at most one of the two corresponding lower-order bits may be clear (free). |
| 691 |
* |
| 692 |
* 2. Order-0 (bitmap) validation, performed on bit pairs: |
| 693 |
* - If either bit in a pair is set (1, allocated), then all corresponding higher-order bits |
| 694 |
* must not be free (0). |
| 695 |
* - If both bits in a pair are clear (0, free), then exactly one of the corresponding |
| 696 |
* higher-order bits must be free (0). |
| 697 |
* |
| 698 |
* 3. Preallocation (pa) list validation: |
| 699 |
* For each preallocated block (pa) in the group: |
| 700 |
* - Verify that pa_pstart falls within the bounds of this block group. |
| 701 |
* - Ensure the corresponding bit(s) in the order-0 bitmap are marked as allocated (1). |
| 702 |
*/ |
| 703 |
static void __mb_check_buddy(struct ext4_buddy *e4b, char *file, |
| 704 |
const char *function, int line) |
| 705 |
{ |
| 706 |
struct super_block *sb = e4b->bd_sb; |
| 707 |
int order = e4b->bd_blkbits + 1; |
| 708 |
int max; |
| 709 |
int max2; |
| 710 |
int i; |
| 711 |
int j; |
| 712 |
int k; |
| 713 |
int count; |
| 714 |
struct ext4_group_info *grp; |
| 715 |
int fragments = 0; |
| 716 |
int fstart; |
| 717 |
struct list_head *cur; |
| 718 |
void *buddy; |
| 719 |
void *buddy2; |
| 720 |
|
| 721 |
if (e4b->bd_info->bb_check_counter++ % 10) |
| 722 |
return; |
| 723 |
|
| 724 |
while (order > 1) { |
| 725 |
buddy = mb_find_buddy(e4b, order, &max); |
| 726 |
MB_CHECK_ASSERT(buddy); |
| 727 |
buddy2 = mb_find_buddy(e4b, order - 1, &max2); |
| 728 |
MB_CHECK_ASSERT(buddy2); |
| 729 |
MB_CHECK_ASSERT(buddy != buddy2); |
| 730 |
MB_CHECK_ASSERT(max * 2 == max2); |
| 731 |
|
| 732 |
count = 0; |
| 733 |
for (i = 0; i < max; i++) { |
| 734 |
|
| 735 |
if (mb_test_bit(i, buddy)) { |
| 736 |
/* only single bit in buddy2 may be 0 */ |
| 737 |
if (!mb_test_bit(i << 1, buddy2)) { |
| 738 |
MB_CHECK_ASSERT( |
| 739 |
mb_test_bit((i<<1)+1, buddy2)); |
| 740 |
} |
| 741 |
continue; |
| 742 |
} |
| 743 |
|
| 744 |
count++; |
| 745 |
} |
| 746 |
MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); |
| 747 |
order--; |
| 748 |
} |
| 749 |
|
| 750 |
fstart = -1; |
| 751 |
buddy = mb_find_buddy(e4b, 0, &max); |
| 752 |
for (i = 0; i < max; i++) { |
| 753 |
if (!mb_test_bit(i, buddy)) { |
| 754 |
MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); |
| 755 |
if (fstart == -1) { |
| 756 |
fragments++; |
| 757 |
fstart = i; |
| 758 |
} |
| 759 |
} else { |
| 760 |
fstart = -1; |
| 761 |
} |
| 762 |
if (!(i & 1)) { |
| 763 |
int in_use, zero_bit_count = 0; |
| 764 |
|
| 765 |
in_use = mb_test_bit(i, buddy) || mb_test_bit(i + 1, buddy); |
| 766 |
for (j = 1; j < e4b->bd_blkbits + 2; j++) { |
| 767 |
buddy2 = mb_find_buddy(e4b, j, &max2); |
| 768 |
k = i >> j; |
| 769 |
MB_CHECK_ASSERT(k < max2); |
| 770 |
if (!mb_test_bit(k, buddy2)) |
| 771 |
zero_bit_count++; |
| 772 |
} |
| 773 |
MB_CHECK_ASSERT(zero_bit_count == !in_use); |
| 774 |
} |
| 775 |
} |
| 776 |
MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); |
| 777 |
MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); |
| 778 |
|
| 779 |
grp = ext4_get_group_info(sb, e4b->bd_group); |
| 780 |
if (!grp) |
| 781 |
return; |
| 782 |
list_for_each(cur, &grp->bb_prealloc_list) { |
| 783 |
ext4_group_t groupnr; |
| 784 |
struct ext4_prealloc_space *pa; |
| 785 |
pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
| 786 |
if (!pa->pa_len) |
| 787 |
continue; |
| 788 |
ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); |
| 789 |
MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
| 790 |
for (i = 0; i < pa->pa_len; i++) |
| 791 |
MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
| 792 |
} |
| 793 |
} |
| 794 |
#undef MB_CHECK_ASSERT |
| 795 |
#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ |
| 796 |
__FILE__, __func__, __LINE__) |
| 797 |
#else |
| 798 |
#define mb_check_buddy(e4b) |
| 799 |
#endif |
| 800 |
|
| 801 |
/* |
| 802 |
* Divide blocks started from @first with length @len into |
| 803 |
* smaller chunks with power of 2 blocks. |
| 804 |
* Clear the bits in bitmap which the blocks of the chunk(s) covered, |
| 805 |
* then increase bb_counters[] for corresponded chunk size. |
| 806 |
*/ |
| 807 |
static void ext4_mb_mark_free_simple(struct super_block *sb, |
| 808 |
void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
| 809 |
struct ext4_group_info *grp) |
| 810 |
{ |
| 811 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 812 |
ext4_grpblk_t min; |
| 813 |
ext4_grpblk_t max; |
| 814 |
ext4_grpblk_t chunk; |
| 815 |
unsigned int border; |
| 816 |
|
| 817 |
BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
| 818 |
|
| 819 |
border = 2 << sb->s_blocksize_bits; |
| 820 |
|
| 821 |
while (len > 0) { |
| 822 |
/* find how many blocks can be covered since this position */ |
| 823 |
max = ffs(first | border) - 1; |
| 824 |
|
| 825 |
/* find how many blocks of power 2 we need to mark */ |
| 826 |
min = fls(len) - 1; |
| 827 |
|
| 828 |
if (max < min) |
| 829 |
min = max; |
| 830 |
chunk = 1 << min; |
| 831 |
|
| 832 |
/* mark multiblock chunks only */ |
| 833 |
grp->bb_counters[min]++; |
| 834 |
if (min > 0) |
| 835 |
mb_clear_bit(first >> min, |
| 836 |
buddy + sbi->s_mb_offsets[min]); |
| 837 |
|
| 838 |
len -= chunk; |
| 839 |
first += chunk; |
| 840 |
} |
| 841 |
} |
| 842 |
|
| 843 |
static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) |
| 844 |
{ |
| 845 |
int order; |
| 846 |
|
| 847 |
/* |
| 848 |
* We don't bother with a special lists groups with only 1 block free |
| 849 |
* extents and for completely empty groups. |
| 850 |
*/ |
| 851 |
order = fls(len) - 2; |
| 852 |
if (order < 0) |
| 853 |
return 0; |
| 854 |
if (order == MB_NUM_ORDERS(sb)) |
| 855 |
order--; |
| 856 |
if (WARN_ON_ONCE(order > MB_NUM_ORDERS(sb))) |
| 857 |
order = MB_NUM_ORDERS(sb) - 1; |
| 858 |
return order; |
| 859 |
} |
| 860 |
|
| 861 |
/* Move group to appropriate avg_fragment_size list */ |
| 862 |
static void |
| 863 |
mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) |
| 864 |
{ |
| 865 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 866 |
int new, old; |
| 867 |
|
| 868 |
if (!test_opt2(sb, MB_OPTIMIZE_SCAN)) |
| 869 |
return; |
| 870 |
|
| 871 |
old = grp->bb_avg_fragment_size_order; |
| 872 |
new = grp->bb_fragments == 0 ? -1 : |
| 873 |
mb_avg_fragment_size_order(sb, grp->bb_free / grp->bb_fragments); |
| 874 |
if (new == old) |
| 875 |
return; |
| 876 |
|
| 877 |
if (old >= 0) |
| 878 |
xa_erase(&sbi->s_mb_avg_fragment_size[old], grp->bb_group); |
| 879 |
|
| 880 |
grp->bb_avg_fragment_size_order = new; |
| 881 |
if (new >= 0) { |
| 882 |
/* |
| 883 |
* Cannot use __GFP_NOFAIL because we hold the group lock. |
| 884 |
* Although allocation for insertion may fails, it's not fatal |
| 885 |
* as we have linear traversal to fall back on. |
| 886 |
*/ |
| 887 |
int err = xa_insert(&sbi->s_mb_avg_fragment_size[new], |
| 888 |
grp->bb_group, grp, GFP_ATOMIC); |
| 889 |
if (err) |
| 890 |
mb_debug(sb, "insert group: %u to s_mb_avg_fragment_size[%d] failed, err %d", |
| 891 |
grp->bb_group, new, err); |
| 892 |
} |
| 893 |
} |
| 894 |
|
| 895 |
static ext4_group_t ext4_get_allocation_groups_count( |
| 896 |
struct ext4_allocation_context *ac) |
| 897 |
{ |
| 898 |
ext4_group_t ngroups = ext4_get_groups_count(ac->ac_sb); |
| 899 |
|
| 900 |
/* non-extent files are limited to low blocks/groups */ |
| 901 |
if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
| 902 |
ngroups = EXT4_SB(ac->ac_sb)->s_blockfile_groups; |
| 903 |
|
| 904 |
/* Pairs with smp_wmb() in ext4_update_super() */ |
| 905 |
smp_rmb(); |
| 906 |
|
| 907 |
return ngroups; |
| 908 |
} |
| 909 |
|
| 910 |
static int ext4_mb_scan_groups_xa_range(struct ext4_allocation_context *ac, |
| 911 |
struct xarray *xa, |
| 912 |
ext4_group_t start, ext4_group_t end) |
| 913 |
{ |
| 914 |
struct super_block *sb = ac->ac_sb; |
| 915 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 916 |
enum criteria cr = ac->ac_criteria; |
| 917 |
ext4_group_t ngroups = ext4_get_allocation_groups_count(ac); |
| 918 |
unsigned long group = start; |
| 919 |
struct ext4_group_info *grp; |
| 920 |
|
| 921 |
if (WARN_ON_ONCE(end > ngroups || start >= end)) |
| 922 |
return 0; |
| 923 |
|
| 924 |
xa_for_each_range(xa, group, grp, start, end - 1) { |
| 925 |
int err; |
| 926 |
|
| 927 |
if (sbi->s_mb_stats) |
| 928 |
atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]); |
| 929 |
|
| 930 |
err = ext4_mb_scan_group(ac, grp->bb_group); |
| 931 |
if (err || ac->ac_status != AC_STATUS_CONTINUE) |
| 932 |
return err; |
| 933 |
|
| 934 |
cond_resched(); |
| 935 |
} |
| 936 |
|
| 937 |
return 0; |
| 938 |
} |
| 939 |
|
| 940 |
/* |
| 941 |
* Find a suitable group of given order from the largest free orders xarray. |
| 942 |
*/ |
| 943 |
static inline int |
| 944 |
ext4_mb_scan_groups_largest_free_order_range(struct ext4_allocation_context *ac, |
| 945 |
int order, ext4_group_t start, |
| 946 |
ext4_group_t end) |
| 947 |
{ |
| 948 |
struct xarray *xa = &EXT4_SB(ac->ac_sb)->s_mb_largest_free_orders[order]; |
| 949 |
|
| 950 |
if (xa_empty(xa)) |
| 951 |
return 0; |
| 952 |
|
| 953 |
return ext4_mb_scan_groups_xa_range(ac, xa, start, end); |
| 954 |
} |
| 955 |
|
| 956 |
/* |
| 957 |
* Choose next group by traversing largest_free_order lists. Updates *new_cr if |
| 958 |
* cr level needs an update. |
| 959 |
*/ |
| 960 |
static int ext4_mb_scan_groups_p2_aligned(struct ext4_allocation_context *ac, |
| 961 |
ext4_group_t group) |
| 962 |
{ |
| 963 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 964 |
int i; |
| 965 |
int ret = 0; |
| 966 |
ext4_group_t start, end; |
| 967 |
|
| 968 |
start = group; |
| 969 |
end = ext4_get_allocation_groups_count(ac); |
| 970 |
wrap_around: |
| 971 |
for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { |
| 972 |
ret = ext4_mb_scan_groups_largest_free_order_range(ac, i, |
| 973 |
start, end); |
| 974 |
if (ret || ac->ac_status != AC_STATUS_CONTINUE) |
| 975 |
return ret; |
| 976 |
} |
| 977 |
if (start) { |
| 978 |
end = start; |
| 979 |
start = 0; |
| 980 |
goto wrap_around; |
| 981 |
} |
| 982 |
|
| 983 |
if (sbi->s_mb_stats) |
| 984 |
atomic64_inc(&sbi->s_bal_cX_failed[ac->ac_criteria]); |
| 985 |
|
| 986 |
/* Increment cr and search again if no group is found */ |
| 987 |
ac->ac_criteria = CR_GOAL_LEN_FAST; |
| 988 |
return ret; |
| 989 |
} |
| 990 |
|
| 991 |
/* |
| 992 |
* Find a suitable group of given order from the average fragments xarray. |
| 993 |
*/ |
| 994 |
static int |
| 995 |
ext4_mb_scan_groups_avg_frag_order_range(struct ext4_allocation_context *ac, |
| 996 |
int order, ext4_group_t start, |
| 997 |
ext4_group_t end) |
| 998 |
{ |
| 999 |
struct xarray *xa = &EXT4_SB(ac->ac_sb)->s_mb_avg_fragment_size[order]; |
| 1000 |
|
| 1001 |
if (xa_empty(xa)) |
| 1002 |
return 0; |
| 1003 |
|
| 1004 |
return ext4_mb_scan_groups_xa_range(ac, xa, start, end); |
| 1005 |
} |
| 1006 |
|
| 1007 |
/* |
| 1008 |
* Choose next group by traversing average fragment size list of suitable |
| 1009 |
* order. Updates *new_cr if cr level needs an update. |
| 1010 |
*/ |
| 1011 |
static int ext4_mb_scan_groups_goal_fast(struct ext4_allocation_context *ac, |
| 1012 |
ext4_group_t group) |
| 1013 |
{ |
| 1014 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 1015 |
int i, ret = 0; |
| 1016 |
ext4_group_t start, end; |
| 1017 |
|
| 1018 |
start = group; |
| 1019 |
end = ext4_get_allocation_groups_count(ac); |
| 1020 |
wrap_around: |
| 1021 |
i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); |
| 1022 |
for (; i < MB_NUM_ORDERS(ac->ac_sb); i++) { |
| 1023 |
ret = ext4_mb_scan_groups_avg_frag_order_range(ac, i, |
| 1024 |
start, end); |
| 1025 |
if (ret || ac->ac_status != AC_STATUS_CONTINUE) |
| 1026 |
return ret; |
| 1027 |
} |
| 1028 |
if (start) { |
| 1029 |
end = start; |
| 1030 |
start = 0; |
| 1031 |
goto wrap_around; |
| 1032 |
} |
| 1033 |
|
| 1034 |
if (sbi->s_mb_stats) |
| 1035 |
atomic64_inc(&sbi->s_bal_cX_failed[ac->ac_criteria]); |
| 1036 |
/* |
| 1037 |
* CR_BEST_AVAIL_LEN works based on the concept that we have |
| 1038 |
* a larger normalized goal len request which can be trimmed to |
| 1039 |
* a smaller goal len such that it can still satisfy original |
| 1040 |
* request len. However, allocation request for non-regular |
| 1041 |
* files never gets normalized. |
| 1042 |
* See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA). |
| 1043 |
*/ |
| 1044 |
if (ac->ac_flags & EXT4_MB_HINT_DATA) |
| 1045 |
ac->ac_criteria = CR_BEST_AVAIL_LEN; |
| 1046 |
else |
| 1047 |
ac->ac_criteria = CR_GOAL_LEN_SLOW; |
| 1048 |
|
| 1049 |
return ret; |
| 1050 |
} |
| 1051 |
|
| 1052 |
/* |
| 1053 |
* We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment |
| 1054 |
* order we have and proactively trim the goal request length to that order to |
| 1055 |
* find a suitable group faster. |
| 1056 |
* |
| 1057 |
* This optimizes allocation speed at the cost of slightly reduced |
| 1058 |
* preallocations. However, we make sure that we don't trim the request too |
| 1059 |
* much and fall to CR_GOAL_LEN_SLOW in that case. |
| 1060 |
*/ |
| 1061 |
static int ext4_mb_scan_groups_best_avail(struct ext4_allocation_context *ac, |
| 1062 |
ext4_group_t group) |
| 1063 |
{ |
| 1064 |
int ret = 0; |
| 1065 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 1066 |
int i, order, min_order; |
| 1067 |
unsigned long num_stripe_clusters = 0; |
| 1068 |
ext4_group_t start, end; |
| 1069 |
|
| 1070 |
/* |
| 1071 |
* mb_avg_fragment_size_order() returns order in a way that makes |
| 1072 |
* retrieving back the length using (1 << order) inaccurate. Hence, use |
| 1073 |
* fls() instead since we need to know the actual length while modifying |
| 1074 |
* goal length. |
| 1075 |
*/ |
| 1076 |
order = fls(ac->ac_g_ex.fe_len) - 1; |
| 1077 |
if (WARN_ON_ONCE(order - 1 > MB_NUM_ORDERS(ac->ac_sb))) |
| 1078 |
order = MB_NUM_ORDERS(ac->ac_sb); |
| 1079 |
min_order = order - sbi->s_mb_best_avail_max_trim_order; |
| 1080 |
if (min_order < 0) |
| 1081 |
min_order = 0; |
| 1082 |
|
| 1083 |
if (sbi->s_stripe > 0) { |
| 1084 |
/* |
| 1085 |
* We are assuming that stripe size is always a multiple of |
| 1086 |
* cluster ratio otherwise __ext4_fill_super exists early. |
| 1087 |
*/ |
| 1088 |
num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe); |
| 1089 |
if (1 << min_order < num_stripe_clusters) |
| 1090 |
/* |
| 1091 |
* We consider 1 order less because later we round |
| 1092 |
* up the goal len to num_stripe_clusters |
| 1093 |
*/ |
| 1094 |
min_order = fls(num_stripe_clusters) - 1; |
| 1095 |
} |
| 1096 |
|
| 1097 |
if (1 << min_order < ac->ac_o_ex.fe_len) |
| 1098 |
min_order = fls(ac->ac_o_ex.fe_len); |
| 1099 |
|
| 1100 |
start = group; |
| 1101 |
end = ext4_get_allocation_groups_count(ac); |
| 1102 |
wrap_around: |
| 1103 |
for (i = order; i >= min_order; i--) { |
| 1104 |
int frag_order; |
| 1105 |
/* |
| 1106 |
* Scale down goal len to make sure we find something |
| 1107 |
* in the free fragments list. Basically, reduce |
| 1108 |
* preallocations. |
| 1109 |
*/ |
| 1110 |
ac->ac_g_ex.fe_len = 1 << i; |
| 1111 |
|
| 1112 |
if (num_stripe_clusters > 0) { |
| 1113 |
/* |
| 1114 |
* Try to round up the adjusted goal length to |
| 1115 |
* stripe size (in cluster units) multiple for |
| 1116 |
* efficiency. |
| 1117 |
*/ |
| 1118 |
ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len, |
| 1119 |
num_stripe_clusters); |
| 1120 |
} |
| 1121 |
|
| 1122 |
frag_order = mb_avg_fragment_size_order(ac->ac_sb, |
| 1123 |
ac->ac_g_ex.fe_len); |
| 1124 |
|
| 1125 |
ret = ext4_mb_scan_groups_avg_frag_order_range(ac, frag_order, |
| 1126 |
start, end); |
| 1127 |
if (ret || ac->ac_status != AC_STATUS_CONTINUE) |
| 1128 |
return ret; |
| 1129 |
} |
| 1130 |
if (start) { |
| 1131 |
end = start; |
| 1132 |
start = 0; |
| 1133 |
goto wrap_around; |
| 1134 |
} |
| 1135 |
|
| 1136 |
/* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */ |
| 1137 |
ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; |
| 1138 |
if (sbi->s_mb_stats) |
| 1139 |
atomic64_inc(&sbi->s_bal_cX_failed[ac->ac_criteria]); |
| 1140 |
ac->ac_criteria = CR_GOAL_LEN_SLOW; |
| 1141 |
|
| 1142 |
return ret; |
| 1143 |
} |
| 1144 |
|
| 1145 |
static inline int should_optimize_scan(struct ext4_allocation_context *ac) |
| 1146 |
{ |
| 1147 |
if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) |
| 1148 |
return 0; |
| 1149 |
if (ac->ac_criteria >= CR_GOAL_LEN_SLOW) |
| 1150 |
return 0; |
| 1151 |
return 1; |
| 1152 |
} |
| 1153 |
|
| 1154 |
/* |
| 1155 |
* next linear group for allocation. |
| 1156 |
*/ |
| 1157 |
static void next_linear_group(ext4_group_t *group, ext4_group_t ngroups) |
| 1158 |
{ |
| 1159 |
/* |
| 1160 |
* Artificially restricted ngroups for non-extent |
| 1161 |
* files makes group > ngroups possible on first loop. |
| 1162 |
*/ |
| 1163 |
*group = *group + 1 >= ngroups ? 0 : *group + 1; |
| 1164 |
} |
| 1165 |
|
| 1166 |
static int ext4_mb_scan_groups_linear(struct ext4_allocation_context *ac, |
| 1167 |
ext4_group_t ngroups, ext4_group_t *start, ext4_group_t count) |
| 1168 |
{ |
| 1169 |
int ret, i; |
| 1170 |
enum criteria cr = ac->ac_criteria; |
| 1171 |
struct super_block *sb = ac->ac_sb; |
| 1172 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 1173 |
ext4_group_t group = *start; |
| 1174 |
|
| 1175 |
for (i = 0; i < count; i++, next_linear_group(&group, ngroups)) { |
| 1176 |
ret = ext4_mb_scan_group(ac, group); |
| 1177 |
if (ret || ac->ac_status != AC_STATUS_CONTINUE) |
| 1178 |
return ret; |
| 1179 |
cond_resched(); |
| 1180 |
} |
| 1181 |
|
| 1182 |
*start = group; |
| 1183 |
if (count == ngroups) |
| 1184 |
ac->ac_criteria++; |
| 1185 |
|
| 1186 |
/* Processed all groups and haven't found blocks */ |
| 1187 |
if (sbi->s_mb_stats && i == ngroups) |
| 1188 |
atomic64_inc(&sbi->s_bal_cX_failed[cr]); |
| 1189 |
|
| 1190 |
return 0; |
| 1191 |
} |
| 1192 |
|
| 1193 |
static int ext4_mb_scan_groups(struct ext4_allocation_context *ac) |
| 1194 |
{ |
| 1195 |
int ret = 0; |
| 1196 |
ext4_group_t start; |
| 1197 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 1198 |
ext4_group_t ngroups = ext4_get_allocation_groups_count(ac); |
| 1199 |
|
| 1200 |
/* searching for the right group start from the goal value specified */ |
| 1201 |
start = ac->ac_g_ex.fe_group; |
| 1202 |
if (start >= ngroups) |
| 1203 |
start = 0; |
| 1204 |
ac->ac_prefetch_grp = start; |
| 1205 |
ac->ac_prefetch_nr = 0; |
| 1206 |
|
| 1207 |
if (!should_optimize_scan(ac)) |
| 1208 |
return ext4_mb_scan_groups_linear(ac, ngroups, &start, ngroups); |
| 1209 |
|
| 1210 |
/* |
| 1211 |
* Optimized scanning can return non adjacent groups which can cause |
| 1212 |
* seek overhead for rotational disks. So try few linear groups before |
| 1213 |
* trying optimized scan. |
| 1214 |
*/ |
| 1215 |
if (sbi->s_mb_max_linear_groups) |
| 1216 |
ret = ext4_mb_scan_groups_linear(ac, ngroups, &start, |
| 1217 |
sbi->s_mb_max_linear_groups); |
| 1218 |
if (ret || ac->ac_status != AC_STATUS_CONTINUE) |
| 1219 |
return ret; |
| 1220 |
|
| 1221 |
switch (ac->ac_criteria) { |
| 1222 |
case CR_POWER2_ALIGNED: |
| 1223 |
return ext4_mb_scan_groups_p2_aligned(ac, start); |
| 1224 |
case CR_GOAL_LEN_FAST: |
| 1225 |
return ext4_mb_scan_groups_goal_fast(ac, start); |
| 1226 |
case CR_BEST_AVAIL_LEN: |
| 1227 |
return ext4_mb_scan_groups_best_avail(ac, start); |
| 1228 |
default: |
| 1229 |
/* |
| 1230 |
* TODO: For CR_GOAL_LEN_SLOW, we can arrange groups in an |
| 1231 |
* rb tree sorted by bb_free. But until that happens, we should |
| 1232 |
* never come here. |
| 1233 |
*/ |
| 1234 |
WARN_ON(1); |
| 1235 |
} |
| 1236 |
|
| 1237 |
return 0; |
| 1238 |
} |
| 1239 |
|
| 1240 |
/* |
| 1241 |
* Cache the order of the largest free extent we have available in this block |
| 1242 |
* group. |
| 1243 |
*/ |
| 1244 |
static void |
| 1245 |
mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) |
| 1246 |
{ |
| 1247 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 1248 |
int new, old = grp->bb_largest_free_order; |
| 1249 |
|
| 1250 |
for (new = MB_NUM_ORDERS(sb) - 1; new >= 0; new--) |
| 1251 |
if (grp->bb_counters[new] > 0) |
| 1252 |
break; |
| 1253 |
|
| 1254 |
/* No need to move between order lists? */ |
| 1255 |
if (new == old) |
| 1256 |
return; |
| 1257 |
|
| 1258 |
if (old >= 0) { |
| 1259 |
struct xarray *xa = &sbi->s_mb_largest_free_orders[old]; |
| 1260 |
|
| 1261 |
if (!xa_empty(xa) && xa_load(xa, grp->bb_group)) |
| 1262 |
xa_erase(xa, grp->bb_group); |
| 1263 |
} |
| 1264 |
|
| 1265 |
grp->bb_largest_free_order = new; |
| 1266 |
if (test_opt2(sb, MB_OPTIMIZE_SCAN) && new >= 0 && grp->bb_free) { |
| 1267 |
/* |
| 1268 |
* Cannot use __GFP_NOFAIL because we hold the group lock. |
| 1269 |
* Although allocation for insertion may fails, it's not fatal |
| 1270 |
* as we have linear traversal to fall back on. |
| 1271 |
*/ |
| 1272 |
int err = xa_insert(&sbi->s_mb_largest_free_orders[new], |
| 1273 |
grp->bb_group, grp, GFP_ATOMIC); |
| 1274 |
if (err) |
| 1275 |
mb_debug(sb, "insert group: %u to s_mb_largest_free_orders[%d] failed, err %d", |
| 1276 |
grp->bb_group, new, err); |
| 1277 |
} |
| 1278 |
} |
| 1279 |
|
| 1280 |
static noinline_for_stack |
| 1281 |
void ext4_mb_generate_buddy(struct super_block *sb, |
| 1282 |
void *buddy, void *bitmap, ext4_group_t group, |
| 1283 |
struct ext4_group_info *grp) |
| 1284 |
{ |
| 1285 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 1286 |
ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
| 1287 |
ext4_grpblk_t i = 0; |
| 1288 |
ext4_grpblk_t first; |
| 1289 |
ext4_grpblk_t len; |
| 1290 |
unsigned free = 0; |
| 1291 |
unsigned fragments = 0; |
| 1292 |
unsigned long long period = get_cycles(); |
| 1293 |
|
| 1294 |
/* initialize buddy from bitmap which is aggregation |
| 1295 |
* of on-disk bitmap and preallocations */ |
| 1296 |
i = mb_find_next_zero_bit(bitmap, max, 0); |
| 1297 |
grp->bb_first_free = i; |
| 1298 |
while (i < max) { |
| 1299 |
fragments++; |
| 1300 |
first = i; |
| 1301 |
i = mb_find_next_bit(bitmap, max, i); |
| 1302 |
len = i - first; |
| 1303 |
free += len; |
| 1304 |
if (len > 1) |
| 1305 |
ext4_mb_mark_free_simple(sb, buddy, first, len, grp); |
| 1306 |
else |
| 1307 |
grp->bb_counters[0]++; |
| 1308 |
if (i < max) |
| 1309 |
i = mb_find_next_zero_bit(bitmap, max, i); |
| 1310 |
} |
| 1311 |
grp->bb_fragments = fragments; |
| 1312 |
|
| 1313 |
if (free != grp->bb_free) { |
| 1314 |
ext4_grp_locked_error(sb, group, 0, 0, |
| 1315 |
"block bitmap and bg descriptor " |
| 1316 |
"inconsistent: %u vs %u free clusters", |
| 1317 |
free, grp->bb_free); |
| 1318 |
/* |
| 1319 |
* If we intend to continue, we consider group descriptor |
| 1320 |
* corrupt and update bb_free using bitmap value |
| 1321 |
*/ |
| 1322 |
grp->bb_free = free; |
| 1323 |
ext4_mark_group_bitmap_corrupted(sb, group, |
| 1324 |
EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
| 1325 |
} |
| 1326 |
mb_set_largest_free_order(sb, grp); |
| 1327 |
mb_update_avg_fragment_size(sb, grp); |
| 1328 |
|
| 1329 |
clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); |
| 1330 |
|
| 1331 |
period = get_cycles() - period; |
| 1332 |
atomic_inc(&sbi->s_mb_buddies_generated); |
| 1333 |
atomic64_add(period, &sbi->s_mb_generation_time); |
| 1334 |
} |
| 1335 |
|
| 1336 |
static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
| 1337 |
{ |
| 1338 |
int count; |
| 1339 |
int order = 1; |
| 1340 |
void *buddy; |
| 1341 |
|
| 1342 |
while ((buddy = mb_find_buddy(e4b, order++, &count))) |
| 1343 |
mb_set_bits(buddy, 0, count); |
| 1344 |
|
| 1345 |
e4b->bd_info->bb_fragments = 0; |
| 1346 |
memset(e4b->bd_info->bb_counters, 0, |
| 1347 |
sizeof(*e4b->bd_info->bb_counters) * |
| 1348 |
(e4b->bd_sb->s_blocksize_bits + 2)); |
| 1349 |
|
| 1350 |
ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, |
| 1351 |
e4b->bd_bitmap, e4b->bd_group, e4b->bd_info); |
| 1352 |
} |
| 1353 |
|
| 1354 |
/* The buddy information is attached the buddy cache inode |
| 1355 |
* for convenience. The information regarding each group |
| 1356 |
* is loaded via ext4_mb_load_buddy. The information involve |
| 1357 |
* block bitmap and buddy information. The information are |
| 1358 |
* stored in the inode as |
| 1359 |
* |
| 1360 |
* { folio } |
| 1361 |
* [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
| 1362 |
* |
| 1363 |
* |
| 1364 |
* one block each for bitmap and buddy information. |
| 1365 |
* So for each group we take up 2 blocks. A folio can |
| 1366 |
* contain blocks_per_folio (folio_size / blocksize) blocks. |
| 1367 |
* So it can have information regarding groups_per_folio which |
| 1368 |
* is blocks_per_folio/2 |
| 1369 |
* |
| 1370 |
* Locking note: This routine takes the block group lock of all groups |
| 1371 |
* for this folio; do not hold this lock when calling this routine! |
| 1372 |
*/ |
| 1373 |
static int ext4_mb_init_cache(struct folio *folio, char *incore, gfp_t gfp) |
| 1374 |
{ |
| 1375 |
ext4_group_t ngroups; |
| 1376 |
unsigned int blocksize; |
| 1377 |
int blocks_per_folio; |
| 1378 |
int groups_per_folio; |
| 1379 |
int err = 0; |
| 1380 |
int i; |
| 1381 |
ext4_group_t first_group, group; |
| 1382 |
int first_block; |
| 1383 |
struct super_block *sb; |
| 1384 |
struct buffer_head *bhs; |
| 1385 |
struct buffer_head **bh = NULL; |
| 1386 |
struct inode *inode; |
| 1387 |
char *data; |
| 1388 |
char *bitmap; |
| 1389 |
struct ext4_group_info *grinfo; |
| 1390 |
|
| 1391 |
inode = folio->mapping->host; |
| 1392 |
sb = inode->i_sb; |
| 1393 |
ngroups = ext4_get_groups_count(sb); |
| 1394 |
blocksize = i_blocksize(inode); |
| 1395 |
blocks_per_folio = folio_size(folio) / blocksize; |
| 1396 |
WARN_ON_ONCE(!blocks_per_folio); |
| 1397 |
groups_per_folio = DIV_ROUND_UP(blocks_per_folio, 2); |
| 1398 |
|
| 1399 |
mb_debug(sb, "init folio %lu\n", folio->index); |
| 1400 |
|
| 1401 |
/* allocate buffer_heads to read bitmaps */ |
| 1402 |
if (groups_per_folio > 1) { |
| 1403 |
i = sizeof(struct buffer_head *) * groups_per_folio; |
| 1404 |
bh = kzalloc(i, gfp); |
| 1405 |
if (bh == NULL) |
| 1406 |
return -ENOMEM; |
| 1407 |
} else |
| 1408 |
bh = &bhs; |
| 1409 |
|
| 1410 |
/* read all groups the folio covers into the cache */ |
| 1411 |
first_group = EXT4_PG_TO_LBLK(inode, folio->index) / 2; |
| 1412 |
for (i = 0, group = first_group; i < groups_per_folio; i++, group++) { |
| 1413 |
if (group >= ngroups) |
| 1414 |
break; |
| 1415 |
|
| 1416 |
grinfo = ext4_get_group_info(sb, group); |
| 1417 |
if (!grinfo) |
| 1418 |
continue; |
| 1419 |
/* |
| 1420 |
* If folio is uptodate then we came here after online resize |
| 1421 |
* which added some new uninitialized group info structs, so |
| 1422 |
* we must skip all initialized uptodate buddies on the folio, |
| 1423 |
* which may be currently in use by an allocating task. |
| 1424 |
*/ |
| 1425 |
if (folio_test_uptodate(folio) && |
| 1426 |
!EXT4_MB_GRP_NEED_INIT(grinfo)) { |
| 1427 |
bh[i] = NULL; |
| 1428 |
continue; |
| 1429 |
} |
| 1430 |
bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); |
| 1431 |
if (IS_ERR(bh[i])) { |
| 1432 |
err = PTR_ERR(bh[i]); |
| 1433 |
bh[i] = NULL; |
| 1434 |
goto out; |
| 1435 |
} |
| 1436 |
mb_debug(sb, "read bitmap for group %u\n", group); |
| 1437 |
} |
| 1438 |
|
| 1439 |
/* wait for I/O completion */ |
| 1440 |
for (i = 0, group = first_group; i < groups_per_folio; i++, group++) { |
| 1441 |
int err2; |
| 1442 |
|
| 1443 |
if (!bh[i]) |
| 1444 |
continue; |
| 1445 |
err2 = ext4_wait_block_bitmap(sb, group, bh[i]); |
| 1446 |
if (!err) |
| 1447 |
err = err2; |
| 1448 |
} |
| 1449 |
|
| 1450 |
first_block = EXT4_PG_TO_LBLK(inode, folio->index); |
| 1451 |
for (i = 0; i < blocks_per_folio; i++) { |
| 1452 |
group = (first_block + i) >> 1; |
| 1453 |
if (group >= ngroups) |
| 1454 |
break; |
| 1455 |
|
| 1456 |
if (!bh[group - first_group]) |
| 1457 |
/* skip initialized uptodate buddy */ |
| 1458 |
continue; |
| 1459 |
|
| 1460 |
if (!buffer_verified(bh[group - first_group])) |
| 1461 |
/* Skip faulty bitmaps */ |
| 1462 |
continue; |
| 1463 |
err = 0; |
| 1464 |
|
| 1465 |
/* |
| 1466 |
* data carry information regarding this |
| 1467 |
* particular group in the format specified |
| 1468 |
* above |
| 1469 |
* |
| 1470 |
*/ |
| 1471 |
data = folio_address(folio) + (i * blocksize); |
| 1472 |
bitmap = bh[group - first_group]->b_data; |
| 1473 |
|
| 1474 |
/* |
| 1475 |
* We place the buddy block and bitmap block |
| 1476 |
* close together |
| 1477 |
*/ |
| 1478 |
grinfo = ext4_get_group_info(sb, group); |
| 1479 |
if (!grinfo) { |
| 1480 |
err = -EFSCORRUPTED; |
| 1481 |
goto out; |
| 1482 |
} |
| 1483 |
if ((first_block + i) & 1) { |
| 1484 |
/* this is block of buddy */ |
| 1485 |
BUG_ON(incore == NULL); |
| 1486 |
mb_debug(sb, "put buddy for group %u in folio %lu/%x\n", |
| 1487 |
group, folio->index, i * blocksize); |
| 1488 |
trace_ext4_mb_buddy_bitmap_load(sb, group); |
| 1489 |
grinfo->bb_fragments = 0; |
| 1490 |
memset(grinfo->bb_counters, 0, |
| 1491 |
sizeof(*grinfo->bb_counters) * |
| 1492 |
(MB_NUM_ORDERS(sb))); |
| 1493 |
/* |
| 1494 |
* incore got set to the group block bitmap below |
| 1495 |
*/ |
| 1496 |
ext4_lock_group(sb, group); |
| 1497 |
/* init the buddy */ |
| 1498 |
memset(data, 0xff, blocksize); |
| 1499 |
ext4_mb_generate_buddy(sb, data, incore, group, grinfo); |
| 1500 |
ext4_unlock_group(sb, group); |
| 1501 |
incore = NULL; |
| 1502 |
} else { |
| 1503 |
/* this is block of bitmap */ |
| 1504 |
BUG_ON(incore != NULL); |
| 1505 |
mb_debug(sb, "put bitmap for group %u in folio %lu/%x\n", |
| 1506 |
group, folio->index, i * blocksize); |
| 1507 |
trace_ext4_mb_bitmap_load(sb, group); |
| 1508 |
|
| 1509 |
/* see comments in ext4_mb_put_pa() */ |
| 1510 |
ext4_lock_group(sb, group); |
| 1511 |
memcpy(data, bitmap, blocksize); |
| 1512 |
|
| 1513 |
/* mark all preallocated blks used in in-core bitmap */ |
| 1514 |
ext4_mb_generate_from_pa(sb, data, group); |
| 1515 |
WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root)); |
| 1516 |
ext4_unlock_group(sb, group); |
| 1517 |
|
| 1518 |
/* set incore so that the buddy information can be |
| 1519 |
* generated using this |
| 1520 |
*/ |
| 1521 |
incore = data; |
| 1522 |
} |
| 1523 |
} |
| 1524 |
folio_mark_uptodate(folio); |
| 1525 |
|
| 1526 |
out: |
| 1527 |
if (bh) { |
| 1528 |
for (i = 0; i < groups_per_folio; i++) |
| 1529 |
brelse(bh[i]); |
| 1530 |
if (bh != &bhs) |
| 1531 |
kfree(bh); |
| 1532 |
} |
| 1533 |
return err; |
| 1534 |
} |
| 1535 |
|
| 1536 |
/* |
| 1537 |
* Lock the buddy and bitmap folios. This makes sure other parallel init_group |
| 1538 |
* on the same buddy folio doesn't happen while holding the buddy folio lock. |
| 1539 |
* Return locked buddy and bitmap folios on e4b struct. If buddy and bitmap |
| 1540 |
* are on the same folio e4b->bd_buddy_folio is NULL and return value is 0. |
| 1541 |
*/ |
| 1542 |
static int ext4_mb_get_buddy_folio_lock(struct super_block *sb, |
| 1543 |
ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) |
| 1544 |
{ |
| 1545 |
struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
| 1546 |
int block, pnum; |
| 1547 |
struct folio *folio; |
| 1548 |
|
| 1549 |
e4b->bd_buddy_folio = NULL; |
| 1550 |
e4b->bd_bitmap_folio = NULL; |
| 1551 |
|
| 1552 |
/* |
| 1553 |
* the buddy cache inode stores the block bitmap |
| 1554 |
* and buddy information in consecutive blocks. |
| 1555 |
* So for each group we need two blocks. |
| 1556 |
*/ |
| 1557 |
block = group * 2; |
| 1558 |
pnum = EXT4_LBLK_TO_PG(inode, block); |
| 1559 |
folio = __filemap_get_folio(inode->i_mapping, pnum, |
| 1560 |
FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); |
| 1561 |
if (IS_ERR(folio)) |
| 1562 |
return PTR_ERR(folio); |
| 1563 |
BUG_ON(folio->mapping != inode->i_mapping); |
| 1564 |
WARN_ON_ONCE(folio_size(folio) < sb->s_blocksize); |
| 1565 |
e4b->bd_bitmap_folio = folio; |
| 1566 |
e4b->bd_bitmap = folio_address(folio) + |
| 1567 |
offset_in_folio(folio, EXT4_LBLK_TO_B(inode, block)); |
| 1568 |
|
| 1569 |
block++; |
| 1570 |
pnum = EXT4_LBLK_TO_PG(inode, block); |
| 1571 |
if (folio_contains(folio, pnum)) { |
| 1572 |
/* buddy and bitmap are on the same folio */ |
| 1573 |
return 0; |
| 1574 |
} |
| 1575 |
|
| 1576 |
/* we need another folio for the buddy */ |
| 1577 |
folio = __filemap_get_folio(inode->i_mapping, pnum, |
| 1578 |
FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); |
| 1579 |
if (IS_ERR(folio)) |
| 1580 |
return PTR_ERR(folio); |
| 1581 |
BUG_ON(folio->mapping != inode->i_mapping); |
| 1582 |
WARN_ON_ONCE(folio_size(folio) < sb->s_blocksize); |
| 1583 |
e4b->bd_buddy_folio = folio; |
| 1584 |
return 0; |
| 1585 |
} |
| 1586 |
|
| 1587 |
static void ext4_mb_put_buddy_folio_lock(struct ext4_buddy *e4b) |
| 1588 |
{ |
| 1589 |
if (e4b->bd_bitmap_folio) { |
| 1590 |
folio_unlock(e4b->bd_bitmap_folio); |
| 1591 |
folio_put(e4b->bd_bitmap_folio); |
| 1592 |
} |
| 1593 |
if (e4b->bd_buddy_folio) { |
| 1594 |
folio_unlock(e4b->bd_buddy_folio); |
| 1595 |
folio_put(e4b->bd_buddy_folio); |
| 1596 |
} |
| 1597 |
} |
| 1598 |
|
| 1599 |
/* |
| 1600 |
* Locking note: This routine calls ext4_mb_init_cache(), which takes the |
| 1601 |
* block group lock of all groups for this folio; do not hold the BG lock when |
| 1602 |
* calling this routine! |
| 1603 |
*/ |
| 1604 |
static noinline_for_stack |
| 1605 |
int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) |
| 1606 |
{ |
| 1607 |
|
| 1608 |
struct ext4_group_info *this_grp; |
| 1609 |
struct ext4_buddy e4b; |
| 1610 |
struct folio *folio; |
| 1611 |
int ret = 0; |
| 1612 |
|
| 1613 |
might_sleep(); |
| 1614 |
mb_debug(sb, "init group %u\n", group); |
| 1615 |
this_grp = ext4_get_group_info(sb, group); |
| 1616 |
if (!this_grp) |
| 1617 |
return -EFSCORRUPTED; |
| 1618 |
|
| 1619 |
/* |
| 1620 |
* This ensures that we don't reinit the buddy cache |
| 1621 |
* folio which map to the group from which we are already |
| 1622 |
* allocating. If we are looking at the buddy cache we would |
| 1623 |
* have taken a reference using ext4_mb_load_buddy and that |
| 1624 |
* would have pinned buddy folio to page cache. |
| 1625 |
* The call to ext4_mb_get_buddy_folio_lock will mark the |
| 1626 |
* folio accessed. |
| 1627 |
*/ |
| 1628 |
ret = ext4_mb_get_buddy_folio_lock(sb, group, &e4b, gfp); |
| 1629 |
if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { |
| 1630 |
/* |
| 1631 |
* somebody initialized the group |
| 1632 |
* return without doing anything |
| 1633 |
*/ |
| 1634 |
goto err; |
| 1635 |
} |
| 1636 |
|
| 1637 |
folio = e4b.bd_bitmap_folio; |
| 1638 |
ret = ext4_mb_init_cache(folio, NULL, gfp); |
| 1639 |
if (ret) |
| 1640 |
goto err; |
| 1641 |
if (!folio_test_uptodate(folio)) { |
| 1642 |
ret = -EIO; |
| 1643 |
goto err; |
| 1644 |
} |
| 1645 |
|
| 1646 |
if (e4b.bd_buddy_folio == NULL) { |
| 1647 |
/* |
| 1648 |
* If both the bitmap and buddy are in |
| 1649 |
* the same folio we don't need to force |
| 1650 |
* init the buddy |
| 1651 |
*/ |
| 1652 |
ret = 0; |
| 1653 |
goto err; |
| 1654 |
} |
| 1655 |
/* init buddy cache */ |
| 1656 |
folio = e4b.bd_buddy_folio; |
| 1657 |
ret = ext4_mb_init_cache(folio, e4b.bd_bitmap, gfp); |
| 1658 |
if (ret) |
| 1659 |
goto err; |
| 1660 |
if (!folio_test_uptodate(folio)) { |
| 1661 |
ret = -EIO; |
| 1662 |
goto err; |
| 1663 |
} |
| 1664 |
err: |
| 1665 |
ext4_mb_put_buddy_folio_lock(&e4b); |
| 1666 |
return ret; |
| 1667 |
} |
| 1668 |
|
| 1669 |
/* |
| 1670 |
* Locking note: This routine calls ext4_mb_init_cache(), which takes the |
| 1671 |
* block group lock of all groups for this folio; do not hold the BG lock when |
| 1672 |
* calling this routine! |
| 1673 |
*/ |
| 1674 |
static noinline_for_stack int |
| 1675 |
ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, |
| 1676 |
struct ext4_buddy *e4b, gfp_t gfp) |
| 1677 |
{ |
| 1678 |
int block; |
| 1679 |
int pnum; |
| 1680 |
struct folio *folio; |
| 1681 |
int ret; |
| 1682 |
struct ext4_group_info *grp; |
| 1683 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 1684 |
struct inode *inode = sbi->s_buddy_cache; |
| 1685 |
|
| 1686 |
might_sleep(); |
| 1687 |
mb_debug(sb, "load group %u\n", group); |
| 1688 |
|
| 1689 |
grp = ext4_get_group_info(sb, group); |
| 1690 |
if (!grp) |
| 1691 |
return -EFSCORRUPTED; |
| 1692 |
|
| 1693 |
e4b->bd_blkbits = sb->s_blocksize_bits; |
| 1694 |
e4b->bd_info = grp; |
| 1695 |
e4b->bd_sb = sb; |
| 1696 |
e4b->bd_group = group; |
| 1697 |
e4b->bd_buddy_folio = NULL; |
| 1698 |
e4b->bd_bitmap_folio = NULL; |
| 1699 |
|
| 1700 |
if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
| 1701 |
/* |
| 1702 |
* we need full data about the group |
| 1703 |
* to make a good selection |
| 1704 |
*/ |
| 1705 |
ret = ext4_mb_init_group(sb, group, gfp); |
| 1706 |
if (ret) |
| 1707 |
return ret; |
| 1708 |
} |
| 1709 |
|
| 1710 |
/* |
| 1711 |
* the buddy cache inode stores the block bitmap |
| 1712 |
* and buddy information in consecutive blocks. |
| 1713 |
* So for each group we need two blocks. |
| 1714 |
*/ |
| 1715 |
block = group * 2; |
| 1716 |
pnum = EXT4_LBLK_TO_PG(inode, block); |
| 1717 |
|
| 1718 |
/* Avoid locking the folio in the fast path ... */ |
| 1719 |
folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0); |
| 1720 |
if (IS_ERR(folio) || !folio_test_uptodate(folio) || folio_test_locked(folio)) { |
| 1721 |
/* |
| 1722 |
* folio_test_locked is employed to detect ongoing folio |
| 1723 |
* migrations, since concurrent migrations can lead to |
| 1724 |
* bitmap inconsistency. And if we are not uptodate that |
| 1725 |
* implies somebody just created the folio but is yet to |
| 1726 |
* initialize it. We can drop the folio reference and |
| 1727 |
* try to get the folio with lock in both cases to avoid |
| 1728 |
* concurrency. |
| 1729 |
*/ |
| 1730 |
if (!IS_ERR(folio)) |
| 1731 |
folio_put(folio); |
| 1732 |
folio = __filemap_get_folio(inode->i_mapping, pnum, |
| 1733 |
FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); |
| 1734 |
if (!IS_ERR(folio)) { |
| 1735 |
if (WARN_RATELIMIT(folio->mapping != inode->i_mapping, |
| 1736 |
"ext4: bitmap's mapping != inode->i_mapping\n")) { |
| 1737 |
/* should never happen */ |
| 1738 |
folio_unlock(folio); |
| 1739 |
ret = -EINVAL; |
| 1740 |
goto err; |
| 1741 |
} |
| 1742 |
if (!folio_test_uptodate(folio)) { |
| 1743 |
ret = ext4_mb_init_cache(folio, NULL, gfp); |
| 1744 |
if (ret) { |
| 1745 |
folio_unlock(folio); |
| 1746 |
goto err; |
| 1747 |
} |
| 1748 |
mb_cmp_bitmaps(e4b, folio_address(folio) + |
| 1749 |
offset_in_folio(folio, |
| 1750 |
EXT4_LBLK_TO_B(inode, block))); |
| 1751 |
} |
| 1752 |
folio_unlock(folio); |
| 1753 |
} |
| 1754 |
} |
| 1755 |
if (IS_ERR(folio)) { |
| 1756 |
ret = PTR_ERR(folio); |
| 1757 |
goto err; |
| 1758 |
} |
| 1759 |
if (!folio_test_uptodate(folio)) { |
| 1760 |
ret = -EIO; |
| 1761 |
goto err; |
| 1762 |
} |
| 1763 |
|
| 1764 |
/* Folios marked accessed already */ |
| 1765 |
e4b->bd_bitmap_folio = folio; |
| 1766 |
e4b->bd_bitmap = folio_address(folio) + |
| 1767 |
offset_in_folio(folio, EXT4_LBLK_TO_B(inode, block)); |
| 1768 |
|
| 1769 |
block++; |
| 1770 |
pnum = EXT4_LBLK_TO_PG(inode, block); |
| 1771 |
/* buddy and bitmap are on the same folio? */ |
| 1772 |
if (folio_contains(folio, pnum)) { |
| 1773 |
folio_get(folio); |
| 1774 |
goto update_buddy; |
| 1775 |
} |
| 1776 |
|
| 1777 |
/* we need another folio for the buddy */ |
| 1778 |
folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0); |
| 1779 |
if (IS_ERR(folio) || !folio_test_uptodate(folio) || folio_test_locked(folio)) { |
| 1780 |
if (!IS_ERR(folio)) |
| 1781 |
folio_put(folio); |
| 1782 |
folio = __filemap_get_folio(inode->i_mapping, pnum, |
| 1783 |
FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); |
| 1784 |
if (!IS_ERR(folio)) { |
| 1785 |
if (WARN_RATELIMIT(folio->mapping != inode->i_mapping, |
| 1786 |
"ext4: buddy bitmap's mapping != inode->i_mapping\n")) { |
| 1787 |
/* should never happen */ |
| 1788 |
folio_unlock(folio); |
| 1789 |
ret = -EINVAL; |
| 1790 |
goto err; |
| 1791 |
} |
| 1792 |
if (!folio_test_uptodate(folio)) { |
| 1793 |
ret = ext4_mb_init_cache(folio, e4b->bd_bitmap, |
| 1794 |
gfp); |
| 1795 |
if (ret) { |
| 1796 |
folio_unlock(folio); |
| 1797 |
goto err; |
| 1798 |
} |
| 1799 |
} |
| 1800 |
folio_unlock(folio); |
| 1801 |
} |
| 1802 |
} |
| 1803 |
if (IS_ERR(folio)) { |
| 1804 |
ret = PTR_ERR(folio); |
| 1805 |
goto err; |
| 1806 |
} |
| 1807 |
if (!folio_test_uptodate(folio)) { |
| 1808 |
ret = -EIO; |
| 1809 |
goto err; |
| 1810 |
} |
| 1811 |
|
| 1812 |
update_buddy: |
| 1813 |
/* Folios marked accessed already */ |
| 1814 |
e4b->bd_buddy_folio = folio; |
| 1815 |
e4b->bd_buddy = folio_address(folio) + |
| 1816 |
offset_in_folio(folio, EXT4_LBLK_TO_B(inode, block)); |
| 1817 |
|
| 1818 |
return 0; |
| 1819 |
|
| 1820 |
err: |
| 1821 |
if (!IS_ERR_OR_NULL(folio)) |
| 1822 |
folio_put(folio); |
| 1823 |
if (e4b->bd_bitmap_folio) |
| 1824 |
folio_put(e4b->bd_bitmap_folio); |
| 1825 |
|
| 1826 |
e4b->bd_buddy = NULL; |
| 1827 |
e4b->bd_bitmap = NULL; |
| 1828 |
return ret; |
| 1829 |
} |
| 1830 |
|
| 1831 |
static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, |
| 1832 |
struct ext4_buddy *e4b) |
| 1833 |
{ |
| 1834 |
return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); |
| 1835 |
} |
| 1836 |
|
| 1837 |
static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
| 1838 |
{ |
| 1839 |
if (e4b->bd_bitmap_folio) |
| 1840 |
folio_put(e4b->bd_bitmap_folio); |
| 1841 |
if (e4b->bd_buddy_folio) |
| 1842 |
folio_put(e4b->bd_buddy_folio); |
| 1843 |
} |
| 1844 |
|
| 1845 |
|
| 1846 |
static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) |
| 1847 |
{ |
| 1848 |
int order = 1, max; |
| 1849 |
void *bb; |
| 1850 |
|
| 1851 |
BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
| 1852 |
BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
| 1853 |
|
| 1854 |
while (order <= e4b->bd_blkbits + 1) { |
| 1855 |
bb = mb_find_buddy(e4b, order, &max); |
| 1856 |
if (!mb_test_bit(block >> order, bb)) { |
| 1857 |
/* this block is part of buddy of order 'order' */ |
| 1858 |
return order; |
| 1859 |
} |
| 1860 |
order++; |
| 1861 |
} |
| 1862 |
return 0; |
| 1863 |
} |
| 1864 |
|
| 1865 |
static void mb_clear_bits(void *bm, int cur, int len) |
| 1866 |
{ |
| 1867 |
__u32 *addr; |
| 1868 |
|
| 1869 |
len = cur + len; |
| 1870 |
while (cur < len) { |
| 1871 |
if ((cur & 31) == 0 && (len - cur) >= 32) { |
| 1872 |
/* fast path: clear whole word at once */ |
| 1873 |
addr = bm + (cur >> 3); |
| 1874 |
*addr = 0; |
| 1875 |
cur += 32; |
| 1876 |
continue; |
| 1877 |
} |
| 1878 |
mb_clear_bit(cur, bm); |
| 1879 |
cur++; |
| 1880 |
} |
| 1881 |
} |
| 1882 |
|
| 1883 |
/* clear bits in given range |
| 1884 |
* will return first found zero bit if any, -1 otherwise |
| 1885 |
*/ |
| 1886 |
static int mb_test_and_clear_bits(void *bm, int cur, int len) |
| 1887 |
{ |
| 1888 |
__u32 *addr; |
| 1889 |
int zero_bit = -1; |
| 1890 |
|
| 1891 |
len = cur + len; |
| 1892 |
while (cur < len) { |
| 1893 |
if ((cur & 31) == 0 && (len - cur) >= 32) { |
| 1894 |
/* fast path: clear whole word at once */ |
| 1895 |
addr = bm + (cur >> 3); |
| 1896 |
if (*addr != (__u32)(-1) && zero_bit == -1) |
| 1897 |
zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); |
| 1898 |
*addr = 0; |
| 1899 |
cur += 32; |
| 1900 |
continue; |
| 1901 |
} |
| 1902 |
if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) |
| 1903 |
zero_bit = cur; |
| 1904 |
cur++; |
| 1905 |
} |
| 1906 |
|
| 1907 |
return zero_bit; |
| 1908 |
} |
| 1909 |
|
| 1910 |
void mb_set_bits(void *bm, int cur, int len) |
| 1911 |
{ |
| 1912 |
__u32 *addr; |
| 1913 |
|
| 1914 |
len = cur + len; |
| 1915 |
while (cur < len) { |
| 1916 |
if ((cur & 31) == 0 && (len - cur) >= 32) { |
| 1917 |
/* fast path: set whole word at once */ |
| 1918 |
addr = bm + (cur >> 3); |
| 1919 |
*addr = 0xffffffff; |
| 1920 |
cur += 32; |
| 1921 |
continue; |
| 1922 |
} |
| 1923 |
mb_set_bit(cur, bm); |
| 1924 |
cur++; |
| 1925 |
} |
| 1926 |
} |
| 1927 |
|
| 1928 |
static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) |
| 1929 |
{ |
| 1930 |
if (mb_test_bit(*bit + side, bitmap)) { |
| 1931 |
mb_clear_bit(*bit, bitmap); |
| 1932 |
(*bit) -= side; |
| 1933 |
return 1; |
| 1934 |
} |
| 1935 |
else { |
| 1936 |
(*bit) += side; |
| 1937 |
mb_set_bit(*bit, bitmap); |
| 1938 |
return -1; |
| 1939 |
} |
| 1940 |
} |
| 1941 |
|
| 1942 |
static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) |
| 1943 |
{ |
| 1944 |
int max; |
| 1945 |
int order = 1; |
| 1946 |
void *buddy = mb_find_buddy(e4b, order, &max); |
| 1947 |
|
| 1948 |
while (buddy) { |
| 1949 |
void *buddy2; |
| 1950 |
|
| 1951 |
/* Bits in range [first; last] are known to be set since |
| 1952 |
* corresponding blocks were allocated. Bits in range |
| 1953 |
* (first; last) will stay set because they form buddies on |
| 1954 |
* upper layer. We just deal with borders if they don't |
| 1955 |
* align with upper layer and then go up. |
| 1956 |
* Releasing entire group is all about clearing |
| 1957 |
* single bit of highest order buddy. |
| 1958 |
*/ |
| 1959 |
|
| 1960 |
/* Example: |
| 1961 |
* --------------------------------- |
| 1962 |
* | 1 | 1 | 1 | 1 | |
| 1963 |
* --------------------------------- |
| 1964 |
* | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | |
| 1965 |
* --------------------------------- |
| 1966 |
* 0 1 2 3 4 5 6 7 |
| 1967 |
* \_____________________/ |
| 1968 |
* |
| 1969 |
* Neither [1] nor [6] is aligned to above layer. |
| 1970 |
* Left neighbour [0] is free, so mark it busy, |
| 1971 |
* decrease bb_counters and extend range to |
| 1972 |
* [0; 6] |
| 1973 |
* Right neighbour [7] is busy. It can't be coaleasced with [6], so |
| 1974 |
* mark [6] free, increase bb_counters and shrink range to |
| 1975 |
* [0; 5]. |
| 1976 |
* Then shift range to [0; 2], go up and do the same. |
| 1977 |
*/ |
| 1978 |
|
| 1979 |
|
| 1980 |
if (first & 1) |
| 1981 |
e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); |
| 1982 |
if (!(last & 1)) |
| 1983 |
e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); |
| 1984 |
if (first > last) |
| 1985 |
break; |
| 1986 |
order++; |
| 1987 |
|
| 1988 |
buddy2 = mb_find_buddy(e4b, order, &max); |
| 1989 |
if (!buddy2) { |
| 1990 |
mb_clear_bits(buddy, first, last - first + 1); |
| 1991 |
e4b->bd_info->bb_counters[order - 1] += last - first + 1; |
| 1992 |
break; |
| 1993 |
} |
| 1994 |
first >>= 1; |
| 1995 |
last >>= 1; |
| 1996 |
buddy = buddy2; |
| 1997 |
} |
| 1998 |
} |
| 1999 |
|
| 2000 |
static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
| 2001 |
int first, int count) |
| 2002 |
{ |
| 2003 |
int left_is_free = 0; |
| 2004 |
int right_is_free = 0; |
| 2005 |
int block; |
| 2006 |
int last = first + count - 1; |
| 2007 |
struct super_block *sb = e4b->bd_sb; |
| 2008 |
|
| 2009 |
if (WARN_ON(count == 0)) |
| 2010 |
return; |
| 2011 |
BUG_ON(last >= (sb->s_blocksize << 3)); |
| 2012 |
assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
| 2013 |
/* Don't bother if the block group is corrupt. */ |
| 2014 |
if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
| 2015 |
return; |
| 2016 |
|
| 2017 |
mb_check_buddy(e4b); |
| 2018 |
mb_free_blocks_double(inode, e4b, first, count); |
| 2019 |
|
| 2020 |
/* access memory sequentially: check left neighbour, |
| 2021 |
* clear range and then check right neighbour |
| 2022 |
*/ |
| 2023 |
if (first != 0) |
| 2024 |
left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
| 2025 |
block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); |
| 2026 |
if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) |
| 2027 |
right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); |
| 2028 |
|
| 2029 |
if (unlikely(block != -1)) { |
| 2030 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 2031 |
ext4_fsblk_t blocknr; |
| 2032 |
|
| 2033 |
/* |
| 2034 |
* Fastcommit replay can free already freed blocks which |
| 2035 |
* corrupts allocation info. Regenerate it. |
| 2036 |
*/ |
| 2037 |
if (sbi->s_mount_state & EXT4_FC_REPLAY) { |
| 2038 |
mb_regenerate_buddy(e4b); |
| 2039 |
goto check; |
| 2040 |
} |
| 2041 |
|
| 2042 |
blocknr = ext4_group_first_block_no(sb, e4b->bd_group); |
| 2043 |
blocknr += EXT4_C2B(sbi, block); |
| 2044 |
ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
| 2045 |
EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
| 2046 |
ext4_grp_locked_error(sb, e4b->bd_group, |
| 2047 |
inode ? inode->i_ino : 0, blocknr, |
| 2048 |
"freeing already freed block (bit %u); block bitmap corrupt.", |
| 2049 |
block); |
| 2050 |
return; |
| 2051 |
} |
| 2052 |
|
| 2053 |
this_cpu_inc(discard_pa_seq); |
| 2054 |
e4b->bd_info->bb_free += count; |
| 2055 |
if (first < e4b->bd_info->bb_first_free) |
| 2056 |
e4b->bd_info->bb_first_free = first; |
| 2057 |
|
| 2058 |
/* let's maintain fragments counter */ |
| 2059 |
if (left_is_free && right_is_free) |
| 2060 |
e4b->bd_info->bb_fragments--; |
| 2061 |
else if (!left_is_free && !right_is_free) |
| 2062 |
e4b->bd_info->bb_fragments++; |
| 2063 |
|
| 2064 |
/* buddy[0] == bd_bitmap is a special case, so handle |
| 2065 |
* it right away and let mb_buddy_mark_free stay free of |
| 2066 |
* zero order checks. |
| 2067 |
* Check if neighbours are to be coaleasced, |
| 2068 |
* adjust bitmap bb_counters and borders appropriately. |
| 2069 |
*/ |
| 2070 |
if (first & 1) { |
| 2071 |
first += !left_is_free; |
| 2072 |
e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; |
| 2073 |
} |
| 2074 |
if (!(last & 1)) { |
| 2075 |
last -= !right_is_free; |
| 2076 |
e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; |
| 2077 |
} |
| 2078 |
|
| 2079 |
if (first <= last) |
| 2080 |
mb_buddy_mark_free(e4b, first >> 1, last >> 1); |
| 2081 |
|
| 2082 |
mb_set_largest_free_order(sb, e4b->bd_info); |
| 2083 |
mb_update_avg_fragment_size(sb, e4b->bd_info); |
| 2084 |
check: |
| 2085 |
mb_check_buddy(e4b); |
| 2086 |
} |
| 2087 |
|
| 2088 |
static int mb_find_extent(struct ext4_buddy *e4b, int block, |
| 2089 |
int needed, struct ext4_free_extent *ex) |
| 2090 |
{ |
| 2091 |
int max, order, next; |
| 2092 |
void *buddy; |
| 2093 |
|
| 2094 |
assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
| 2095 |
BUG_ON(ex == NULL); |
| 2096 |
|
| 2097 |
buddy = mb_find_buddy(e4b, 0, &max); |
| 2098 |
BUG_ON(buddy == NULL); |
| 2099 |
BUG_ON(block >= max); |
| 2100 |
if (mb_test_bit(block, buddy)) { |
| 2101 |
ex->fe_len = 0; |
| 2102 |
ex->fe_start = 0; |
| 2103 |
ex->fe_group = 0; |
| 2104 |
return 0; |
| 2105 |
} |
| 2106 |
|
| 2107 |
/* find actual order */ |
| 2108 |
order = mb_find_order_for_block(e4b, block); |
| 2109 |
|
| 2110 |
ex->fe_len = (1 << order) - (block & ((1 << order) - 1)); |
| 2111 |
ex->fe_start = block; |
| 2112 |
ex->fe_group = e4b->bd_group; |
| 2113 |
|
| 2114 |
block = block >> order; |
| 2115 |
|
| 2116 |
while (needed > ex->fe_len && |
| 2117 |
mb_find_buddy(e4b, order, &max)) { |
| 2118 |
|
| 2119 |
if (block + 1 >= max) |
| 2120 |
break; |
| 2121 |
|
| 2122 |
next = (block + 1) * (1 << order); |
| 2123 |
if (mb_test_bit(next, e4b->bd_bitmap)) |
| 2124 |
break; |
| 2125 |
|
| 2126 |
order = mb_find_order_for_block(e4b, next); |
| 2127 |
|
| 2128 |
block = next >> order; |
| 2129 |
ex->fe_len += 1 << order; |
| 2130 |
} |
| 2131 |
|
| 2132 |
if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { |
| 2133 |
/* Should never happen! (but apparently sometimes does?!?) */ |
| 2134 |
WARN_ON(1); |
| 2135 |
ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, |
| 2136 |
"corruption or bug in mb_find_extent " |
| 2137 |
"block=%d, order=%d needed=%d ex=%u/%d/%d@%u", |
| 2138 |
block, order, needed, ex->fe_group, ex->fe_start, |
| 2139 |
ex->fe_len, ex->fe_logical); |
| 2140 |
ex->fe_len = 0; |
| 2141 |
ex->fe_start = 0; |
| 2142 |
ex->fe_group = 0; |
| 2143 |
} |
| 2144 |
return ex->fe_len; |
| 2145 |
} |
| 2146 |
|
| 2147 |
static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) |
| 2148 |
{ |
| 2149 |
int ord; |
| 2150 |
int mlen = 0; |
| 2151 |
int max = 0; |
| 2152 |
int start = ex->fe_start; |
| 2153 |
int len = ex->fe_len; |
| 2154 |
unsigned ret = 0; |
| 2155 |
int len0 = len; |
| 2156 |
void *buddy; |
| 2157 |
int ord_start, ord_end; |
| 2158 |
|
| 2159 |
BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); |
| 2160 |
BUG_ON(e4b->bd_group != ex->fe_group); |
| 2161 |
assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
| 2162 |
mb_check_buddy(e4b); |
| 2163 |
mb_mark_used_double(e4b, start, len); |
| 2164 |
|
| 2165 |
this_cpu_inc(discard_pa_seq); |
| 2166 |
e4b->bd_info->bb_free -= len; |
| 2167 |
if (e4b->bd_info->bb_first_free == start) |
| 2168 |
e4b->bd_info->bb_first_free += len; |
| 2169 |
|
| 2170 |
/* let's maintain fragments counter */ |
| 2171 |
if (start != 0) |
| 2172 |
mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
| 2173 |
if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
| 2174 |
max = !mb_test_bit(start + len, e4b->bd_bitmap); |
| 2175 |
if (mlen && max) |
| 2176 |
e4b->bd_info->bb_fragments++; |
| 2177 |
else if (!mlen && !max) |
| 2178 |
e4b->bd_info->bb_fragments--; |
| 2179 |
|
| 2180 |
/* let's maintain buddy itself */ |
| 2181 |
while (len) { |
| 2182 |
ord = mb_find_order_for_block(e4b, start); |
| 2183 |
|
| 2184 |
if (((start >> ord) << ord) == start && len >= (1 << ord)) { |
| 2185 |
/* the whole chunk may be allocated at once! */ |
| 2186 |
mlen = 1 << ord; |
| 2187 |
buddy = mb_find_buddy(e4b, ord, &max); |
| 2188 |
BUG_ON((start >> ord) >= max); |
| 2189 |
mb_set_bit(start >> ord, buddy); |
| 2190 |
e4b->bd_info->bb_counters[ord]--; |
| 2191 |
start += mlen; |
| 2192 |
len -= mlen; |
| 2193 |
BUG_ON(len < 0); |
| 2194 |
continue; |
| 2195 |
} |
| 2196 |
|
| 2197 |
/* store for history */ |
| 2198 |
if (ret == 0) |
| 2199 |
ret = len | (ord << 16); |
| 2200 |
|
| 2201 |
BUG_ON(ord <= 0); |
| 2202 |
buddy = mb_find_buddy(e4b, ord, &max); |
| 2203 |
mb_set_bit(start >> ord, buddy); |
| 2204 |
e4b->bd_info->bb_counters[ord]--; |
| 2205 |
|
| 2206 |
ord_start = (start >> ord) << ord; |
| 2207 |
ord_end = ord_start + (1 << ord); |
| 2208 |
/* first chunk */ |
| 2209 |
if (start > ord_start) |
| 2210 |
ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy, |
| 2211 |
ord_start, start - ord_start, |
| 2212 |
e4b->bd_info); |
| 2213 |
|
| 2214 |
/* last chunk */ |
| 2215 |
if (start + len < ord_end) { |
| 2216 |
ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy, |
| 2217 |
start + len, |
| 2218 |
ord_end - (start + len), |
| 2219 |
e4b->bd_info); |
| 2220 |
break; |
| 2221 |
} |
| 2222 |
len = start + len - ord_end; |
| 2223 |
start = ord_end; |
| 2224 |
} |
| 2225 |
mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
| 2226 |
|
| 2227 |
mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); |
| 2228 |
mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
| 2229 |
mb_check_buddy(e4b); |
| 2230 |
|
| 2231 |
return ret; |
| 2232 |
} |
| 2233 |
|
| 2234 |
/* |
| 2235 |
* Must be called under group lock! |
| 2236 |
*/ |
| 2237 |
static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, |
| 2238 |
struct ext4_buddy *e4b) |
| 2239 |
{ |
| 2240 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 2241 |
int ret; |
| 2242 |
|
| 2243 |
BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); |
| 2244 |
BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
| 2245 |
|
| 2246 |
ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); |
| 2247 |
ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; |
| 2248 |
ret = mb_mark_used(e4b, &ac->ac_b_ex); |
| 2249 |
|
| 2250 |
/* preallocation can change ac_b_ex, thus we store actually |
| 2251 |
* allocated blocks for history */ |
| 2252 |
ac->ac_f_ex = ac->ac_b_ex; |
| 2253 |
|
| 2254 |
ac->ac_status = AC_STATUS_FOUND; |
| 2255 |
ac->ac_tail = ret & 0xffff; |
| 2256 |
ac->ac_buddy = ret >> 16; |
| 2257 |
|
| 2258 |
/* |
| 2259 |
* take the folio reference. We want the folio to be pinned |
| 2260 |
* so that we don't get a ext4_mb_init_cache_call for this |
| 2261 |
* group until we update the bitmap. That would mean we |
| 2262 |
* double allocate blocks. The reference is dropped |
| 2263 |
* in ext4_mb_release_context |
| 2264 |
*/ |
| 2265 |
ac->ac_bitmap_folio = e4b->bd_bitmap_folio; |
| 2266 |
folio_get(ac->ac_bitmap_folio); |
| 2267 |
ac->ac_buddy_folio = e4b->bd_buddy_folio; |
| 2268 |
folio_get(ac->ac_buddy_folio); |
| 2269 |
/* store last allocated for subsequent stream allocation */ |
| 2270 |
if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
| 2271 |
int hash = ac->ac_inode->i_ino % sbi->s_mb_nr_global_goals; |
| 2272 |
|
| 2273 |
WRITE_ONCE(sbi->s_mb_last_groups[hash], ac->ac_f_ex.fe_group); |
| 2274 |
} |
| 2275 |
|
| 2276 |
/* |
| 2277 |
* As we've just preallocated more space than |
| 2278 |
* user requested originally, we store allocated |
| 2279 |
* space in a special descriptor. |
| 2280 |
*/ |
| 2281 |
if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) |
| 2282 |
ext4_mb_new_preallocation(ac); |
| 2283 |
|
| 2284 |
} |
| 2285 |
|
| 2286 |
static void ext4_mb_check_limits(struct ext4_allocation_context *ac, |
| 2287 |
struct ext4_buddy *e4b, |
| 2288 |
int finish_group) |
| 2289 |
{ |
| 2290 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 2291 |
struct ext4_free_extent *bex = &ac->ac_b_ex; |
| 2292 |
struct ext4_free_extent *gex = &ac->ac_g_ex; |
| 2293 |
|
| 2294 |
if (ac->ac_status == AC_STATUS_FOUND) |
| 2295 |
return; |
| 2296 |
/* |
| 2297 |
* We don't want to scan for a whole year |
| 2298 |
*/ |
| 2299 |
if (ac->ac_found > sbi->s_mb_max_to_scan && |
| 2300 |
!(ac->ac_flags & EXT4_MB_HINT_FIRST)) { |
| 2301 |
ac->ac_status = AC_STATUS_BREAK; |
| 2302 |
return; |
| 2303 |
} |
| 2304 |
|
| 2305 |
/* |
| 2306 |
* Haven't found good chunk so far, let's continue |
| 2307 |
*/ |
| 2308 |
if (bex->fe_len < gex->fe_len) |
| 2309 |
return; |
| 2310 |
|
| 2311 |
if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan) |
| 2312 |
ext4_mb_use_best_found(ac, e4b); |
| 2313 |
} |
| 2314 |
|
| 2315 |
/* |
| 2316 |
* The routine checks whether found extent is good enough. If it is, |
| 2317 |
* then the extent gets marked used and flag is set to the context |
| 2318 |
* to stop scanning. Otherwise, the extent is compared with the |
| 2319 |
* previous found extent and if new one is better, then it's stored |
| 2320 |
* in the context. Later, the best found extent will be used, if |
| 2321 |
* mballoc can't find good enough extent. |
| 2322 |
* |
| 2323 |
* The algorithm used is roughly as follows: |
| 2324 |
* |
| 2325 |
* * If free extent found is exactly as big as goal, then |
| 2326 |
* stop the scan and use it immediately |
| 2327 |
* |
| 2328 |
* * If free extent found is smaller than goal, then keep retrying |
| 2329 |
* upto a max of sbi->s_mb_max_to_scan times (default 200). After |
| 2330 |
* that stop scanning and use whatever we have. |
| 2331 |
* |
| 2332 |
* * If free extent found is bigger than goal, then keep retrying |
| 2333 |
* upto a max of sbi->s_mb_min_to_scan times (default 10) before |
| 2334 |
* stopping the scan and using the extent. |
| 2335 |
* |
| 2336 |
* |
| 2337 |
* FIXME: real allocation policy is to be designed yet! |
| 2338 |
*/ |
| 2339 |
static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, |
| 2340 |
struct ext4_free_extent *ex, |
| 2341 |
struct ext4_buddy *e4b) |
| 2342 |
{ |
| 2343 |
struct ext4_free_extent *bex = &ac->ac_b_ex; |
| 2344 |
struct ext4_free_extent *gex = &ac->ac_g_ex; |
| 2345 |
|
| 2346 |
BUG_ON(ex->fe_len <= 0); |
| 2347 |
BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
| 2348 |
BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
| 2349 |
BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
| 2350 |
|
| 2351 |
ac->ac_found++; |
| 2352 |
ac->ac_cX_found[ac->ac_criteria]++; |
| 2353 |
|
| 2354 |
/* |
| 2355 |
* The special case - take what you catch first |
| 2356 |
*/ |
| 2357 |
if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { |
| 2358 |
*bex = *ex; |
| 2359 |
ext4_mb_use_best_found(ac, e4b); |
| 2360 |
return; |
| 2361 |
} |
| 2362 |
|
| 2363 |
/* |
| 2364 |
* Let's check whether the chuck is good enough |
| 2365 |
*/ |
| 2366 |
if (ex->fe_len == gex->fe_len) { |
| 2367 |
*bex = *ex; |
| 2368 |
ext4_mb_use_best_found(ac, e4b); |
| 2369 |
return; |
| 2370 |
} |
| 2371 |
|
| 2372 |
/* |
| 2373 |
* If this is first found extent, just store it in the context |
| 2374 |
*/ |
| 2375 |
if (bex->fe_len == 0) { |
| 2376 |
*bex = *ex; |
| 2377 |
return; |
| 2378 |
} |
| 2379 |
|
| 2380 |
/* |
| 2381 |
* If new found extent is better, store it in the context |
| 2382 |
*/ |
| 2383 |
if (bex->fe_len < gex->fe_len) { |
| 2384 |
/* if the request isn't satisfied, any found extent |
| 2385 |
* larger than previous best one is better */ |
| 2386 |
if (ex->fe_len > bex->fe_len) |
| 2387 |
*bex = *ex; |
| 2388 |
} else if (ex->fe_len > gex->fe_len) { |
| 2389 |
/* if the request is satisfied, then we try to find |
| 2390 |
* an extent that still satisfy the request, but is |
| 2391 |
* smaller than previous one */ |
| 2392 |
if (ex->fe_len < bex->fe_len) |
| 2393 |
*bex = *ex; |
| 2394 |
} |
| 2395 |
|
| 2396 |
ext4_mb_check_limits(ac, e4b, 0); |
| 2397 |
} |
| 2398 |
|
| 2399 |
static noinline_for_stack |
| 2400 |
void ext4_mb_try_best_found(struct ext4_allocation_context *ac, |
| 2401 |
struct ext4_buddy *e4b) |
| 2402 |
{ |
| 2403 |
struct ext4_free_extent ex = ac->ac_b_ex; |
| 2404 |
ext4_group_t group = ex.fe_group; |
| 2405 |
int max; |
| 2406 |
int err; |
| 2407 |
|
| 2408 |
BUG_ON(ex.fe_len <= 0); |
| 2409 |
err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); |
| 2410 |
if (err) |
| 2411 |
return; |
| 2412 |
|
| 2413 |
ext4_lock_group(ac->ac_sb, group); |
| 2414 |
if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
| 2415 |
goto out; |
| 2416 |
|
| 2417 |
max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
| 2418 |
|
| 2419 |
if (max > 0) { |
| 2420 |
ac->ac_b_ex = ex; |
| 2421 |
ext4_mb_use_best_found(ac, e4b); |
| 2422 |
} |
| 2423 |
|
| 2424 |
out: |
| 2425 |
ext4_unlock_group(ac->ac_sb, group); |
| 2426 |
ext4_mb_unload_buddy(e4b); |
| 2427 |
} |
| 2428 |
|
| 2429 |
static noinline_for_stack |
| 2430 |
int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, |
| 2431 |
struct ext4_buddy *e4b) |
| 2432 |
{ |
| 2433 |
ext4_group_t group = ac->ac_g_ex.fe_group; |
| 2434 |
int max; |
| 2435 |
int err; |
| 2436 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 2437 |
struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
| 2438 |
struct ext4_free_extent ex; |
| 2439 |
|
| 2440 |
if (!grp) |
| 2441 |
return -EFSCORRUPTED; |
| 2442 |
if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY))) |
| 2443 |
return 0; |
| 2444 |
if (grp->bb_free == 0) |
| 2445 |
return 0; |
| 2446 |
|
| 2447 |
err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); |
| 2448 |
if (err) { |
| 2449 |
if (EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info) && |
| 2450 |
!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
| 2451 |
return 0; |
| 2452 |
return err; |
| 2453 |
} |
| 2454 |
|
| 2455 |
ext4_lock_group(ac->ac_sb, group); |
| 2456 |
if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
| 2457 |
goto out; |
| 2458 |
|
| 2459 |
max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
| 2460 |
ac->ac_g_ex.fe_len, &ex); |
| 2461 |
ex.fe_logical = 0xDEADFA11; /* debug value */ |
| 2462 |
|
| 2463 |
if (max >= ac->ac_g_ex.fe_len && |
| 2464 |
ac->ac_g_ex.fe_len == EXT4_NUM_B2C(sbi, sbi->s_stripe)) { |
| 2465 |
ext4_fsblk_t start; |
| 2466 |
|
| 2467 |
start = ext4_grp_offs_to_block(ac->ac_sb, &ex); |
| 2468 |
/* use do_div to get remainder (would be 64-bit modulo) */ |
| 2469 |
if (do_div(start, sbi->s_stripe) == 0) { |
| 2470 |
ac->ac_found++; |
| 2471 |
ac->ac_b_ex = ex; |
| 2472 |
ext4_mb_use_best_found(ac, e4b); |
| 2473 |
} |
| 2474 |
} else if (max >= ac->ac_g_ex.fe_len) { |
| 2475 |
BUG_ON(ex.fe_len <= 0); |
| 2476 |
BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); |
| 2477 |
BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); |
| 2478 |
ac->ac_found++; |
| 2479 |
ac->ac_b_ex = ex; |
| 2480 |
ext4_mb_use_best_found(ac, e4b); |
| 2481 |
} else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { |
| 2482 |
/* Sometimes, caller may want to merge even small |
| 2483 |
* number of blocks to an existing extent */ |
| 2484 |
BUG_ON(ex.fe_len <= 0); |
| 2485 |
BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); |
| 2486 |
BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); |
| 2487 |
ac->ac_found++; |
| 2488 |
ac->ac_b_ex = ex; |
| 2489 |
ext4_mb_use_best_found(ac, e4b); |
| 2490 |
} |
| 2491 |
out: |
| 2492 |
ext4_unlock_group(ac->ac_sb, group); |
| 2493 |
ext4_mb_unload_buddy(e4b); |
| 2494 |
|
| 2495 |
return 0; |
| 2496 |
} |
| 2497 |
|
| 2498 |
/* |
| 2499 |
* The routine scans buddy structures (not bitmap!) from given order |
| 2500 |
* to max order and tries to find big enough chunk to satisfy the req |
| 2501 |
*/ |
| 2502 |
static noinline_for_stack |
| 2503 |
void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, |
| 2504 |
struct ext4_buddy *e4b) |
| 2505 |
{ |
| 2506 |
struct super_block *sb = ac->ac_sb; |
| 2507 |
struct ext4_group_info *grp = e4b->bd_info; |
| 2508 |
void *buddy; |
| 2509 |
int i; |
| 2510 |
int k; |
| 2511 |
int max; |
| 2512 |
|
| 2513 |
BUG_ON(ac->ac_2order <= 0); |
| 2514 |
for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { |
| 2515 |
if (grp->bb_counters[i] == 0) |
| 2516 |
continue; |
| 2517 |
|
| 2518 |
buddy = mb_find_buddy(e4b, i, &max); |
| 2519 |
if (WARN_RATELIMIT(buddy == NULL, |
| 2520 |
"ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i)) |
| 2521 |
continue; |
| 2522 |
|
| 2523 |
k = mb_find_next_zero_bit(buddy, max, 0); |
| 2524 |
if (k >= max) { |
| 2525 |
ext4_mark_group_bitmap_corrupted(ac->ac_sb, |
| 2526 |
e4b->bd_group, |
| 2527 |
EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
| 2528 |
ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, |
| 2529 |
"%d free clusters of order %d. But found 0", |
| 2530 |
grp->bb_counters[i], i); |
| 2531 |
break; |
| 2532 |
} |
| 2533 |
ac->ac_found++; |
| 2534 |
ac->ac_cX_found[ac->ac_criteria]++; |
| 2535 |
|
| 2536 |
ac->ac_b_ex.fe_len = 1 << i; |
| 2537 |
ac->ac_b_ex.fe_start = k << i; |
| 2538 |
ac->ac_b_ex.fe_group = e4b->bd_group; |
| 2539 |
|
| 2540 |
ext4_mb_use_best_found(ac, e4b); |
| 2541 |
|
| 2542 |
BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); |
| 2543 |
|
| 2544 |
if (EXT4_SB(sb)->s_mb_stats) |
| 2545 |
atomic_inc(&EXT4_SB(sb)->s_bal_2orders); |
| 2546 |
|
| 2547 |
break; |
| 2548 |
} |
| 2549 |
} |
| 2550 |
|
| 2551 |
/* |
| 2552 |
* The routine scans the group and measures all found extents. |
| 2553 |
* In order to optimize scanning, caller must pass number of |
| 2554 |
* free blocks in the group, so the routine can know upper limit. |
| 2555 |
*/ |
| 2556 |
static noinline_for_stack |
| 2557 |
void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, |
| 2558 |
struct ext4_buddy *e4b) |
| 2559 |
{ |
| 2560 |
struct super_block *sb = ac->ac_sb; |
| 2561 |
void *bitmap = e4b->bd_bitmap; |
| 2562 |
struct ext4_free_extent ex; |
| 2563 |
int i, j, freelen; |
| 2564 |
int free; |
| 2565 |
|
| 2566 |
free = e4b->bd_info->bb_free; |
| 2567 |
if (WARN_ON(free <= 0)) |
| 2568 |
return; |
| 2569 |
|
| 2570 |
i = e4b->bd_info->bb_first_free; |
| 2571 |
|
| 2572 |
while (free && ac->ac_status == AC_STATUS_CONTINUE) { |
| 2573 |
i = mb_find_next_zero_bit(bitmap, |
| 2574 |
EXT4_CLUSTERS_PER_GROUP(sb), i); |
| 2575 |
if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { |
| 2576 |
/* |
| 2577 |
* IF we have corrupt bitmap, we won't find any |
| 2578 |
* free blocks even though group info says we |
| 2579 |
* have free blocks |
| 2580 |
*/ |
| 2581 |
ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
| 2582 |
EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
| 2583 |
ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
| 2584 |
"%d free clusters as per " |
| 2585 |
"group info. But bitmap says 0", |
| 2586 |
free); |
| 2587 |
break; |
| 2588 |
} |
| 2589 |
|
| 2590 |
if (!ext4_mb_cr_expensive(ac->ac_criteria)) { |
| 2591 |
/* |
| 2592 |
* In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are |
| 2593 |
* sure that this group will have a large enough |
| 2594 |
* continuous free extent, so skip over the smaller free |
| 2595 |
* extents |
| 2596 |
*/ |
| 2597 |
j = mb_find_next_bit(bitmap, |
| 2598 |
EXT4_CLUSTERS_PER_GROUP(sb), i); |
| 2599 |
freelen = j - i; |
| 2600 |
|
| 2601 |
if (freelen < ac->ac_g_ex.fe_len) { |
| 2602 |
i = j; |
| 2603 |
free -= freelen; |
| 2604 |
continue; |
| 2605 |
} |
| 2606 |
} |
| 2607 |
|
| 2608 |
mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
| 2609 |
if (WARN_ON(ex.fe_len <= 0)) |
| 2610 |
break; |
| 2611 |
if (free < ex.fe_len) { |
| 2612 |
ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, |
| 2613 |
EXT4_GROUP_INFO_BBITMAP_CORRUPT); |
| 2614 |
ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
| 2615 |
"%d free clusters as per " |
| 2616 |
"group info. But got %d blocks", |
| 2617 |
free, ex.fe_len); |
| 2618 |
/* |
| 2619 |
* The number of free blocks differs. This mostly |
| 2620 |
* indicate that the bitmap is corrupt. So exit |
| 2621 |
* without claiming the space. |
| 2622 |
*/ |
| 2623 |
break; |
| 2624 |
} |
| 2625 |
ex.fe_logical = 0xDEADC0DE; /* debug value */ |
| 2626 |
ext4_mb_measure_extent(ac, &ex, e4b); |
| 2627 |
|
| 2628 |
i += ex.fe_len; |
| 2629 |
free -= ex.fe_len; |
| 2630 |
} |
| 2631 |
|
| 2632 |
ext4_mb_check_limits(ac, e4b, 1); |
| 2633 |
} |
| 2634 |
|
| 2635 |
/* |
| 2636 |
* This is a special case for storages like raid5 |
| 2637 |
* we try to find stripe-aligned chunks for stripe-size-multiple requests |
| 2638 |
*/ |
| 2639 |
static noinline_for_stack |
| 2640 |
void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, |
| 2641 |
struct ext4_buddy *e4b) |
| 2642 |
{ |
| 2643 |
struct super_block *sb = ac->ac_sb; |
| 2644 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 2645 |
void *bitmap = e4b->bd_bitmap; |
| 2646 |
struct ext4_free_extent ex; |
| 2647 |
ext4_fsblk_t first_group_block; |
| 2648 |
ext4_fsblk_t a; |
| 2649 |
ext4_grpblk_t i, stripe; |
| 2650 |
int max; |
| 2651 |
|
| 2652 |
BUG_ON(sbi->s_stripe == 0); |
| 2653 |
|
| 2654 |
/* find first stripe-aligned block in group */ |
| 2655 |
first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
| 2656 |
|
| 2657 |
a = first_group_block + sbi->s_stripe - 1; |
| 2658 |
do_div(a, sbi->s_stripe); |
| 2659 |
i = (a * sbi->s_stripe) - first_group_block; |
| 2660 |
|
| 2661 |
stripe = EXT4_NUM_B2C(sbi, sbi->s_stripe); |
| 2662 |
i = EXT4_B2C(sbi, i); |
| 2663 |
while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
| 2664 |
if (!mb_test_bit(i, bitmap)) { |
| 2665 |
max = mb_find_extent(e4b, i, stripe, &ex); |
| 2666 |
if (max >= stripe) { |
| 2667 |
ac->ac_found++; |
| 2668 |
ac->ac_cX_found[ac->ac_criteria]++; |
| 2669 |
ex.fe_logical = 0xDEADF00D; /* debug value */ |
| 2670 |
ac->ac_b_ex = ex; |
| 2671 |
ext4_mb_use_best_found(ac, e4b); |
| 2672 |
break; |
| 2673 |
} |
| 2674 |
} |
| 2675 |
i += stripe; |
| 2676 |
} |
| 2677 |
} |
| 2678 |
|
| 2679 |
static void __ext4_mb_scan_group(struct ext4_allocation_context *ac) |
| 2680 |
{ |
| 2681 |
bool is_stripe_aligned; |
| 2682 |
struct ext4_sb_info *sbi; |
| 2683 |
enum criteria cr = ac->ac_criteria; |
| 2684 |
|
| 2685 |
ac->ac_groups_scanned++; |
| 2686 |
if (cr == CR_POWER2_ALIGNED) |
| 2687 |
return ext4_mb_simple_scan_group(ac, ac->ac_e4b); |
| 2688 |
|
| 2689 |
sbi = EXT4_SB(ac->ac_sb); |
| 2690 |
is_stripe_aligned = false; |
| 2691 |
if ((sbi->s_stripe >= sbi->s_cluster_ratio) && |
| 2692 |
!(ac->ac_g_ex.fe_len % EXT4_NUM_B2C(sbi, sbi->s_stripe))) |
| 2693 |
is_stripe_aligned = true; |
| 2694 |
|
| 2695 |
if ((cr == CR_GOAL_LEN_FAST || cr == CR_BEST_AVAIL_LEN) && |
| 2696 |
is_stripe_aligned) |
| 2697 |
ext4_mb_scan_aligned(ac, ac->ac_e4b); |
| 2698 |
|
| 2699 |
if (ac->ac_status == AC_STATUS_CONTINUE) |
| 2700 |
ext4_mb_complex_scan_group(ac, ac->ac_e4b); |
| 2701 |
} |
| 2702 |
|
| 2703 |
/* |
| 2704 |
* This is also called BEFORE we load the buddy bitmap. |
| 2705 |
* Returns either 1 or 0 indicating that the group is either suitable |
| 2706 |
* for the allocation or not. |
| 2707 |
*/ |
| 2708 |
static bool ext4_mb_good_group(struct ext4_allocation_context *ac, |
| 2709 |
ext4_group_t group, enum criteria cr) |
| 2710 |
{ |
| 2711 |
ext4_grpblk_t free, fragments; |
| 2712 |
int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
| 2713 |
struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
| 2714 |
|
| 2715 |
BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS); |
| 2716 |
|
| 2717 |
if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
| 2718 |
return false; |
| 2719 |
|
| 2720 |
free = grp->bb_free; |
| 2721 |
if (free == 0) |
| 2722 |
return false; |
| 2723 |
|
| 2724 |
fragments = grp->bb_fragments; |
| 2725 |
if (fragments == 0) |
| 2726 |
return false; |
| 2727 |
|
| 2728 |
switch (cr) { |
| 2729 |
case CR_POWER2_ALIGNED: |
| 2730 |
BUG_ON(ac->ac_2order == 0); |
| 2731 |
|
| 2732 |
/* Avoid using the first bg of a flexgroup for data files */ |
| 2733 |
if ((ac->ac_flags & EXT4_MB_HINT_DATA) && |
| 2734 |
(flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && |
| 2735 |
((group % flex_size) == 0)) |
| 2736 |
return false; |
| 2737 |
|
| 2738 |
if (free < ac->ac_g_ex.fe_len) |
| 2739 |
return false; |
| 2740 |
|
| 2741 |
if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) |
| 2742 |
return true; |
| 2743 |
|
| 2744 |
if (grp->bb_largest_free_order < ac->ac_2order) |
| 2745 |
return false; |
| 2746 |
|
| 2747 |
return true; |
| 2748 |
case CR_GOAL_LEN_FAST: |
| 2749 |
case CR_BEST_AVAIL_LEN: |
| 2750 |
if ((free / fragments) >= ac->ac_g_ex.fe_len) |
| 2751 |
return true; |
| 2752 |
break; |
| 2753 |
case CR_GOAL_LEN_SLOW: |
| 2754 |
if (free >= ac->ac_g_ex.fe_len) |
| 2755 |
return true; |
| 2756 |
break; |
| 2757 |
case CR_ANY_FREE: |
| 2758 |
return true; |
| 2759 |
default: |
| 2760 |
BUG(); |
| 2761 |
} |
| 2762 |
|
| 2763 |
return false; |
| 2764 |
} |
| 2765 |
|
| 2766 |
/* |
| 2767 |
* This could return negative error code if something goes wrong |
| 2768 |
* during ext4_mb_init_group(). This should not be called with |
| 2769 |
* ext4_lock_group() held. |
| 2770 |
* |
| 2771 |
* Note: because we are conditionally operating with the group lock in |
| 2772 |
* the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this |
| 2773 |
* function using __acquire and __release. This means we need to be |
| 2774 |
* super careful before messing with the error path handling via "goto |
| 2775 |
* out"! |
| 2776 |
*/ |
| 2777 |
static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, |
| 2778 |
ext4_group_t group, enum criteria cr) |
| 2779 |
{ |
| 2780 |
struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
| 2781 |
struct super_block *sb = ac->ac_sb; |
| 2782 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 2783 |
bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; |
| 2784 |
ext4_grpblk_t free; |
| 2785 |
int ret = 0; |
| 2786 |
|
| 2787 |
if (!grp) |
| 2788 |
return -EFSCORRUPTED; |
| 2789 |
if (sbi->s_mb_stats) |
| 2790 |
atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); |
| 2791 |
if (should_lock) { |
| 2792 |
ext4_lock_group(sb, group); |
| 2793 |
__release(ext4_group_lock_ptr(sb, group)); |
| 2794 |
} |
| 2795 |
free = grp->bb_free; |
| 2796 |
if (free == 0) |
| 2797 |
goto out; |
| 2798 |
/* |
| 2799 |
* In all criterias except CR_ANY_FREE we try to avoid groups that |
| 2800 |
* can't possibly satisfy the full goal request due to insufficient |
| 2801 |
* free blocks. |
| 2802 |
*/ |
| 2803 |
if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len) |
| 2804 |
goto out; |
| 2805 |
if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
| 2806 |
goto out; |
| 2807 |
if (should_lock) { |
| 2808 |
__acquire(ext4_group_lock_ptr(sb, group)); |
| 2809 |
ext4_unlock_group(sb, group); |
| 2810 |
} |
| 2811 |
|
| 2812 |
/* We only do this if the grp has never been initialized */ |
| 2813 |
if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
| 2814 |
struct ext4_group_desc *gdp = |
| 2815 |
ext4_get_group_desc(sb, group, NULL); |
| 2816 |
int ret; |
| 2817 |
|
| 2818 |
/* |
| 2819 |
* CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic |
| 2820 |
* search to find large good chunks almost for free. If buddy |
| 2821 |
* data is not ready, then this optimization makes no sense. But |
| 2822 |
* we never skip the first block group in a flex_bg, since this |
| 2823 |
* gets used for metadata block allocation, and we want to make |
| 2824 |
* sure we locate metadata blocks in the first block group in |
| 2825 |
* the flex_bg if possible. |
| 2826 |
*/ |
| 2827 |
if (!ext4_mb_cr_expensive(cr) && |
| 2828 |
(!sbi->s_log_groups_per_flex || |
| 2829 |
((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && |
| 2830 |
!(ext4_has_group_desc_csum(sb) && |
| 2831 |
(gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) |
| 2832 |
return 0; |
| 2833 |
ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
| 2834 |
if (ret) |
| 2835 |
return ret; |
| 2836 |
} |
| 2837 |
|
| 2838 |
if (should_lock) { |
| 2839 |
ext4_lock_group(sb, group); |
| 2840 |
__release(ext4_group_lock_ptr(sb, group)); |
| 2841 |
} |
| 2842 |
ret = ext4_mb_good_group(ac, group, cr); |
| 2843 |
out: |
| 2844 |
if (should_lock) { |
| 2845 |
__acquire(ext4_group_lock_ptr(sb, group)); |
| 2846 |
ext4_unlock_group(sb, group); |
| 2847 |
} |
| 2848 |
return ret; |
| 2849 |
} |
| 2850 |
|
| 2851 |
/* |
| 2852 |
* Start prefetching @nr block bitmaps starting at @group. |
| 2853 |
* Return the next group which needs to be prefetched. |
| 2854 |
*/ |
| 2855 |
ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, |
| 2856 |
unsigned int nr, int *cnt) |
| 2857 |
{ |
| 2858 |
ext4_group_t ngroups = ext4_get_groups_count(sb); |
| 2859 |
struct buffer_head *bh; |
| 2860 |
struct blk_plug plug; |
| 2861 |
|
| 2862 |
blk_start_plug(&plug); |
| 2863 |
while (nr-- > 0) { |
| 2864 |
struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, |
| 2865 |
NULL); |
| 2866 |
struct ext4_group_info *grp = ext4_get_group_info(sb, group); |
| 2867 |
|
| 2868 |
/* |
| 2869 |
* Prefetch block groups with free blocks; but don't |
| 2870 |
* bother if it is marked uninitialized on disk, since |
| 2871 |
* it won't require I/O to read. Also only try to |
| 2872 |
* prefetch once, so we avoid getblk() call, which can |
| 2873 |
* be expensive. |
| 2874 |
*/ |
| 2875 |
if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) && |
| 2876 |
EXT4_MB_GRP_NEED_INIT(grp) && |
| 2877 |
ext4_free_group_clusters(sb, gdp) > 0 ) { |
| 2878 |
bh = ext4_read_block_bitmap_nowait(sb, group, true); |
| 2879 |
if (bh && !IS_ERR(bh)) { |
| 2880 |
if (!buffer_uptodate(bh) && cnt) |
| 2881 |
(*cnt)++; |
| 2882 |
brelse(bh); |
| 2883 |
} |
| 2884 |
} |
| 2885 |
if (++group >= ngroups) |
| 2886 |
group = 0; |
| 2887 |
} |
| 2888 |
blk_finish_plug(&plug); |
| 2889 |
return group; |
| 2890 |
} |
| 2891 |
|
| 2892 |
/* |
| 2893 |
* Batch reads of the block allocation bitmaps to get |
| 2894 |
* multiple READs in flight; limit prefetching at inexpensive |
| 2895 |
* CR, otherwise mballoc can spend a lot of time loading |
| 2896 |
* imperfect groups |
| 2897 |
*/ |
| 2898 |
static void ext4_mb_might_prefetch(struct ext4_allocation_context *ac, |
| 2899 |
ext4_group_t group) |
| 2900 |
{ |
| 2901 |
struct ext4_sb_info *sbi; |
| 2902 |
|
| 2903 |
if (ac->ac_prefetch_grp != group) |
| 2904 |
return; |
| 2905 |
|
| 2906 |
sbi = EXT4_SB(ac->ac_sb); |
| 2907 |
if (ext4_mb_cr_expensive(ac->ac_criteria) || |
| 2908 |
ac->ac_prefetch_ios < sbi->s_mb_prefetch_limit) { |
| 2909 |
unsigned int nr = sbi->s_mb_prefetch; |
| 2910 |
|
| 2911 |
if (ext4_has_feature_flex_bg(ac->ac_sb)) { |
| 2912 |
nr = 1 << sbi->s_log_groups_per_flex; |
| 2913 |
nr -= group & (nr - 1); |
| 2914 |
nr = umin(nr, sbi->s_mb_prefetch); |
| 2915 |
} |
| 2916 |
|
| 2917 |
ac->ac_prefetch_nr = nr; |
| 2918 |
ac->ac_prefetch_grp = ext4_mb_prefetch(ac->ac_sb, group, nr, |
| 2919 |
&ac->ac_prefetch_ios); |
| 2920 |
} |
| 2921 |
} |
| 2922 |
|
| 2923 |
/* |
| 2924 |
* Prefetching reads the block bitmap into the buffer cache; but we |
| 2925 |
* need to make sure that the buddy bitmap in the page cache has been |
| 2926 |
* initialized. Note that ext4_mb_init_group() will block if the I/O |
| 2927 |
* is not yet completed, or indeed if it was not initiated by |
| 2928 |
* ext4_mb_prefetch did not start the I/O. |
| 2929 |
* |
| 2930 |
* TODO: We should actually kick off the buddy bitmap setup in a work |
| 2931 |
* queue when the buffer I/O is completed, so that we don't block |
| 2932 |
* waiting for the block allocation bitmap read to finish when |
| 2933 |
* ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). |
| 2934 |
*/ |
| 2935 |
void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, |
| 2936 |
unsigned int nr) |
| 2937 |
{ |
| 2938 |
struct ext4_group_desc *gdp; |
| 2939 |
struct ext4_group_info *grp; |
| 2940 |
|
| 2941 |
while (nr-- > 0) { |
| 2942 |
if (!group) |
| 2943 |
group = ext4_get_groups_count(sb); |
| 2944 |
group--; |
| 2945 |
gdp = ext4_get_group_desc(sb, group, NULL); |
| 2946 |
grp = ext4_get_group_info(sb, group); |
| 2947 |
|
| 2948 |
if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) && |
| 2949 |
ext4_free_group_clusters(sb, gdp) > 0) { |
| 2950 |
if (ext4_mb_init_group(sb, group, GFP_NOFS)) |
| 2951 |
break; |
| 2952 |
} |
| 2953 |
} |
| 2954 |
} |
| 2955 |
|
| 2956 |
static int ext4_mb_scan_group(struct ext4_allocation_context *ac, |
| 2957 |
ext4_group_t group) |
| 2958 |
{ |
| 2959 |
int ret; |
| 2960 |
struct super_block *sb = ac->ac_sb; |
| 2961 |
enum criteria cr = ac->ac_criteria; |
| 2962 |
|
| 2963 |
ext4_mb_might_prefetch(ac, group); |
| 2964 |
|
| 2965 |
/* prevent unnecessary buddy loading. */ |
| 2966 |
if (cr < CR_ANY_FREE && spin_is_locked(ext4_group_lock_ptr(sb, group))) |
| 2967 |
return 0; |
| 2968 |
|
| 2969 |
/* This now checks without needing the buddy folio */ |
| 2970 |
ret = ext4_mb_good_group_nolock(ac, group, cr); |
| 2971 |
if (ret <= 0) { |
| 2972 |
if (!ac->ac_first_err) |
| 2973 |
ac->ac_first_err = ret; |
| 2974 |
return 0; |
| 2975 |
} |
| 2976 |
|
| 2977 |
ret = ext4_mb_load_buddy(sb, group, ac->ac_e4b); |
| 2978 |
if (ret) |
| 2979 |
return ret; |
| 2980 |
|
| 2981 |
/* skip busy group */ |
| 2982 |
if (cr >= CR_ANY_FREE) |
| 2983 |
ext4_lock_group(sb, group); |
| 2984 |
else if (!ext4_try_lock_group(sb, group)) |
| 2985 |
goto out_unload; |
| 2986 |
|
| 2987 |
/* We need to check again after locking the block group. */ |
| 2988 |
if (unlikely(!ext4_mb_good_group(ac, group, cr))) |
| 2989 |
goto out_unlock; |
| 2990 |
|
| 2991 |
__ext4_mb_scan_group(ac); |
| 2992 |
|
| 2993 |
out_unlock: |
| 2994 |
ext4_unlock_group(sb, group); |
| 2995 |
out_unload: |
| 2996 |
ext4_mb_unload_buddy(ac->ac_e4b); |
| 2997 |
return ret; |
| 2998 |
} |
| 2999 |
|
| 3000 |
static noinline_for_stack int |
| 3001 |
ext4_mb_regular_allocator(struct ext4_allocation_context *ac) |
| 3002 |
{ |
| 3003 |
ext4_group_t i; |
| 3004 |
int err = 0; |
| 3005 |
struct super_block *sb = ac->ac_sb; |
| 3006 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3007 |
struct ext4_buddy e4b; |
| 3008 |
|
| 3009 |
BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
| 3010 |
|
| 3011 |
/* first, try the goal */ |
| 3012 |
err = ext4_mb_find_by_goal(ac, &e4b); |
| 3013 |
if (err || ac->ac_status == AC_STATUS_FOUND) |
| 3014 |
goto out; |
| 3015 |
|
| 3016 |
if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
| 3017 |
goto out; |
| 3018 |
|
| 3019 |
/* |
| 3020 |
* ac->ac_2order is set only if the fe_len is a power of 2 |
| 3021 |
* if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED |
| 3022 |
* so that we try exact allocation using buddy. |
| 3023 |
*/ |
| 3024 |
i = fls(ac->ac_g_ex.fe_len); |
| 3025 |
ac->ac_2order = 0; |
| 3026 |
/* |
| 3027 |
* We search using buddy data only if the order of the request |
| 3028 |
* is greater than equal to the sbi_s_mb_order2_reqs |
| 3029 |
* You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
| 3030 |
* We also support searching for power-of-two requests only for |
| 3031 |
* requests upto maximum buddy size we have constructed. |
| 3032 |
*/ |
| 3033 |
if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { |
| 3034 |
if (is_power_of_2(ac->ac_g_ex.fe_len)) |
| 3035 |
ac->ac_2order = array_index_nospec(i - 1, |
| 3036 |
MB_NUM_ORDERS(sb)); |
| 3037 |
} |
| 3038 |
|
| 3039 |
/* if stream allocation is enabled, use global goal */ |
| 3040 |
if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
| 3041 |
int hash = ac->ac_inode->i_ino % sbi->s_mb_nr_global_goals; |
| 3042 |
|
| 3043 |
ac->ac_g_ex.fe_group = READ_ONCE(sbi->s_mb_last_groups[hash]); |
| 3044 |
ac->ac_g_ex.fe_start = -1; |
| 3045 |
ac->ac_flags &= ~EXT4_MB_HINT_TRY_GOAL; |
| 3046 |
} |
| 3047 |
|
| 3048 |
/* |
| 3049 |
* Let's just scan groups to find more-less suitable blocks We |
| 3050 |
* start with CR_GOAL_LEN_FAST, unless it is power of 2 |
| 3051 |
* aligned, in which case let's do that faster approach first. |
| 3052 |
*/ |
| 3053 |
ac->ac_criteria = CR_GOAL_LEN_FAST; |
| 3054 |
if (ac->ac_2order) |
| 3055 |
ac->ac_criteria = CR_POWER2_ALIGNED; |
| 3056 |
|
| 3057 |
ac->ac_e4b = &e4b; |
| 3058 |
ac->ac_prefetch_ios = 0; |
| 3059 |
ac->ac_first_err = 0; |
| 3060 |
repeat: |
| 3061 |
while (ac->ac_criteria < EXT4_MB_NUM_CRS) { |
| 3062 |
err = ext4_mb_scan_groups(ac); |
| 3063 |
if (err) |
| 3064 |
goto out; |
| 3065 |
|
| 3066 |
if (ac->ac_status != AC_STATUS_CONTINUE) |
| 3067 |
break; |
| 3068 |
} |
| 3069 |
|
| 3070 |
if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && |
| 3071 |
!(ac->ac_flags & EXT4_MB_HINT_FIRST)) { |
| 3072 |
/* |
| 3073 |
* We've been searching too long. Let's try to allocate |
| 3074 |
* the best chunk we've found so far |
| 3075 |
*/ |
| 3076 |
ext4_mb_try_best_found(ac, &e4b); |
| 3077 |
if (ac->ac_status != AC_STATUS_FOUND) { |
| 3078 |
int lost; |
| 3079 |
|
| 3080 |
/* |
| 3081 |
* Someone more lucky has already allocated it. |
| 3082 |
* The only thing we can do is just take first |
| 3083 |
* found block(s) |
| 3084 |
*/ |
| 3085 |
lost = atomic_inc_return(&sbi->s_mb_lost_chunks); |
| 3086 |
mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", |
| 3087 |
ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, |
| 3088 |
ac->ac_b_ex.fe_len, lost); |
| 3089 |
|
| 3090 |
ac->ac_b_ex.fe_group = 0; |
| 3091 |
ac->ac_b_ex.fe_start = 0; |
| 3092 |
ac->ac_b_ex.fe_len = 0; |
| 3093 |
ac->ac_status = AC_STATUS_CONTINUE; |
| 3094 |
ac->ac_flags |= EXT4_MB_HINT_FIRST; |
| 3095 |
ac->ac_criteria = CR_ANY_FREE; |
| 3096 |
goto repeat; |
| 3097 |
} |
| 3098 |
} |
| 3099 |
|
| 3100 |
if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) { |
| 3101 |
atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); |
| 3102 |
if (ac->ac_flags & EXT4_MB_STREAM_ALLOC && |
| 3103 |
ac->ac_b_ex.fe_group == ac->ac_g_ex.fe_group) |
| 3104 |
atomic_inc(&sbi->s_bal_stream_goals); |
| 3105 |
} |
| 3106 |
out: |
| 3107 |
if (!err && ac->ac_status != AC_STATUS_FOUND && ac->ac_first_err) |
| 3108 |
err = ac->ac_first_err; |
| 3109 |
|
| 3110 |
mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", |
| 3111 |
ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, |
| 3112 |
ac->ac_flags, ac->ac_criteria, err); |
| 3113 |
|
| 3114 |
if (ac->ac_prefetch_nr) |
| 3115 |
ext4_mb_prefetch_fini(sb, ac->ac_prefetch_grp, ac->ac_prefetch_nr); |
| 3116 |
|
| 3117 |
return err; |
| 3118 |
} |
| 3119 |
|
| 3120 |
static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
| 3121 |
{ |
| 3122 |
struct super_block *sb = pde_data(file_inode(seq->file)); |
| 3123 |
ext4_group_t group; |
| 3124 |
|
| 3125 |
if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
| 3126 |
return NULL; |
| 3127 |
group = *pos + 1; |
| 3128 |
return (void *) ((unsigned long) group); |
| 3129 |
} |
| 3130 |
|
| 3131 |
static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) |
| 3132 |
{ |
| 3133 |
struct super_block *sb = pde_data(file_inode(seq->file)); |
| 3134 |
ext4_group_t group; |
| 3135 |
|
| 3136 |
++*pos; |
| 3137 |
if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
| 3138 |
return NULL; |
| 3139 |
group = *pos + 1; |
| 3140 |
return (void *) ((unsigned long) group); |
| 3141 |
} |
| 3142 |
|
| 3143 |
static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) |
| 3144 |
{ |
| 3145 |
struct super_block *sb = pde_data(file_inode(seq->file)); |
| 3146 |
ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
| 3147 |
int i, err; |
| 3148 |
char nbuf[16]; |
| 3149 |
struct ext4_buddy e4b; |
| 3150 |
struct ext4_group_info *grinfo; |
| 3151 |
unsigned char blocksize_bits = min_t(unsigned char, |
| 3152 |
sb->s_blocksize_bits, |
| 3153 |
EXT4_MAX_BLOCK_LOG_SIZE); |
| 3154 |
DEFINE_RAW_FLEX(struct ext4_group_info, sg, bb_counters, |
| 3155 |
EXT4_MAX_BLOCK_LOG_SIZE + 2); |
| 3156 |
|
| 3157 |
group--; |
| 3158 |
if (group == 0) |
| 3159 |
seq_puts(seq, "#group: free frags first [" |
| 3160 |
" 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " |
| 3161 |
" 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); |
| 3162 |
|
| 3163 |
i = (blocksize_bits + 2) * sizeof(sg->bb_counters[0]) + |
| 3164 |
sizeof(struct ext4_group_info); |
| 3165 |
|
| 3166 |
grinfo = ext4_get_group_info(sb, group); |
| 3167 |
if (!grinfo) |
| 3168 |
return 0; |
| 3169 |
/* Load the group info in memory only if not already loaded. */ |
| 3170 |
if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { |
| 3171 |
err = ext4_mb_load_buddy(sb, group, &e4b); |
| 3172 |
if (err) { |
| 3173 |
seq_printf(seq, "#%-5u: %s\n", group, ext4_decode_error(NULL, err, nbuf)); |
| 3174 |
return 0; |
| 3175 |
} |
| 3176 |
ext4_mb_unload_buddy(&e4b); |
| 3177 |
} |
| 3178 |
|
| 3179 |
/* |
| 3180 |
* We care only about free space counters in the group info and |
| 3181 |
* these are safe to access even after the buddy has been unloaded |
| 3182 |
*/ |
| 3183 |
memcpy(sg, grinfo, i); |
| 3184 |
seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg->bb_free, |
| 3185 |
sg->bb_fragments, sg->bb_first_free); |
| 3186 |
for (i = 0; i <= 13; i++) |
| 3187 |
seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? |
| 3188 |
sg->bb_counters[i] : 0); |
| 3189 |
seq_puts(seq, " ]"); |
| 3190 |
if (EXT4_MB_GRP_BBITMAP_CORRUPT(sg)) |
| 3191 |
seq_puts(seq, " Block bitmap corrupted!"); |
| 3192 |
seq_putc(seq, '\n'); |
| 3193 |
return 0; |
| 3194 |
} |
| 3195 |
|
| 3196 |
static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) |
| 3197 |
{ |
| 3198 |
} |
| 3199 |
|
| 3200 |
const struct seq_operations ext4_mb_seq_groups_ops = { |
| 3201 |
.start = ext4_mb_seq_groups_start, |
| 3202 |
.next = ext4_mb_seq_groups_next, |
| 3203 |
.stop = ext4_mb_seq_groups_stop, |
| 3204 |
.show = ext4_mb_seq_groups_show, |
| 3205 |
}; |
| 3206 |
|
| 3207 |
int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) |
| 3208 |
{ |
| 3209 |
struct super_block *sb = seq->private; |
| 3210 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3211 |
|
| 3212 |
seq_puts(seq, "mballoc:\n"); |
| 3213 |
if (!sbi->s_mb_stats) { |
| 3214 |
seq_puts(seq, "\tmb stats collection turned off.\n"); |
| 3215 |
seq_puts( |
| 3216 |
seq, |
| 3217 |
"\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); |
| 3218 |
return 0; |
| 3219 |
} |
| 3220 |
seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); |
| 3221 |
seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); |
| 3222 |
|
| 3223 |
seq_printf(seq, "\tgroups_scanned: %u\n", |
| 3224 |
atomic_read(&sbi->s_bal_groups_scanned)); |
| 3225 |
|
| 3226 |
/* CR_POWER2_ALIGNED stats */ |
| 3227 |
seq_puts(seq, "\tcr_p2_aligned_stats:\n"); |
| 3228 |
seq_printf(seq, "\t\thits: %llu\n", |
| 3229 |
atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED])); |
| 3230 |
seq_printf( |
| 3231 |
seq, "\t\tgroups_considered: %llu\n", |
| 3232 |
atomic64_read( |
| 3233 |
&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED])); |
| 3234 |
seq_printf(seq, "\t\textents_scanned: %u\n", |
| 3235 |
atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED])); |
| 3236 |
seq_printf(seq, "\t\tuseless_loops: %llu\n", |
| 3237 |
atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED])); |
| 3238 |
|
| 3239 |
/* CR_GOAL_LEN_FAST stats */ |
| 3240 |
seq_puts(seq, "\tcr_goal_fast_stats:\n"); |
| 3241 |
seq_printf(seq, "\t\thits: %llu\n", |
| 3242 |
atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST])); |
| 3243 |
seq_printf(seq, "\t\tgroups_considered: %llu\n", |
| 3244 |
atomic64_read( |
| 3245 |
&sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST])); |
| 3246 |
seq_printf(seq, "\t\textents_scanned: %u\n", |
| 3247 |
atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST])); |
| 3248 |
seq_printf(seq, "\t\tuseless_loops: %llu\n", |
| 3249 |
atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST])); |
| 3250 |
|
| 3251 |
/* CR_BEST_AVAIL_LEN stats */ |
| 3252 |
seq_puts(seq, "\tcr_best_avail_stats:\n"); |
| 3253 |
seq_printf(seq, "\t\thits: %llu\n", |
| 3254 |
atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN])); |
| 3255 |
seq_printf( |
| 3256 |
seq, "\t\tgroups_considered: %llu\n", |
| 3257 |
atomic64_read( |
| 3258 |
&sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN])); |
| 3259 |
seq_printf(seq, "\t\textents_scanned: %u\n", |
| 3260 |
atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN])); |
| 3261 |
seq_printf(seq, "\t\tuseless_loops: %llu\n", |
| 3262 |
atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN])); |
| 3263 |
|
| 3264 |
/* CR_GOAL_LEN_SLOW stats */ |
| 3265 |
seq_puts(seq, "\tcr_goal_slow_stats:\n"); |
| 3266 |
seq_printf(seq, "\t\thits: %llu\n", |
| 3267 |
atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW])); |
| 3268 |
seq_printf(seq, "\t\tgroups_considered: %llu\n", |
| 3269 |
atomic64_read( |
| 3270 |
&sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW])); |
| 3271 |
seq_printf(seq, "\t\textents_scanned: %u\n", |
| 3272 |
atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW])); |
| 3273 |
seq_printf(seq, "\t\tuseless_loops: %llu\n", |
| 3274 |
atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW])); |
| 3275 |
|
| 3276 |
/* CR_ANY_FREE stats */ |
| 3277 |
seq_puts(seq, "\tcr_any_free_stats:\n"); |
| 3278 |
seq_printf(seq, "\t\thits: %llu\n", |
| 3279 |
atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE])); |
| 3280 |
seq_printf( |
| 3281 |
seq, "\t\tgroups_considered: %llu\n", |
| 3282 |
atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE])); |
| 3283 |
seq_printf(seq, "\t\textents_scanned: %u\n", |
| 3284 |
atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE])); |
| 3285 |
seq_printf(seq, "\t\tuseless_loops: %llu\n", |
| 3286 |
atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE])); |
| 3287 |
|
| 3288 |
/* Aggregates */ |
| 3289 |
seq_printf(seq, "\textents_scanned: %u\n", |
| 3290 |
atomic_read(&sbi->s_bal_ex_scanned)); |
| 3291 |
seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); |
| 3292 |
seq_printf(seq, "\t\tstream_goal_hits: %u\n", |
| 3293 |
atomic_read(&sbi->s_bal_stream_goals)); |
| 3294 |
seq_printf(seq, "\t\tlen_goal_hits: %u\n", |
| 3295 |
atomic_read(&sbi->s_bal_len_goals)); |
| 3296 |
seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); |
| 3297 |
seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); |
| 3298 |
seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); |
| 3299 |
seq_printf(seq, "\tbuddies_generated: %u/%u\n", |
| 3300 |
atomic_read(&sbi->s_mb_buddies_generated), |
| 3301 |
ext4_get_groups_count(sb)); |
| 3302 |
seq_printf(seq, "\tbuddies_time_used: %llu\n", |
| 3303 |
atomic64_read(&sbi->s_mb_generation_time)); |
| 3304 |
seq_printf(seq, "\tpreallocated: %u\n", |
| 3305 |
atomic_read(&sbi->s_mb_preallocated)); |
| 3306 |
seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded)); |
| 3307 |
return 0; |
| 3308 |
} |
| 3309 |
|
| 3310 |
static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) |
| 3311 |
{ |
| 3312 |
struct super_block *sb = pde_data(file_inode(seq->file)); |
| 3313 |
unsigned long position; |
| 3314 |
|
| 3315 |
if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
| 3316 |
return NULL; |
| 3317 |
position = *pos + 1; |
| 3318 |
return (void *) ((unsigned long) position); |
| 3319 |
} |
| 3320 |
|
| 3321 |
static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) |
| 3322 |
{ |
| 3323 |
struct super_block *sb = pde_data(file_inode(seq->file)); |
| 3324 |
unsigned long position; |
| 3325 |
|
| 3326 |
++*pos; |
| 3327 |
if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) |
| 3328 |
return NULL; |
| 3329 |
position = *pos + 1; |
| 3330 |
return (void *) ((unsigned long) position); |
| 3331 |
} |
| 3332 |
|
| 3333 |
static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) |
| 3334 |
{ |
| 3335 |
struct super_block *sb = pde_data(file_inode(seq->file)); |
| 3336 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3337 |
unsigned long position = ((unsigned long) v); |
| 3338 |
struct ext4_group_info *grp; |
| 3339 |
unsigned int count; |
| 3340 |
unsigned long idx; |
| 3341 |
|
| 3342 |
position--; |
| 3343 |
if (position >= MB_NUM_ORDERS(sb)) { |
| 3344 |
position -= MB_NUM_ORDERS(sb); |
| 3345 |
if (position == 0) |
| 3346 |
seq_puts(seq, "avg_fragment_size_lists:\n"); |
| 3347 |
|
| 3348 |
count = 0; |
| 3349 |
xa_for_each(&sbi->s_mb_avg_fragment_size[position], idx, grp) |
| 3350 |
count++; |
| 3351 |
seq_printf(seq, "\tlist_order_%u_groups: %u\n", |
| 3352 |
(unsigned int)position, count); |
| 3353 |
return 0; |
| 3354 |
} |
| 3355 |
|
| 3356 |
if (position == 0) { |
| 3357 |
seq_printf(seq, "optimize_scan: %d\n", |
| 3358 |
test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); |
| 3359 |
seq_puts(seq, "max_free_order_lists:\n"); |
| 3360 |
} |
| 3361 |
count = 0; |
| 3362 |
xa_for_each(&sbi->s_mb_largest_free_orders[position], idx, grp) |
| 3363 |
count++; |
| 3364 |
seq_printf(seq, "\tlist_order_%u_groups: %u\n", |
| 3365 |
(unsigned int)position, count); |
| 3366 |
|
| 3367 |
return 0; |
| 3368 |
} |
| 3369 |
|
| 3370 |
static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) |
| 3371 |
{ |
| 3372 |
} |
| 3373 |
|
| 3374 |
const struct seq_operations ext4_mb_seq_structs_summary_ops = { |
| 3375 |
.start = ext4_mb_seq_structs_summary_start, |
| 3376 |
.next = ext4_mb_seq_structs_summary_next, |
| 3377 |
.stop = ext4_mb_seq_structs_summary_stop, |
| 3378 |
.show = ext4_mb_seq_structs_summary_show, |
| 3379 |
}; |
| 3380 |
|
| 3381 |
static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
| 3382 |
{ |
| 3383 |
int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; |
| 3384 |
struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; |
| 3385 |
|
| 3386 |
BUG_ON(!cachep); |
| 3387 |
return cachep; |
| 3388 |
} |
| 3389 |
|
| 3390 |
/* |
| 3391 |
* Allocate the top-level s_group_info array for the specified number |
| 3392 |
* of groups |
| 3393 |
*/ |
| 3394 |
int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) |
| 3395 |
{ |
| 3396 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3397 |
unsigned size; |
| 3398 |
struct ext4_group_info ***old_groupinfo, ***new_groupinfo; |
| 3399 |
|
| 3400 |
size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> |
| 3401 |
EXT4_DESC_PER_BLOCK_BITS(sb); |
| 3402 |
if (size <= sbi->s_group_info_size) |
| 3403 |
return 0; |
| 3404 |
|
| 3405 |
size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); |
| 3406 |
new_groupinfo = kvzalloc(size, GFP_KERNEL); |
| 3407 |
if (!new_groupinfo) { |
| 3408 |
ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); |
| 3409 |
return -ENOMEM; |
| 3410 |
} |
| 3411 |
rcu_read_lock(); |
| 3412 |
old_groupinfo = rcu_dereference(sbi->s_group_info); |
| 3413 |
if (old_groupinfo) |
| 3414 |
memcpy(new_groupinfo, old_groupinfo, |
| 3415 |
sbi->s_group_info_size * sizeof(*sbi->s_group_info)); |
| 3416 |
rcu_read_unlock(); |
| 3417 |
rcu_assign_pointer(sbi->s_group_info, new_groupinfo); |
| 3418 |
sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); |
| 3419 |
if (old_groupinfo) |
| 3420 |
ext4_kvfree_array_rcu(old_groupinfo); |
| 3421 |
ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", |
| 3422 |
sbi->s_group_info_size); |
| 3423 |
return 0; |
| 3424 |
} |
| 3425 |
|
| 3426 |
/* Create and initialize ext4_group_info data for the given group. */ |
| 3427 |
int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
| 3428 |
struct ext4_group_desc *desc) |
| 3429 |
{ |
| 3430 |
int i; |
| 3431 |
int metalen = 0; |
| 3432 |
int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); |
| 3433 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3434 |
struct ext4_group_info **meta_group_info; |
| 3435 |
struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
| 3436 |
|
| 3437 |
/* |
| 3438 |
* First check if this group is the first of a reserved block. |
| 3439 |
* If it's true, we have to allocate a new table of pointers |
| 3440 |
* to ext4_group_info structures |
| 3441 |
*/ |
| 3442 |
if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
| 3443 |
metalen = sizeof(*meta_group_info) << |
| 3444 |
EXT4_DESC_PER_BLOCK_BITS(sb); |
| 3445 |
meta_group_info = kmalloc(metalen, GFP_NOFS); |
| 3446 |
if (meta_group_info == NULL) { |
| 3447 |
ext4_msg(sb, KERN_ERR, "can't allocate mem " |
| 3448 |
"for a buddy group"); |
| 3449 |
return -ENOMEM; |
| 3450 |
} |
| 3451 |
rcu_read_lock(); |
| 3452 |
rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; |
| 3453 |
rcu_read_unlock(); |
| 3454 |
} |
| 3455 |
|
| 3456 |
meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); |
| 3457 |
i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); |
| 3458 |
|
| 3459 |
meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); |
| 3460 |
if (meta_group_info[i] == NULL) { |
| 3461 |
ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
| 3462 |
goto exit_group_info; |
| 3463 |
} |
| 3464 |
set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, |
| 3465 |
&(meta_group_info[i]->bb_state)); |
| 3466 |
|
| 3467 |
/* |
| 3468 |
* initialize bb_free to be able to skip |
| 3469 |
* empty groups without initialization |
| 3470 |
*/ |
| 3471 |
if (ext4_has_group_desc_csum(sb) && |
| 3472 |
(desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { |
| 3473 |
meta_group_info[i]->bb_free = |
| 3474 |
ext4_free_clusters_after_init(sb, group, desc); |
| 3475 |
} else { |
| 3476 |
meta_group_info[i]->bb_free = |
| 3477 |
ext4_free_group_clusters(sb, desc); |
| 3478 |
} |
| 3479 |
|
| 3480 |
INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); |
| 3481 |
init_rwsem(&meta_group_info[i]->alloc_sem); |
| 3482 |
meta_group_info[i]->bb_free_root = RB_ROOT; |
| 3483 |
meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
| 3484 |
meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ |
| 3485 |
meta_group_info[i]->bb_group = group; |
| 3486 |
|
| 3487 |
mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); |
| 3488 |
return 0; |
| 3489 |
|
| 3490 |
exit_group_info: |
| 3491 |
/* If a meta_group_info table has been allocated, release it now */ |
| 3492 |
if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
| 3493 |
struct ext4_group_info ***group_info; |
| 3494 |
|
| 3495 |
rcu_read_lock(); |
| 3496 |
group_info = rcu_dereference(sbi->s_group_info); |
| 3497 |
kfree(group_info[idx]); |
| 3498 |
group_info[idx] = NULL; |
| 3499 |
rcu_read_unlock(); |
| 3500 |
} |
| 3501 |
return -ENOMEM; |
| 3502 |
} /* ext4_mb_add_groupinfo */ |
| 3503 |
|
| 3504 |
static int ext4_mb_init_backend(struct super_block *sb) |
| 3505 |
{ |
| 3506 |
ext4_group_t ngroups = ext4_get_groups_count(sb); |
| 3507 |
ext4_group_t i; |
| 3508 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3509 |
int err; |
| 3510 |
struct ext4_group_desc *desc; |
| 3511 |
struct ext4_group_info ***group_info; |
| 3512 |
struct kmem_cache *cachep; |
| 3513 |
|
| 3514 |
err = ext4_mb_alloc_groupinfo(sb, ngroups); |
| 3515 |
if (err) |
| 3516 |
return err; |
| 3517 |
|
| 3518 |
sbi->s_buddy_cache = new_inode(sb); |
| 3519 |
if (sbi->s_buddy_cache == NULL) { |
| 3520 |
ext4_msg(sb, KERN_ERR, "can't get new inode"); |
| 3521 |
goto err_freesgi; |
| 3522 |
} |
| 3523 |
/* To avoid potentially colliding with an valid on-disk inode number, |
| 3524 |
* use EXT4_BAD_INO for the buddy cache inode number. This inode is |
| 3525 |
* not in the inode hash, so it should never be found by iget(), but |
| 3526 |
* this will avoid confusion if it ever shows up during debugging. */ |
| 3527 |
sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; |
| 3528 |
EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
| 3529 |
ext4_set_inode_mapping_order(sbi->s_buddy_cache); |
| 3530 |
|
| 3531 |
for (i = 0; i < ngroups; i++) { |
| 3532 |
cond_resched(); |
| 3533 |
desc = ext4_get_group_desc(sb, i, NULL); |
| 3534 |
if (desc == NULL) { |
| 3535 |
ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
| 3536 |
goto err_freebuddy; |
| 3537 |
} |
| 3538 |
if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
| 3539 |
goto err_freebuddy; |
| 3540 |
} |
| 3541 |
|
| 3542 |
if (ext4_has_feature_flex_bg(sb)) { |
| 3543 |
/* a single flex group is supposed to be read by a single IO. |
| 3544 |
* 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is |
| 3545 |
* unsigned integer, so the maximum shift is 32. |
| 3546 |
*/ |
| 3547 |
if (sbi->s_es->s_log_groups_per_flex >= 32) { |
| 3548 |
ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); |
| 3549 |
goto err_freebuddy; |
| 3550 |
} |
| 3551 |
sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, |
| 3552 |
BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); |
| 3553 |
sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ |
| 3554 |
} else { |
| 3555 |
sbi->s_mb_prefetch = 32; |
| 3556 |
} |
| 3557 |
if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) |
| 3558 |
sbi->s_mb_prefetch = ext4_get_groups_count(sb); |
| 3559 |
/* |
| 3560 |
* now many real IOs to prefetch within a single allocation at |
| 3561 |
* CR_POWER2_ALIGNED. Given CR_POWER2_ALIGNED is an CPU-related |
| 3562 |
* optimization we shouldn't try to load too many groups, at some point |
| 3563 |
* we should start to use what we've got in memory. |
| 3564 |
* with an average random access time 5ms, it'd take a second to get |
| 3565 |
* 200 groups (* N with flex_bg), so let's make this limit 4 |
| 3566 |
*/ |
| 3567 |
sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; |
| 3568 |
if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) |
| 3569 |
sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); |
| 3570 |
|
| 3571 |
return 0; |
| 3572 |
|
| 3573 |
err_freebuddy: |
| 3574 |
cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
| 3575 |
while (i-- > 0) { |
| 3576 |
struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
| 3577 |
|
| 3578 |
if (grp) |
| 3579 |
kmem_cache_free(cachep, grp); |
| 3580 |
} |
| 3581 |
i = sbi->s_group_info_size; |
| 3582 |
rcu_read_lock(); |
| 3583 |
group_info = rcu_dereference(sbi->s_group_info); |
| 3584 |
while (i-- > 0) |
| 3585 |
kfree(group_info[i]); |
| 3586 |
rcu_read_unlock(); |
| 3587 |
iput(sbi->s_buddy_cache); |
| 3588 |
err_freesgi: |
| 3589 |
kvfree(rcu_access_pointer(sbi->s_group_info)); |
| 3590 |
return -ENOMEM; |
| 3591 |
} |
| 3592 |
|
| 3593 |
static void ext4_groupinfo_destroy_slabs(void) |
| 3594 |
{ |
| 3595 |
int i; |
| 3596 |
|
| 3597 |
for (i = 0; i < NR_GRPINFO_CACHES; i++) { |
| 3598 |
kmem_cache_destroy(ext4_groupinfo_caches[i]); |
| 3599 |
ext4_groupinfo_caches[i] = NULL; |
| 3600 |
} |
| 3601 |
} |
| 3602 |
|
| 3603 |
static int ext4_groupinfo_create_slab(size_t size) |
| 3604 |
{ |
| 3605 |
static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); |
| 3606 |
int slab_size; |
| 3607 |
int blocksize_bits = order_base_2(size); |
| 3608 |
int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; |
| 3609 |
struct kmem_cache *cachep; |
| 3610 |
|
| 3611 |
if (cache_index >= NR_GRPINFO_CACHES) |
| 3612 |
return -EINVAL; |
| 3613 |
|
| 3614 |
if (unlikely(cache_index < 0)) |
| 3615 |
cache_index = 0; |
| 3616 |
|
| 3617 |
mutex_lock(&ext4_grpinfo_slab_create_mutex); |
| 3618 |
if (ext4_groupinfo_caches[cache_index]) { |
| 3619 |
mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
| 3620 |
return 0; /* Already created */ |
| 3621 |
} |
| 3622 |
|
| 3623 |
slab_size = offsetof(struct ext4_group_info, |
| 3624 |
bb_counters[blocksize_bits + 2]); |
| 3625 |
|
| 3626 |
cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], |
| 3627 |
slab_size, 0, SLAB_RECLAIM_ACCOUNT, |
| 3628 |
NULL); |
| 3629 |
|
| 3630 |
ext4_groupinfo_caches[cache_index] = cachep; |
| 3631 |
|
| 3632 |
mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
| 3633 |
if (!cachep) { |
| 3634 |
printk(KERN_EMERG |
| 3635 |
"EXT4-fs: no memory for groupinfo slab cache\n"); |
| 3636 |
return -ENOMEM; |
| 3637 |
} |
| 3638 |
|
| 3639 |
return 0; |
| 3640 |
} |
| 3641 |
|
| 3642 |
static void ext4_discard_work(struct work_struct *work) |
| 3643 |
{ |
| 3644 |
struct ext4_sb_info *sbi = container_of(work, |
| 3645 |
struct ext4_sb_info, s_discard_work); |
| 3646 |
struct super_block *sb = sbi->s_sb; |
| 3647 |
struct ext4_free_data *fd, *nfd; |
| 3648 |
struct ext4_buddy e4b; |
| 3649 |
LIST_HEAD(discard_list); |
| 3650 |
ext4_group_t grp, load_grp; |
| 3651 |
int err = 0; |
| 3652 |
|
| 3653 |
spin_lock(&sbi->s_md_lock); |
| 3654 |
list_splice_init(&sbi->s_discard_list, &discard_list); |
| 3655 |
spin_unlock(&sbi->s_md_lock); |
| 3656 |
|
| 3657 |
load_grp = UINT_MAX; |
| 3658 |
list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { |
| 3659 |
/* |
| 3660 |
* If filesystem is umounting or no memory or suffering |
| 3661 |
* from no space, give up the discard |
| 3662 |
*/ |
| 3663 |
if ((sb->s_flags & SB_ACTIVE) && !err && |
| 3664 |
!atomic_read(&sbi->s_retry_alloc_pending)) { |
| 3665 |
grp = fd->efd_group; |
| 3666 |
if (grp != load_grp) { |
| 3667 |
if (load_grp != UINT_MAX) |
| 3668 |
ext4_mb_unload_buddy(&e4b); |
| 3669 |
|
| 3670 |
err = ext4_mb_load_buddy(sb, grp, &e4b); |
| 3671 |
if (err) { |
| 3672 |
kmem_cache_free(ext4_free_data_cachep, fd); |
| 3673 |
load_grp = UINT_MAX; |
| 3674 |
continue; |
| 3675 |
} else { |
| 3676 |
load_grp = grp; |
| 3677 |
} |
| 3678 |
} |
| 3679 |
|
| 3680 |
ext4_lock_group(sb, grp); |
| 3681 |
ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, |
| 3682 |
fd->efd_start_cluster + fd->efd_count - 1, 1); |
| 3683 |
ext4_unlock_group(sb, grp); |
| 3684 |
} |
| 3685 |
kmem_cache_free(ext4_free_data_cachep, fd); |
| 3686 |
} |
| 3687 |
|
| 3688 |
if (load_grp != UINT_MAX) |
| 3689 |
ext4_mb_unload_buddy(&e4b); |
| 3690 |
} |
| 3691 |
|
| 3692 |
static inline void ext4_mb_avg_fragment_size_destroy(struct ext4_sb_info *sbi) |
| 3693 |
{ |
| 3694 |
if (!sbi->s_mb_avg_fragment_size) |
| 3695 |
return; |
| 3696 |
|
| 3697 |
for (int i = 0; i < MB_NUM_ORDERS(sbi->s_sb); i++) |
| 3698 |
xa_destroy(&sbi->s_mb_avg_fragment_size[i]); |
| 3699 |
|
| 3700 |
kfree(sbi->s_mb_avg_fragment_size); |
| 3701 |
sbi->s_mb_avg_fragment_size = NULL; |
| 3702 |
} |
| 3703 |
|
| 3704 |
static inline void ext4_mb_largest_free_orders_destroy(struct ext4_sb_info *sbi) |
| 3705 |
{ |
| 3706 |
if (!sbi->s_mb_largest_free_orders) |
| 3707 |
return; |
| 3708 |
|
| 3709 |
for (int i = 0; i < MB_NUM_ORDERS(sbi->s_sb); i++) |
| 3710 |
xa_destroy(&sbi->s_mb_largest_free_orders[i]); |
| 3711 |
|
| 3712 |
kfree(sbi->s_mb_largest_free_orders); |
| 3713 |
sbi->s_mb_largest_free_orders = NULL; |
| 3714 |
} |
| 3715 |
|
| 3716 |
int ext4_mb_init(struct super_block *sb) |
| 3717 |
{ |
| 3718 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3719 |
unsigned i, j; |
| 3720 |
unsigned offset, offset_incr; |
| 3721 |
unsigned max; |
| 3722 |
int ret; |
| 3723 |
|
| 3724 |
i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); |
| 3725 |
|
| 3726 |
sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); |
| 3727 |
if (sbi->s_mb_offsets == NULL) { |
| 3728 |
ret = -ENOMEM; |
| 3729 |
goto out; |
| 3730 |
} |
| 3731 |
|
| 3732 |
i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); |
| 3733 |
sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
| 3734 |
if (sbi->s_mb_maxs == NULL) { |
| 3735 |
ret = -ENOMEM; |
| 3736 |
goto out; |
| 3737 |
} |
| 3738 |
|
| 3739 |
ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
| 3740 |
if (ret < 0) |
| 3741 |
goto out; |
| 3742 |
|
| 3743 |
/* order 0 is regular bitmap */ |
| 3744 |
sbi->s_mb_maxs[0] = sb->s_blocksize << 3; |
| 3745 |
sbi->s_mb_offsets[0] = 0; |
| 3746 |
|
| 3747 |
i = 1; |
| 3748 |
offset = 0; |
| 3749 |
offset_incr = 1 << (sb->s_blocksize_bits - 1); |
| 3750 |
max = sb->s_blocksize << 2; |
| 3751 |
do { |
| 3752 |
sbi->s_mb_offsets[i] = offset; |
| 3753 |
sbi->s_mb_maxs[i] = max; |
| 3754 |
offset += offset_incr; |
| 3755 |
offset_incr = offset_incr >> 1; |
| 3756 |
max = max >> 1; |
| 3757 |
i++; |
| 3758 |
} while (i < MB_NUM_ORDERS(sb)); |
| 3759 |
|
| 3760 |
sbi->s_mb_avg_fragment_size = |
| 3761 |
kmalloc_objs(struct xarray, MB_NUM_ORDERS(sb)); |
| 3762 |
if (!sbi->s_mb_avg_fragment_size) { |
| 3763 |
ret = -ENOMEM; |
| 3764 |
goto out; |
| 3765 |
} |
| 3766 |
for (i = 0; i < MB_NUM_ORDERS(sb); i++) |
| 3767 |
xa_init(&sbi->s_mb_avg_fragment_size[i]); |
| 3768 |
|
| 3769 |
sbi->s_mb_largest_free_orders = |
| 3770 |
kmalloc_objs(struct xarray, MB_NUM_ORDERS(sb)); |
| 3771 |
if (!sbi->s_mb_largest_free_orders) { |
| 3772 |
ret = -ENOMEM; |
| 3773 |
goto out; |
| 3774 |
} |
| 3775 |
for (i = 0; i < MB_NUM_ORDERS(sb); i++) |
| 3776 |
xa_init(&sbi->s_mb_largest_free_orders[i]); |
| 3777 |
|
| 3778 |
spin_lock_init(&sbi->s_md_lock); |
| 3779 |
atomic_set(&sbi->s_mb_free_pending, 0); |
| 3780 |
INIT_LIST_HEAD(&sbi->s_freed_data_list[0]); |
| 3781 |
INIT_LIST_HEAD(&sbi->s_freed_data_list[1]); |
| 3782 |
INIT_LIST_HEAD(&sbi->s_discard_list); |
| 3783 |
INIT_WORK(&sbi->s_discard_work, ext4_discard_work); |
| 3784 |
atomic_set(&sbi->s_retry_alloc_pending, 0); |
| 3785 |
|
| 3786 |
sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; |
| 3787 |
sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; |
| 3788 |
sbi->s_mb_stats = MB_DEFAULT_STATS; |
| 3789 |
sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; |
| 3790 |
sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; |
| 3791 |
sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER; |
| 3792 |
|
| 3793 |
/* |
| 3794 |
* The default group preallocation is 512, which for 4k block |
| 3795 |
* sizes translates to 2 megabytes. However for bigalloc file |
| 3796 |
* systems, this is probably too big (i.e, if the cluster size |
| 3797 |
* is 1 megabyte, then group preallocation size becomes half a |
| 3798 |
* gigabyte!). As a default, we will keep a two megabyte |
| 3799 |
* group pralloc size for cluster sizes up to 64k, and after |
| 3800 |
* that, we will force a minimum group preallocation size of |
| 3801 |
* 32 clusters. This translates to 8 megs when the cluster |
| 3802 |
* size is 256k, and 32 megs when the cluster size is 1 meg, |
| 3803 |
* which seems reasonable as a default. |
| 3804 |
*/ |
| 3805 |
sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> |
| 3806 |
sbi->s_cluster_bits, 32); |
| 3807 |
/* |
| 3808 |
* If there is a s_stripe > 1, then we set the s_mb_group_prealloc |
| 3809 |
* to the lowest multiple of s_stripe which is bigger than |
| 3810 |
* the s_mb_group_prealloc as determined above. We want |
| 3811 |
* the preallocation size to be an exact multiple of the |
| 3812 |
* RAID stripe size so that preallocations don't fragment |
| 3813 |
* the stripes. |
| 3814 |
*/ |
| 3815 |
if (sbi->s_stripe > 1) { |
| 3816 |
sbi->s_mb_group_prealloc = roundup( |
| 3817 |
sbi->s_mb_group_prealloc, EXT4_NUM_B2C(sbi, sbi->s_stripe)); |
| 3818 |
} |
| 3819 |
|
| 3820 |
sbi->s_mb_nr_global_goals = umin(num_possible_cpus(), |
| 3821 |
DIV_ROUND_UP(sbi->s_groups_count, 4)); |
| 3822 |
sbi->s_mb_last_groups = kzalloc_objs(ext4_group_t, |
| 3823 |
sbi->s_mb_nr_global_goals); |
| 3824 |
if (sbi->s_mb_last_groups == NULL) { |
| 3825 |
ret = -ENOMEM; |
| 3826 |
goto out; |
| 3827 |
} |
| 3828 |
|
| 3829 |
sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
| 3830 |
if (sbi->s_locality_groups == NULL) { |
| 3831 |
ret = -ENOMEM; |
| 3832 |
goto out_free_last_groups; |
| 3833 |
} |
| 3834 |
for_each_possible_cpu(i) { |
| 3835 |
struct ext4_locality_group *lg; |
| 3836 |
lg = per_cpu_ptr(sbi->s_locality_groups, i); |
| 3837 |
mutex_init(&lg->lg_mutex); |
| 3838 |
for (j = 0; j < PREALLOC_TB_SIZE; j++) |
| 3839 |
INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); |
| 3840 |
spin_lock_init(&lg->lg_prealloc_lock); |
| 3841 |
} |
| 3842 |
|
| 3843 |
if (bdev_nonrot(sb->s_bdev)) |
| 3844 |
sbi->s_mb_max_linear_groups = 0; |
| 3845 |
else |
| 3846 |
sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; |
| 3847 |
/* init file for buddy data */ |
| 3848 |
ret = ext4_mb_init_backend(sb); |
| 3849 |
if (ret != 0) |
| 3850 |
goto out_free_locality_groups; |
| 3851 |
|
| 3852 |
return 0; |
| 3853 |
|
| 3854 |
out_free_locality_groups: |
| 3855 |
free_percpu(sbi->s_locality_groups); |
| 3856 |
sbi->s_locality_groups = NULL; |
| 3857 |
out_free_last_groups: |
| 3858 |
kfree(sbi->s_mb_last_groups); |
| 3859 |
sbi->s_mb_last_groups = NULL; |
| 3860 |
out: |
| 3861 |
ext4_mb_avg_fragment_size_destroy(sbi); |
| 3862 |
ext4_mb_largest_free_orders_destroy(sbi); |
| 3863 |
kfree(sbi->s_mb_offsets); |
| 3864 |
sbi->s_mb_offsets = NULL; |
| 3865 |
kfree(sbi->s_mb_maxs); |
| 3866 |
sbi->s_mb_maxs = NULL; |
| 3867 |
return ret; |
| 3868 |
} |
| 3869 |
|
| 3870 |
/* need to called with the ext4 group lock held */ |
| 3871 |
static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
| 3872 |
{ |
| 3873 |
struct ext4_prealloc_space *pa; |
| 3874 |
struct list_head *cur, *tmp; |
| 3875 |
int count = 0; |
| 3876 |
|
| 3877 |
list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { |
| 3878 |
pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
| 3879 |
list_del(&pa->pa_group_list); |
| 3880 |
count++; |
| 3881 |
kmem_cache_free(ext4_pspace_cachep, pa); |
| 3882 |
} |
| 3883 |
return count; |
| 3884 |
} |
| 3885 |
|
| 3886 |
void ext4_mb_release(struct super_block *sb) |
| 3887 |
{ |
| 3888 |
ext4_group_t ngroups = ext4_get_groups_count(sb); |
| 3889 |
ext4_group_t i; |
| 3890 |
int num_meta_group_infos; |
| 3891 |
struct ext4_group_info *grinfo, ***group_info; |
| 3892 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 3893 |
struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
| 3894 |
int count; |
| 3895 |
|
| 3896 |
/* |
| 3897 |
* wait the discard work to drain all of ext4_free_data |
| 3898 |
*/ |
| 3899 |
flush_work(&sbi->s_discard_work); |
| 3900 |
WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); |
| 3901 |
|
| 3902 |
group_info = rcu_access_pointer(sbi->s_group_info); |
| 3903 |
if (group_info) { |
| 3904 |
for (i = 0; i < ngroups; i++) { |
| 3905 |
cond_resched(); |
| 3906 |
grinfo = ext4_get_group_info(sb, i); |
| 3907 |
if (!grinfo) |
| 3908 |
continue; |
| 3909 |
mb_group_bb_bitmap_free(grinfo); |
| 3910 |
ext4_lock_group(sb, i); |
| 3911 |
count = ext4_mb_cleanup_pa(grinfo); |
| 3912 |
if (count) |
| 3913 |
mb_debug(sb, "mballoc: %d PAs left\n", |
| 3914 |
count); |
| 3915 |
ext4_unlock_group(sb, i); |
| 3916 |
kmem_cache_free(cachep, grinfo); |
| 3917 |
} |
| 3918 |
num_meta_group_infos = (ngroups + |
| 3919 |
EXT4_DESC_PER_BLOCK(sb) - 1) >> |
| 3920 |
EXT4_DESC_PER_BLOCK_BITS(sb); |
| 3921 |
for (i = 0; i < num_meta_group_infos; i++) |
| 3922 |
kfree(group_info[i]); |
| 3923 |
kvfree(group_info); |
| 3924 |
} |
| 3925 |
ext4_mb_avg_fragment_size_destroy(sbi); |
| 3926 |
ext4_mb_largest_free_orders_destroy(sbi); |
| 3927 |
kfree(sbi->s_mb_offsets); |
| 3928 |
kfree(sbi->s_mb_maxs); |
| 3929 |
iput(sbi->s_buddy_cache); |
| 3930 |
if (sbi->s_mb_stats) { |
| 3931 |
ext4_msg(sb, KERN_INFO, |
| 3932 |
"mballoc: %u blocks %u reqs (%u success)", |
| 3933 |
atomic_read(&sbi->s_bal_allocated), |
| 3934 |
atomic_read(&sbi->s_bal_reqs), |
| 3935 |
atomic_read(&sbi->s_bal_success)); |
| 3936 |
ext4_msg(sb, KERN_INFO, |
| 3937 |
"mballoc: %u extents scanned, %u groups scanned, %u goal hits, " |
| 3938 |
"%u 2^N hits, %u breaks, %u lost", |
| 3939 |
atomic_read(&sbi->s_bal_ex_scanned), |
| 3940 |
atomic_read(&sbi->s_bal_groups_scanned), |
| 3941 |
atomic_read(&sbi->s_bal_goals), |
| 3942 |
atomic_read(&sbi->s_bal_2orders), |
| 3943 |
atomic_read(&sbi->s_bal_breaks), |
| 3944 |
atomic_read(&sbi->s_mb_lost_chunks)); |
| 3945 |
ext4_msg(sb, KERN_INFO, |
| 3946 |
"mballoc: %u generated and it took %llu", |
| 3947 |
atomic_read(&sbi->s_mb_buddies_generated), |
| 3948 |
atomic64_read(&sbi->s_mb_generation_time)); |
| 3949 |
ext4_msg(sb, KERN_INFO, |
| 3950 |
"mballoc: %u preallocated, %u discarded", |
| 3951 |
atomic_read(&sbi->s_mb_preallocated), |
| 3952 |
atomic_read(&sbi->s_mb_discarded)); |
| 3953 |
} |
| 3954 |
|
| 3955 |
free_percpu(sbi->s_locality_groups); |
| 3956 |
kfree(sbi->s_mb_last_groups); |
| 3957 |
} |
| 3958 |
|
| 3959 |
static inline int ext4_issue_discard(struct super_block *sb, |
| 3960 |
ext4_group_t block_group, ext4_grpblk_t cluster, int count) |
| 3961 |
{ |
| 3962 |
ext4_fsblk_t discard_block; |
| 3963 |
|
| 3964 |
discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
| 3965 |
ext4_group_first_block_no(sb, block_group)); |
| 3966 |
count = EXT4_C2B(EXT4_SB(sb), count); |
| 3967 |
trace_ext4_discard_blocks(sb, |
| 3968 |
(unsigned long long) discard_block, count); |
| 3969 |
|
| 3970 |
return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); |
| 3971 |
} |
| 3972 |
|
| 3973 |
static void ext4_free_data_in_buddy(struct super_block *sb, |
| 3974 |
struct ext4_free_data *entry) |
| 3975 |
{ |
| 3976 |
struct ext4_buddy e4b; |
| 3977 |
struct ext4_group_info *db; |
| 3978 |
int err, count = 0; |
| 3979 |
|
| 3980 |
mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", |
| 3981 |
entry->efd_count, entry->efd_group, entry); |
| 3982 |
|
| 3983 |
err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
| 3984 |
/* we expect to find existing buddy because it's pinned */ |
| 3985 |
BUG_ON(err != 0); |
| 3986 |
|
| 3987 |
atomic_sub(entry->efd_count, &EXT4_SB(sb)->s_mb_free_pending); |
| 3988 |
db = e4b.bd_info; |
| 3989 |
/* there are blocks to put in buddy to make them really free */ |
| 3990 |
count += entry->efd_count; |
| 3991 |
ext4_lock_group(sb, entry->efd_group); |
| 3992 |
/* Take it out of per group rb tree */ |
| 3993 |
rb_erase(&entry->efd_node, &(db->bb_free_root)); |
| 3994 |
mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); |
| 3995 |
|
| 3996 |
/* |
| 3997 |
* Clear the trimmed flag for the group so that the next |
| 3998 |
* ext4_trim_fs can trim it. |
| 3999 |
*/ |
| 4000 |
EXT4_MB_GRP_CLEAR_TRIMMED(db); |
| 4001 |
|
| 4002 |
if (!db->bb_free_root.rb_node) { |
| 4003 |
/* No more items in the per group rb tree |
| 4004 |
* balance refcounts from ext4_mb_free_metadata() |
| 4005 |
*/ |
| 4006 |
folio_put(e4b.bd_buddy_folio); |
| 4007 |
folio_put(e4b.bd_bitmap_folio); |
| 4008 |
} |
| 4009 |
ext4_unlock_group(sb, entry->efd_group); |
| 4010 |
ext4_mb_unload_buddy(&e4b); |
| 4011 |
|
| 4012 |
mb_debug(sb, "freed %d blocks in 1 structures\n", count); |
| 4013 |
} |
| 4014 |
|
| 4015 |
/* |
| 4016 |
* This function is called by the jbd2 layer once the commit has finished, |
| 4017 |
* so we know we can free the blocks that were released with that commit. |
| 4018 |
*/ |
| 4019 |
void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) |
| 4020 |
{ |
| 4021 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 4022 |
struct ext4_free_data *entry, *tmp; |
| 4023 |
LIST_HEAD(freed_data_list); |
| 4024 |
struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1]; |
| 4025 |
bool wake; |
| 4026 |
|
| 4027 |
list_replace_init(s_freed_head, &freed_data_list); |
| 4028 |
|
| 4029 |
list_for_each_entry(entry, &freed_data_list, efd_list) |
| 4030 |
ext4_free_data_in_buddy(sb, entry); |
| 4031 |
|
| 4032 |
if (test_opt(sb, DISCARD)) { |
| 4033 |
spin_lock(&sbi->s_md_lock); |
| 4034 |
wake = list_empty(&sbi->s_discard_list); |
| 4035 |
list_splice_tail(&freed_data_list, &sbi->s_discard_list); |
| 4036 |
spin_unlock(&sbi->s_md_lock); |
| 4037 |
if (wake) |
| 4038 |
queue_work(system_dfl_wq, &sbi->s_discard_work); |
| 4039 |
} else { |
| 4040 |
list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) |
| 4041 |
kmem_cache_free(ext4_free_data_cachep, entry); |
| 4042 |
} |
| 4043 |
} |
| 4044 |
|
| 4045 |
int __init ext4_init_mballoc(void) |
| 4046 |
{ |
| 4047 |
ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
| 4048 |
SLAB_RECLAIM_ACCOUNT); |
| 4049 |
if (ext4_pspace_cachep == NULL) |
| 4050 |
goto out; |
| 4051 |
|
| 4052 |
ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
| 4053 |
SLAB_RECLAIM_ACCOUNT); |
| 4054 |
if (ext4_ac_cachep == NULL) |
| 4055 |
goto out_pa_free; |
| 4056 |
|
| 4057 |
ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
| 4058 |
SLAB_RECLAIM_ACCOUNT); |
| 4059 |
if (ext4_free_data_cachep == NULL) |
| 4060 |
goto out_ac_free; |
| 4061 |
|
| 4062 |
return 0; |
| 4063 |
|
| 4064 |
out_ac_free: |
| 4065 |
kmem_cache_destroy(ext4_ac_cachep); |
| 4066 |
out_pa_free: |
| 4067 |
kmem_cache_destroy(ext4_pspace_cachep); |
| 4068 |
out: |
| 4069 |
return -ENOMEM; |
| 4070 |
} |
| 4071 |
|
| 4072 |
void ext4_exit_mballoc(void) |
| 4073 |
{ |
| 4074 |
/* |
| 4075 |
* Wait for completion of call_rcu()'s on ext4_pspace_cachep |
| 4076 |
* before destroying the slab cache. |
| 4077 |
*/ |
| 4078 |
rcu_barrier(); |
| 4079 |
kmem_cache_destroy(ext4_pspace_cachep); |
| 4080 |
kmem_cache_destroy(ext4_ac_cachep); |
| 4081 |
kmem_cache_destroy(ext4_free_data_cachep); |
| 4082 |
ext4_groupinfo_destroy_slabs(); |
| 4083 |
} |
| 4084 |
|
| 4085 |
#define EXT4_MB_BITMAP_MARKED_CHECK 0x0001 |
| 4086 |
#define EXT4_MB_SYNC_UPDATE 0x0002 |
| 4087 |
int |
| 4088 |
ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state, |
| 4089 |
ext4_group_t group, ext4_grpblk_t blkoff, |
| 4090 |
ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed) |
| 4091 |
{ |
| 4092 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 4093 |
struct buffer_head *bitmap_bh = NULL; |
| 4094 |
struct ext4_group_desc *gdp; |
| 4095 |
struct buffer_head *gdp_bh; |
| 4096 |
int err; |
| 4097 |
unsigned int i, already, changed = len; |
| 4098 |
|
| 4099 |
KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context, |
| 4100 |
handle, sb, state, group, blkoff, len, |
| 4101 |
flags, ret_changed); |
| 4102 |
|
| 4103 |
if (ret_changed) |
| 4104 |
*ret_changed = 0; |
| 4105 |
bitmap_bh = ext4_read_block_bitmap(sb, group); |
| 4106 |
if (IS_ERR(bitmap_bh)) |
| 4107 |
return PTR_ERR(bitmap_bh); |
| 4108 |
|
| 4109 |
if (handle) { |
| 4110 |
BUFFER_TRACE(bitmap_bh, "getting write access"); |
| 4111 |
err = ext4_journal_get_write_access(handle, sb, bitmap_bh, |
| 4112 |
EXT4_JTR_NONE); |
| 4113 |
if (err) |
| 4114 |
goto out_err; |
| 4115 |
} |
| 4116 |
|
| 4117 |
err = -EIO; |
| 4118 |
gdp = ext4_get_group_desc(sb, group, &gdp_bh); |
| 4119 |
if (!gdp) |
| 4120 |
goto out_err; |
| 4121 |
|
| 4122 |
if (handle) { |
| 4123 |
BUFFER_TRACE(gdp_bh, "get_write_access"); |
| 4124 |
err = ext4_journal_get_write_access(handle, sb, gdp_bh, |
| 4125 |
EXT4_JTR_NONE); |
| 4126 |
if (err) |
| 4127 |
goto out_err; |
| 4128 |
} |
| 4129 |
|
| 4130 |
ext4_lock_group(sb, group); |
| 4131 |
if (ext4_has_group_desc_csum(sb) && |
| 4132 |
(gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { |
| 4133 |
gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); |
| 4134 |
ext4_free_group_clusters_set(sb, gdp, |
| 4135 |
ext4_free_clusters_after_init(sb, group, gdp)); |
| 4136 |
} |
| 4137 |
|
| 4138 |
if (flags & EXT4_MB_BITMAP_MARKED_CHECK) { |
| 4139 |
already = 0; |
| 4140 |
for (i = 0; i < len; i++) |
| 4141 |
if (mb_test_bit(blkoff + i, bitmap_bh->b_data) == |
| 4142 |
state) |
| 4143 |
already++; |
| 4144 |
changed = len - already; |
| 4145 |
} |
| 4146 |
|
| 4147 |
if (state) { |
| 4148 |
mb_set_bits(bitmap_bh->b_data, blkoff, len); |
| 4149 |
ext4_free_group_clusters_set(sb, gdp, |
| 4150 |
ext4_free_group_clusters(sb, gdp) - changed); |
| 4151 |
} else { |
| 4152 |
mb_clear_bits(bitmap_bh->b_data, blkoff, len); |
| 4153 |
ext4_free_group_clusters_set(sb, gdp, |
| 4154 |
ext4_free_group_clusters(sb, gdp) + changed); |
| 4155 |
} |
| 4156 |
|
| 4157 |
ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); |
| 4158 |
ext4_group_desc_csum_set(sb, group, gdp); |
| 4159 |
ext4_unlock_group(sb, group); |
| 4160 |
if (ret_changed) |
| 4161 |
*ret_changed = changed; |
| 4162 |
|
| 4163 |
if (sbi->s_log_groups_per_flex) { |
| 4164 |
ext4_group_t flex_group = ext4_flex_group(sbi, group); |
| 4165 |
struct flex_groups *fg = sbi_array_rcu_deref(sbi, |
| 4166 |
s_flex_groups, flex_group); |
| 4167 |
|
| 4168 |
if (state) |
| 4169 |
atomic64_sub(changed, &fg->free_clusters); |
| 4170 |
else |
| 4171 |
atomic64_add(changed, &fg->free_clusters); |
| 4172 |
} |
| 4173 |
|
| 4174 |
err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
| 4175 |
if (err) |
| 4176 |
goto out_err; |
| 4177 |
err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
| 4178 |
if (err) |
| 4179 |
goto out_err; |
| 4180 |
|
| 4181 |
if (flags & EXT4_MB_SYNC_UPDATE) { |
| 4182 |
sync_dirty_buffer(bitmap_bh); |
| 4183 |
sync_dirty_buffer(gdp_bh); |
| 4184 |
} |
| 4185 |
|
| 4186 |
out_err: |
| 4187 |
brelse(bitmap_bh); |
| 4188 |
return err; |
| 4189 |
} |
| 4190 |
|
| 4191 |
/* |
| 4192 |
* Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
| 4193 |
* Returns 0 if success or error code |
| 4194 |
*/ |
| 4195 |
static noinline_for_stack int |
| 4196 |
ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, handle_t *handle) |
| 4197 |
{ |
| 4198 |
struct ext4_group_desc *gdp; |
| 4199 |
struct ext4_sb_info *sbi; |
| 4200 |
struct super_block *sb; |
| 4201 |
ext4_fsblk_t block; |
| 4202 |
int err, len; |
| 4203 |
int flags = 0; |
| 4204 |
ext4_grpblk_t changed; |
| 4205 |
|
| 4206 |
BUG_ON(ac->ac_status != AC_STATUS_FOUND); |
| 4207 |
BUG_ON(ac->ac_b_ex.fe_len <= 0); |
| 4208 |
|
| 4209 |
sb = ac->ac_sb; |
| 4210 |
sbi = EXT4_SB(sb); |
| 4211 |
|
| 4212 |
gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL); |
| 4213 |
if (!gdp) |
| 4214 |
return -EIO; |
| 4215 |
ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
| 4216 |
ext4_free_group_clusters(sb, gdp)); |
| 4217 |
|
| 4218 |
block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
| 4219 |
len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
| 4220 |
if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { |
| 4221 |
ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
| 4222 |
"fs metadata", block, block+len); |
| 4223 |
/* File system mounted not to panic on error |
| 4224 |
* Fix the bitmap and return EFSCORRUPTED |
| 4225 |
* We leak some of the blocks here. |
| 4226 |
*/ |
| 4227 |
err = ext4_mb_mark_context(handle, sb, true, |
| 4228 |
ac->ac_b_ex.fe_group, |
| 4229 |
ac->ac_b_ex.fe_start, |
| 4230 |
ac->ac_b_ex.fe_len, |
| 4231 |
0, NULL); |
| 4232 |
if (!err) |
| 4233 |
err = -EFSCORRUPTED; |
| 4234 |
return err; |
| 4235 |
} |
| 4236 |
|
| 4237 |
#ifdef AGGRESSIVE_CHECK |
| 4238 |
flags |= EXT4_MB_BITMAP_MARKED_CHECK; |
| 4239 |
#endif |
| 4240 |
err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group, |
| 4241 |
ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len, |
| 4242 |
flags, &changed); |
| 4243 |
|
| 4244 |
if (err && changed == 0) |
| 4245 |
return err; |
| 4246 |
|
| 4247 |
#ifdef AGGRESSIVE_CHECK |
| 4248 |
BUG_ON(changed != ac->ac_b_ex.fe_len); |
| 4249 |
#endif |
| 4250 |
percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
| 4251 |
|
| 4252 |
return err; |
| 4253 |
} |
| 4254 |
|
| 4255 |
/* |
| 4256 |
* Idempotent helper for Ext4 fast commit replay path to set the state of |
| 4257 |
* blocks in bitmaps and update counters. |
| 4258 |
*/ |
| 4259 |
void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, |
| 4260 |
int len, bool state) |
| 4261 |
{ |
| 4262 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 4263 |
ext4_group_t group; |
| 4264 |
ext4_grpblk_t blkoff; |
| 4265 |
int err = 0; |
| 4266 |
unsigned int clen, thisgrp_len; |
| 4267 |
|
| 4268 |
while (len > 0) { |
| 4269 |
ext4_get_group_no_and_offset(sb, block, &group, &blkoff); |
| 4270 |
|
| 4271 |
/* |
| 4272 |
* Check to see if we are freeing blocks across a group |
| 4273 |
* boundary. |
| 4274 |
* In case of flex_bg, this can happen that (block, len) may |
| 4275 |
* span across more than one group. In that case we need to |
| 4276 |
* get the corresponding group metadata to work with. |
| 4277 |
* For this we have goto again loop. |
| 4278 |
*/ |
| 4279 |
thisgrp_len = min(len, EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); |
| 4280 |
clen = EXT4_NUM_B2C(sbi, thisgrp_len); |
| 4281 |
|
| 4282 |
if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { |
| 4283 |
ext4_error(sb, "Marking blocks in system zone - " |
| 4284 |
"Block = %llu, len = %u", |
| 4285 |
block, thisgrp_len); |
| 4286 |
break; |
| 4287 |
} |
| 4288 |
|
| 4289 |
err = ext4_mb_mark_context(NULL, sb, state, |
| 4290 |
group, blkoff, clen, |
| 4291 |
EXT4_MB_BITMAP_MARKED_CHECK | |
| 4292 |
EXT4_MB_SYNC_UPDATE, |
| 4293 |
NULL); |
| 4294 |
if (err) |
| 4295 |
break; |
| 4296 |
|
| 4297 |
block += thisgrp_len; |
| 4298 |
len -= thisgrp_len; |
| 4299 |
BUG_ON(len < 0); |
| 4300 |
} |
| 4301 |
} |
| 4302 |
|
| 4303 |
/* |
| 4304 |
* here we normalize request for locality group |
| 4305 |
* Group request are normalized to s_mb_group_prealloc, which goes to |
| 4306 |
* s_strip if we set the same via mount option. |
| 4307 |
* s_mb_group_prealloc can be configured via |
| 4308 |
* /sys/fs/ext4/<partition>/mb_group_prealloc |
| 4309 |
* |
| 4310 |
* XXX: should we try to preallocate more than the group has now? |
| 4311 |
*/ |
| 4312 |
static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) |
| 4313 |
{ |
| 4314 |
struct super_block *sb = ac->ac_sb; |
| 4315 |
struct ext4_locality_group *lg = ac->ac_lg; |
| 4316 |
|
| 4317 |
BUG_ON(lg == NULL); |
| 4318 |
ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
| 4319 |
mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); |
| 4320 |
} |
| 4321 |
|
| 4322 |
/* |
| 4323 |
* This function returns the next element to look at during inode |
| 4324 |
* PA rbtree walk. We assume that we have held the inode PA rbtree lock |
| 4325 |
* (ei->i_prealloc_lock) |
| 4326 |
* |
| 4327 |
* new_start The start of the range we want to compare |
| 4328 |
* cur_start The existing start that we are comparing against |
| 4329 |
* node The node of the rb_tree |
| 4330 |
*/ |
| 4331 |
static inline struct rb_node* |
| 4332 |
ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node) |
| 4333 |
{ |
| 4334 |
if (new_start < cur_start) |
| 4335 |
return node->rb_left; |
| 4336 |
else |
| 4337 |
return node->rb_right; |
| 4338 |
} |
| 4339 |
|
| 4340 |
static inline void |
| 4341 |
ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac, |
| 4342 |
ext4_lblk_t start, loff_t end) |
| 4343 |
{ |
| 4344 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4345 |
struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
| 4346 |
struct ext4_prealloc_space *tmp_pa; |
| 4347 |
ext4_lblk_t tmp_pa_start; |
| 4348 |
loff_t tmp_pa_end; |
| 4349 |
struct rb_node *iter; |
| 4350 |
|
| 4351 |
read_lock(&ei->i_prealloc_lock); |
| 4352 |
for (iter = ei->i_prealloc_node.rb_node; iter; |
| 4353 |
iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) { |
| 4354 |
tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
| 4355 |
pa_node.inode_node); |
| 4356 |
tmp_pa_start = tmp_pa->pa_lstart; |
| 4357 |
tmp_pa_end = pa_logical_end(sbi, tmp_pa); |
| 4358 |
|
| 4359 |
spin_lock(&tmp_pa->pa_lock); |
| 4360 |
if (tmp_pa->pa_deleted == 0) |
| 4361 |
BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start)); |
| 4362 |
spin_unlock(&tmp_pa->pa_lock); |
| 4363 |
} |
| 4364 |
read_unlock(&ei->i_prealloc_lock); |
| 4365 |
} |
| 4366 |
|
| 4367 |
/* |
| 4368 |
* Given an allocation context "ac" and a range "start", "end", check |
| 4369 |
* and adjust boundaries if the range overlaps with any of the existing |
| 4370 |
* preallocatoins stored in the corresponding inode of the allocation context. |
| 4371 |
* |
| 4372 |
* Parameters: |
| 4373 |
* ac allocation context |
| 4374 |
* start start of the new range |
| 4375 |
* end end of the new range |
| 4376 |
*/ |
| 4377 |
static inline void |
| 4378 |
ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac, |
| 4379 |
ext4_lblk_t *start, loff_t *end) |
| 4380 |
{ |
| 4381 |
struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
| 4382 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4383 |
struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL; |
| 4384 |
struct rb_node *iter; |
| 4385 |
ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1; |
| 4386 |
loff_t new_end, tmp_pa_end, left_pa_end = -1; |
| 4387 |
|
| 4388 |
new_start = *start; |
| 4389 |
new_end = *end; |
| 4390 |
|
| 4391 |
/* |
| 4392 |
* Adjust the normalized range so that it doesn't overlap with any |
| 4393 |
* existing preallocated blocks(PAs). Make sure to hold the rbtree lock |
| 4394 |
* so it doesn't change underneath us. |
| 4395 |
*/ |
| 4396 |
read_lock(&ei->i_prealloc_lock); |
| 4397 |
|
| 4398 |
/* Step 1: find any one immediate neighboring PA of the normalized range */ |
| 4399 |
for (iter = ei->i_prealloc_node.rb_node; iter; |
| 4400 |
iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, |
| 4401 |
tmp_pa_start, iter)) { |
| 4402 |
tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
| 4403 |
pa_node.inode_node); |
| 4404 |
tmp_pa_start = tmp_pa->pa_lstart; |
| 4405 |
tmp_pa_end = pa_logical_end(sbi, tmp_pa); |
| 4406 |
|
| 4407 |
/* PA must not overlap original request */ |
| 4408 |
spin_lock(&tmp_pa->pa_lock); |
| 4409 |
if (tmp_pa->pa_deleted == 0) |
| 4410 |
BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end || |
| 4411 |
ac->ac_o_ex.fe_logical < tmp_pa_start)); |
| 4412 |
spin_unlock(&tmp_pa->pa_lock); |
| 4413 |
} |
| 4414 |
|
| 4415 |
/* |
| 4416 |
* Step 2: check if the found PA is left or right neighbor and |
| 4417 |
* get the other neighbor |
| 4418 |
*/ |
| 4419 |
if (tmp_pa) { |
| 4420 |
if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) { |
| 4421 |
struct rb_node *tmp; |
| 4422 |
|
| 4423 |
left_pa = tmp_pa; |
| 4424 |
tmp = rb_next(&left_pa->pa_node.inode_node); |
| 4425 |
if (tmp) { |
| 4426 |
right_pa = rb_entry(tmp, |
| 4427 |
struct ext4_prealloc_space, |
| 4428 |
pa_node.inode_node); |
| 4429 |
} |
| 4430 |
} else { |
| 4431 |
struct rb_node *tmp; |
| 4432 |
|
| 4433 |
right_pa = tmp_pa; |
| 4434 |
tmp = rb_prev(&right_pa->pa_node.inode_node); |
| 4435 |
if (tmp) { |
| 4436 |
left_pa = rb_entry(tmp, |
| 4437 |
struct ext4_prealloc_space, |
| 4438 |
pa_node.inode_node); |
| 4439 |
} |
| 4440 |
} |
| 4441 |
} |
| 4442 |
|
| 4443 |
/* Step 3: get the non deleted neighbors */ |
| 4444 |
if (left_pa) { |
| 4445 |
for (iter = &left_pa->pa_node.inode_node;; |
| 4446 |
iter = rb_prev(iter)) { |
| 4447 |
if (!iter) { |
| 4448 |
left_pa = NULL; |
| 4449 |
break; |
| 4450 |
} |
| 4451 |
|
| 4452 |
tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
| 4453 |
pa_node.inode_node); |
| 4454 |
left_pa = tmp_pa; |
| 4455 |
spin_lock(&tmp_pa->pa_lock); |
| 4456 |
if (tmp_pa->pa_deleted == 0) { |
| 4457 |
spin_unlock(&tmp_pa->pa_lock); |
| 4458 |
break; |
| 4459 |
} |
| 4460 |
spin_unlock(&tmp_pa->pa_lock); |
| 4461 |
} |
| 4462 |
} |
| 4463 |
|
| 4464 |
if (right_pa) { |
| 4465 |
for (iter = &right_pa->pa_node.inode_node;; |
| 4466 |
iter = rb_next(iter)) { |
| 4467 |
if (!iter) { |
| 4468 |
right_pa = NULL; |
| 4469 |
break; |
| 4470 |
} |
| 4471 |
|
| 4472 |
tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
| 4473 |
pa_node.inode_node); |
| 4474 |
right_pa = tmp_pa; |
| 4475 |
spin_lock(&tmp_pa->pa_lock); |
| 4476 |
if (tmp_pa->pa_deleted == 0) { |
| 4477 |
spin_unlock(&tmp_pa->pa_lock); |
| 4478 |
break; |
| 4479 |
} |
| 4480 |
spin_unlock(&tmp_pa->pa_lock); |
| 4481 |
} |
| 4482 |
} |
| 4483 |
|
| 4484 |
if (left_pa) { |
| 4485 |
left_pa_end = pa_logical_end(sbi, left_pa); |
| 4486 |
BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical); |
| 4487 |
} |
| 4488 |
|
| 4489 |
if (right_pa) { |
| 4490 |
right_pa_start = right_pa->pa_lstart; |
| 4491 |
BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical); |
| 4492 |
} |
| 4493 |
|
| 4494 |
/* Step 4: trim our normalized range to not overlap with the neighbors */ |
| 4495 |
if (left_pa) { |
| 4496 |
if (left_pa_end > new_start) |
| 4497 |
new_start = left_pa_end; |
| 4498 |
} |
| 4499 |
|
| 4500 |
if (right_pa) { |
| 4501 |
if (right_pa_start < new_end) |
| 4502 |
new_end = right_pa_start; |
| 4503 |
} |
| 4504 |
read_unlock(&ei->i_prealloc_lock); |
| 4505 |
|
| 4506 |
/* XXX: extra loop to check we really don't overlap preallocations */ |
| 4507 |
ext4_mb_pa_assert_overlap(ac, new_start, new_end); |
| 4508 |
|
| 4509 |
*start = new_start; |
| 4510 |
*end = new_end; |
| 4511 |
} |
| 4512 |
|
| 4513 |
/* |
| 4514 |
* Normalization means making request better in terms of |
| 4515 |
* size and alignment |
| 4516 |
*/ |
| 4517 |
static noinline_for_stack void |
| 4518 |
ext4_mb_normalize_request(struct ext4_allocation_context *ac, |
| 4519 |
struct ext4_allocation_request *ar) |
| 4520 |
{ |
| 4521 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4522 |
struct ext4_super_block *es = sbi->s_es; |
| 4523 |
int bsbits, max; |
| 4524 |
loff_t size, start_off, end; |
| 4525 |
loff_t orig_size __maybe_unused; |
| 4526 |
ext4_lblk_t start; |
| 4527 |
|
| 4528 |
/* do normalize only data requests, metadata requests |
| 4529 |
do not need preallocation */ |
| 4530 |
if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) |
| 4531 |
return; |
| 4532 |
|
| 4533 |
/* sometime caller may want exact blocks */ |
| 4534 |
if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
| 4535 |
return; |
| 4536 |
|
| 4537 |
/* caller may indicate that preallocation isn't |
| 4538 |
* required (it's a tail, for example) */ |
| 4539 |
if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) |
| 4540 |
return; |
| 4541 |
|
| 4542 |
if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { |
| 4543 |
ext4_mb_normalize_group_request(ac); |
| 4544 |
return ; |
| 4545 |
} |
| 4546 |
|
| 4547 |
bsbits = ac->ac_sb->s_blocksize_bits; |
| 4548 |
|
| 4549 |
/* first, let's learn actual file size |
| 4550 |
* given current request is allocated */ |
| 4551 |
size = extent_logical_end(sbi, &ac->ac_o_ex); |
| 4552 |
size = size << bsbits; |
| 4553 |
if (size < i_size_read(ac->ac_inode)) |
| 4554 |
size = i_size_read(ac->ac_inode); |
| 4555 |
orig_size = size; |
| 4556 |
|
| 4557 |
/* max size of free chunks */ |
| 4558 |
max = 2 << bsbits; |
| 4559 |
|
| 4560 |
#define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
| 4561 |
(req <= (size) || max <= (chunk_size)) |
| 4562 |
|
| 4563 |
/* first, try to predict filesize */ |
| 4564 |
/* XXX: should this table be tunable? */ |
| 4565 |
start_off = 0; |
| 4566 |
if (size <= 16 * 1024) { |
| 4567 |
size = 16 * 1024; |
| 4568 |
} else if (size <= 32 * 1024) { |
| 4569 |
size = 32 * 1024; |
| 4570 |
} else if (size <= 64 * 1024) { |
| 4571 |
size = 64 * 1024; |
| 4572 |
} else if (size <= 128 * 1024) { |
| 4573 |
size = 128 * 1024; |
| 4574 |
} else if (size <= 256 * 1024) { |
| 4575 |
size = 256 * 1024; |
| 4576 |
} else if (size <= 512 * 1024) { |
| 4577 |
size = 512 * 1024; |
| 4578 |
} else if (size <= 1024 * 1024) { |
| 4579 |
size = 1024 * 1024; |
| 4580 |
} else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
| 4581 |
start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
| 4582 |
(21 - bsbits)) << 21; |
| 4583 |
size = 2 * 1024 * 1024; |
| 4584 |
} else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { |
| 4585 |
start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
| 4586 |
(22 - bsbits)) << 22; |
| 4587 |
size = 4 * 1024 * 1024; |
| 4588 |
} else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len), |
| 4589 |
(8<<20)>>bsbits, max, 8 * 1024)) { |
| 4590 |
start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
| 4591 |
(23 - bsbits)) << 23; |
| 4592 |
size = 8 * 1024 * 1024; |
| 4593 |
} else { |
| 4594 |
start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
| 4595 |
size = (loff_t) EXT4_C2B(sbi, |
| 4596 |
ac->ac_o_ex.fe_len) << bsbits; |
| 4597 |
} |
| 4598 |
size = size >> bsbits; |
| 4599 |
start = start_off >> bsbits; |
| 4600 |
|
| 4601 |
/* |
| 4602 |
* For tiny groups (smaller than 8MB) the chosen allocation |
| 4603 |
* alignment may be larger than group size. Make sure the |
| 4604 |
* alignment does not move allocation to a different group which |
| 4605 |
* makes mballoc fail assertions later. |
| 4606 |
*/ |
| 4607 |
start = max(start, rounddown(ac->ac_o_ex.fe_logical, |
| 4608 |
(ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); |
| 4609 |
|
| 4610 |
/* avoid unnecessary preallocation that may trigger assertions */ |
| 4611 |
if (start + size > EXT_MAX_BLOCKS) |
| 4612 |
size = EXT_MAX_BLOCKS - start; |
| 4613 |
|
| 4614 |
/* don't cover already allocated blocks in selected range */ |
| 4615 |
if (ar->pleft && start <= ar->lleft) { |
| 4616 |
size -= ar->lleft + 1 - start; |
| 4617 |
start = ar->lleft + 1; |
| 4618 |
} |
| 4619 |
if (ar->pright && start + size - 1 >= ar->lright) |
| 4620 |
size -= start + size - ar->lright; |
| 4621 |
|
| 4622 |
/* |
| 4623 |
* Trim allocation request for filesystems with artificially small |
| 4624 |
* groups. |
| 4625 |
*/ |
| 4626 |
if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) |
| 4627 |
size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); |
| 4628 |
|
| 4629 |
end = start + size; |
| 4630 |
|
| 4631 |
ext4_mb_pa_adjust_overlap(ac, &start, &end); |
| 4632 |
|
| 4633 |
size = end - start; |
| 4634 |
|
| 4635 |
/* |
| 4636 |
* In this function "start" and "size" are normalized for better |
| 4637 |
* alignment and length such that we could preallocate more blocks. |
| 4638 |
* This normalization is done such that original request of |
| 4639 |
* ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and |
| 4640 |
* "size" boundaries. |
| 4641 |
* (Note fe_len can be relaxed since FS block allocation API does not |
| 4642 |
* provide gurantee on number of contiguous blocks allocation since that |
| 4643 |
* depends upon free space left, etc). |
| 4644 |
* In case of inode pa, later we use the allocated blocks |
| 4645 |
* [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated |
| 4646 |
* range of goal/best blocks [start, size] to put it at the |
| 4647 |
* ac_o_ex.fe_logical extent of this inode. |
| 4648 |
* (See ext4_mb_use_inode_pa() for more details) |
| 4649 |
*/ |
| 4650 |
if (start + size <= ac->ac_o_ex.fe_logical || |
| 4651 |
start > ac->ac_o_ex.fe_logical) { |
| 4652 |
ext4_msg(ac->ac_sb, KERN_ERR, |
| 4653 |
"start %lu, size %lu, fe_logical %lu", |
| 4654 |
(unsigned long) start, (unsigned long) size, |
| 4655 |
(unsigned long) ac->ac_o_ex.fe_logical); |
| 4656 |
BUG(); |
| 4657 |
} |
| 4658 |
BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
| 4659 |
|
| 4660 |
/* now prepare goal request */ |
| 4661 |
|
| 4662 |
/* XXX: is it better to align blocks WRT to logical |
| 4663 |
* placement or satisfy big request as is */ |
| 4664 |
ac->ac_g_ex.fe_logical = start; |
| 4665 |
ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
| 4666 |
ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; |
| 4667 |
|
| 4668 |
/* define goal start in order to merge */ |
| 4669 |
if (ar->pright && (ar->lright == (start + size)) && |
| 4670 |
ar->pright >= size && |
| 4671 |
ar->pright - size >= le32_to_cpu(es->s_first_data_block)) { |
| 4672 |
/* merge to the right */ |
| 4673 |
ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, |
| 4674 |
&ac->ac_g_ex.fe_group, |
| 4675 |
&ac->ac_g_ex.fe_start); |
| 4676 |
ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; |
| 4677 |
} |
| 4678 |
if (ar->pleft && (ar->lleft + 1 == start) && |
| 4679 |
ar->pleft + 1 < ext4_blocks_count(es)) { |
| 4680 |
/* merge to the left */ |
| 4681 |
ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, |
| 4682 |
&ac->ac_g_ex.fe_group, |
| 4683 |
&ac->ac_g_ex.fe_start); |
| 4684 |
ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; |
| 4685 |
} |
| 4686 |
|
| 4687 |
mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, |
| 4688 |
orig_size, start); |
| 4689 |
} |
| 4690 |
|
| 4691 |
static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) |
| 4692 |
{ |
| 4693 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4694 |
|
| 4695 |
if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { |
| 4696 |
atomic_inc(&sbi->s_bal_reqs); |
| 4697 |
atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); |
| 4698 |
if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
| 4699 |
atomic_inc(&sbi->s_bal_success); |
| 4700 |
|
| 4701 |
atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); |
| 4702 |
for (int i=0; i<EXT4_MB_NUM_CRS; i++) { |
| 4703 |
atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]); |
| 4704 |
} |
| 4705 |
|
| 4706 |
atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); |
| 4707 |
if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && |
| 4708 |
ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) |
| 4709 |
atomic_inc(&sbi->s_bal_goals); |
| 4710 |
/* did we allocate as much as normalizer originally wanted? */ |
| 4711 |
if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len) |
| 4712 |
atomic_inc(&sbi->s_bal_len_goals); |
| 4713 |
|
| 4714 |
if (ac->ac_found > sbi->s_mb_max_to_scan) |
| 4715 |
atomic_inc(&sbi->s_bal_breaks); |
| 4716 |
} |
| 4717 |
|
| 4718 |
if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
| 4719 |
trace_ext4_mballoc_alloc(ac); |
| 4720 |
else |
| 4721 |
trace_ext4_mballoc_prealloc(ac); |
| 4722 |
} |
| 4723 |
|
| 4724 |
/* |
| 4725 |
* Called on failure; free up any blocks from the inode PA for this |
| 4726 |
* context. We don't need this for MB_GROUP_PA because we only change |
| 4727 |
* pa_free in ext4_mb_release_context(), but on failure, we've already |
| 4728 |
* zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. |
| 4729 |
*/ |
| 4730 |
static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) |
| 4731 |
{ |
| 4732 |
struct ext4_prealloc_space *pa = ac->ac_pa; |
| 4733 |
struct ext4_buddy e4b; |
| 4734 |
int err; |
| 4735 |
|
| 4736 |
if (pa == NULL) { |
| 4737 |
if (ac->ac_f_ex.fe_len == 0) |
| 4738 |
return; |
| 4739 |
err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); |
| 4740 |
if (WARN_RATELIMIT(err, |
| 4741 |
"ext4: mb_load_buddy failed (%d)", err)) |
| 4742 |
/* |
| 4743 |
* This should never happen since we pin the |
| 4744 |
* folios in the ext4_allocation_context so |
| 4745 |
* ext4_mb_load_buddy() should never fail. |
| 4746 |
*/ |
| 4747 |
return; |
| 4748 |
ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); |
| 4749 |
mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, |
| 4750 |
ac->ac_f_ex.fe_len); |
| 4751 |
ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); |
| 4752 |
ext4_mb_unload_buddy(&e4b); |
| 4753 |
return; |
| 4754 |
} |
| 4755 |
if (pa->pa_type == MB_INODE_PA) { |
| 4756 |
spin_lock(&pa->pa_lock); |
| 4757 |
pa->pa_free += ac->ac_b_ex.fe_len; |
| 4758 |
spin_unlock(&pa->pa_lock); |
| 4759 |
} |
| 4760 |
} |
| 4761 |
|
| 4762 |
/* |
| 4763 |
* use blocks preallocated to inode |
| 4764 |
*/ |
| 4765 |
static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, |
| 4766 |
struct ext4_prealloc_space *pa) |
| 4767 |
{ |
| 4768 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4769 |
ext4_fsblk_t start; |
| 4770 |
ext4_fsblk_t end; |
| 4771 |
int len; |
| 4772 |
|
| 4773 |
/* found preallocated blocks, use them */ |
| 4774 |
start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); |
| 4775 |
end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
| 4776 |
start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); |
| 4777 |
len = EXT4_NUM_B2C(sbi, end - start); |
| 4778 |
ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
| 4779 |
&ac->ac_b_ex.fe_start); |
| 4780 |
ac->ac_b_ex.fe_len = len; |
| 4781 |
ac->ac_status = AC_STATUS_FOUND; |
| 4782 |
ac->ac_pa = pa; |
| 4783 |
|
| 4784 |
BUG_ON(start < pa->pa_pstart); |
| 4785 |
BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
| 4786 |
BUG_ON(pa->pa_free < len); |
| 4787 |
BUG_ON(ac->ac_b_ex.fe_len <= 0); |
| 4788 |
pa->pa_free -= len; |
| 4789 |
|
| 4790 |
mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); |
| 4791 |
} |
| 4792 |
|
| 4793 |
/* |
| 4794 |
* use blocks preallocated to locality group |
| 4795 |
*/ |
| 4796 |
static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, |
| 4797 |
struct ext4_prealloc_space *pa) |
| 4798 |
{ |
| 4799 |
unsigned int len = ac->ac_o_ex.fe_len; |
| 4800 |
|
| 4801 |
ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
| 4802 |
&ac->ac_b_ex.fe_group, |
| 4803 |
&ac->ac_b_ex.fe_start); |
| 4804 |
ac->ac_b_ex.fe_len = len; |
| 4805 |
ac->ac_status = AC_STATUS_FOUND; |
| 4806 |
ac->ac_pa = pa; |
| 4807 |
|
| 4808 |
/* we don't correct pa_pstart or pa_len here to avoid |
| 4809 |
* possible race when the group is being loaded concurrently |
| 4810 |
* instead we correct pa later, after blocks are marked |
| 4811 |
* in on-disk bitmap -- see ext4_mb_release_context() |
| 4812 |
* Other CPUs are prevented from allocating from this pa by lg_mutex |
| 4813 |
*/ |
| 4814 |
mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", |
| 4815 |
pa->pa_lstart, len, pa); |
| 4816 |
} |
| 4817 |
|
| 4818 |
/* |
| 4819 |
* Return the prealloc space that have minimal distance |
| 4820 |
* from the goal block. @cpa is the prealloc |
| 4821 |
* space that is having currently known minimal distance |
| 4822 |
* from the goal block. |
| 4823 |
*/ |
| 4824 |
static struct ext4_prealloc_space * |
| 4825 |
ext4_mb_check_group_pa(ext4_fsblk_t goal_block, |
| 4826 |
struct ext4_prealloc_space *pa, |
| 4827 |
struct ext4_prealloc_space *cpa) |
| 4828 |
{ |
| 4829 |
ext4_fsblk_t cur_distance, new_distance; |
| 4830 |
|
| 4831 |
if (cpa == NULL) { |
| 4832 |
atomic_inc(&pa->pa_count); |
| 4833 |
return pa; |
| 4834 |
} |
| 4835 |
cur_distance = abs(goal_block - cpa->pa_pstart); |
| 4836 |
new_distance = abs(goal_block - pa->pa_pstart); |
| 4837 |
|
| 4838 |
if (cur_distance <= new_distance) |
| 4839 |
return cpa; |
| 4840 |
|
| 4841 |
/* drop the previous reference */ |
| 4842 |
atomic_dec(&cpa->pa_count); |
| 4843 |
atomic_inc(&pa->pa_count); |
| 4844 |
return pa; |
| 4845 |
} |
| 4846 |
|
| 4847 |
/* |
| 4848 |
* check if found pa meets EXT4_MB_HINT_GOAL_ONLY |
| 4849 |
*/ |
| 4850 |
static bool |
| 4851 |
ext4_mb_pa_goal_check(struct ext4_allocation_context *ac, |
| 4852 |
struct ext4_prealloc_space *pa) |
| 4853 |
{ |
| 4854 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4855 |
ext4_fsblk_t start; |
| 4856 |
|
| 4857 |
if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))) |
| 4858 |
return true; |
| 4859 |
|
| 4860 |
/* |
| 4861 |
* If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted |
| 4862 |
* in ext4_mb_normalize_request and will keep same with ac_o_ex |
| 4863 |
* from ext4_mb_initialize_context. Choose ac_g_ex here to keep |
| 4864 |
* consistent with ext4_mb_find_by_goal. |
| 4865 |
*/ |
| 4866 |
start = pa->pa_pstart + |
| 4867 |
(ac->ac_g_ex.fe_logical - pa->pa_lstart); |
| 4868 |
if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start) |
| 4869 |
return false; |
| 4870 |
|
| 4871 |
if (ac->ac_g_ex.fe_len > pa->pa_len - |
| 4872 |
EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart)) |
| 4873 |
return false; |
| 4874 |
|
| 4875 |
return true; |
| 4876 |
} |
| 4877 |
|
| 4878 |
/* |
| 4879 |
* search goal blocks in preallocated space |
| 4880 |
*/ |
| 4881 |
static noinline_for_stack bool |
| 4882 |
ext4_mb_use_preallocated(struct ext4_allocation_context *ac) |
| 4883 |
{ |
| 4884 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 4885 |
int order, i; |
| 4886 |
struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
| 4887 |
struct ext4_locality_group *lg; |
| 4888 |
struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL; |
| 4889 |
struct rb_node *iter; |
| 4890 |
ext4_fsblk_t goal_block; |
| 4891 |
|
| 4892 |
/* only data can be preallocated */ |
| 4893 |
if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) |
| 4894 |
return false; |
| 4895 |
|
| 4896 |
/* |
| 4897 |
* first, try per-file preallocation by searching the inode pa rbtree. |
| 4898 |
* |
| 4899 |
* Here, we can't do a direct traversal of the tree because |
| 4900 |
* ext4_mb_discard_group_preallocation() can paralelly mark the pa |
| 4901 |
* deleted and that can cause direct traversal to skip some entries. |
| 4902 |
*/ |
| 4903 |
read_lock(&ei->i_prealloc_lock); |
| 4904 |
|
| 4905 |
if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) { |
| 4906 |
goto try_group_pa; |
| 4907 |
} |
| 4908 |
|
| 4909 |
/* |
| 4910 |
* Step 1: Find a pa with logical start immediately adjacent to the |
| 4911 |
* original logical start. This could be on the left or right. |
| 4912 |
* |
| 4913 |
* (tmp_pa->pa_lstart never changes so we can skip locking for it). |
| 4914 |
*/ |
| 4915 |
for (iter = ei->i_prealloc_node.rb_node; iter; |
| 4916 |
iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, |
| 4917 |
tmp_pa->pa_lstart, iter)) { |
| 4918 |
tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
| 4919 |
pa_node.inode_node); |
| 4920 |
} |
| 4921 |
|
| 4922 |
/* |
| 4923 |
* Step 2: The adjacent pa might be to the right of logical start, find |
| 4924 |
* the left adjacent pa. After this step we'd have a valid tmp_pa whose |
| 4925 |
* logical start is towards the left of original request's logical start |
| 4926 |
*/ |
| 4927 |
if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
| 4928 |
struct rb_node *tmp; |
| 4929 |
tmp = rb_prev(&tmp_pa->pa_node.inode_node); |
| 4930 |
|
| 4931 |
if (tmp) { |
| 4932 |
tmp_pa = rb_entry(tmp, struct ext4_prealloc_space, |
| 4933 |
pa_node.inode_node); |
| 4934 |
} else { |
| 4935 |
/* |
| 4936 |
* If there is no adjacent pa to the left then finding |
| 4937 |
* an overlapping pa is not possible hence stop searching |
| 4938 |
* inode pa tree |
| 4939 |
*/ |
| 4940 |
goto try_group_pa; |
| 4941 |
} |
| 4942 |
} |
| 4943 |
|
| 4944 |
BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); |
| 4945 |
|
| 4946 |
/* |
| 4947 |
* Step 3: If the left adjacent pa is deleted, keep moving left to find |
| 4948 |
* the first non deleted adjacent pa. After this step we should have a |
| 4949 |
* valid tmp_pa which is guaranteed to be non deleted. |
| 4950 |
*/ |
| 4951 |
for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) { |
| 4952 |
if (!iter) { |
| 4953 |
/* |
| 4954 |
* no non deleted left adjacent pa, so stop searching |
| 4955 |
* inode pa tree |
| 4956 |
*/ |
| 4957 |
goto try_group_pa; |
| 4958 |
} |
| 4959 |
tmp_pa = rb_entry(iter, struct ext4_prealloc_space, |
| 4960 |
pa_node.inode_node); |
| 4961 |
spin_lock(&tmp_pa->pa_lock); |
| 4962 |
if (tmp_pa->pa_deleted == 0) { |
| 4963 |
/* |
| 4964 |
* We will keep holding the pa_lock from |
| 4965 |
* this point on because we don't want group discard |
| 4966 |
* to delete this pa underneath us. Since group |
| 4967 |
* discard is anyways an ENOSPC operation it |
| 4968 |
* should be okay for it to wait a few more cycles. |
| 4969 |
*/ |
| 4970 |
break; |
| 4971 |
} else { |
| 4972 |
spin_unlock(&tmp_pa->pa_lock); |
| 4973 |
} |
| 4974 |
} |
| 4975 |
|
| 4976 |
BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); |
| 4977 |
BUG_ON(tmp_pa->pa_deleted == 1); |
| 4978 |
|
| 4979 |
/* |
| 4980 |
* Step 4: We now have the non deleted left adjacent pa. Only this |
| 4981 |
* pa can possibly satisfy the request hence check if it overlaps |
| 4982 |
* original logical start and stop searching if it doesn't. |
| 4983 |
*/ |
| 4984 |
if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) { |
| 4985 |
spin_unlock(&tmp_pa->pa_lock); |
| 4986 |
goto try_group_pa; |
| 4987 |
} |
| 4988 |
|
| 4989 |
/* non-extent files can't have physical blocks past 2^32 */ |
| 4990 |
if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
| 4991 |
(tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) > |
| 4992 |
EXT4_MAX_BLOCK_FILE_PHYS)) { |
| 4993 |
/* |
| 4994 |
* Since PAs don't overlap, we won't find any other PA to |
| 4995 |
* satisfy this. |
| 4996 |
*/ |
| 4997 |
spin_unlock(&tmp_pa->pa_lock); |
| 4998 |
goto try_group_pa; |
| 4999 |
} |
| 5000 |
|
| 5001 |
if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) { |
| 5002 |
atomic_inc(&tmp_pa->pa_count); |
| 5003 |
ext4_mb_use_inode_pa(ac, tmp_pa); |
| 5004 |
spin_unlock(&tmp_pa->pa_lock); |
| 5005 |
read_unlock(&ei->i_prealloc_lock); |
| 5006 |
return true; |
| 5007 |
} else { |
| 5008 |
/* |
| 5009 |
* We found a valid overlapping pa but couldn't use it because |
| 5010 |
* it had no free blocks. This should ideally never happen |
| 5011 |
* because: |
| 5012 |
* |
| 5013 |
* 1. When a new inode pa is added to rbtree it must have |
| 5014 |
* pa_free > 0 since otherwise we won't actually need |
| 5015 |
* preallocation. |
| 5016 |
* |
| 5017 |
* 2. An inode pa that is in the rbtree can only have it's |
| 5018 |
* pa_free become zero when another thread calls: |
| 5019 |
* ext4_mb_new_blocks |
| 5020 |
* ext4_mb_use_preallocated |
| 5021 |
* ext4_mb_use_inode_pa |
| 5022 |
* |
| 5023 |
* 3. Further, after the above calls make pa_free == 0, we will |
| 5024 |
* immediately remove it from the rbtree in: |
| 5025 |
* ext4_mb_new_blocks |
| 5026 |
* ext4_mb_release_context |
| 5027 |
* ext4_mb_put_pa |
| 5028 |
* |
| 5029 |
* 4. Since the pa_free becoming 0 and pa_free getting removed |
| 5030 |
* from tree both happen in ext4_mb_new_blocks, which is always |
| 5031 |
* called with i_data_sem held for data allocations, we can be |
| 5032 |
* sure that another process will never see a pa in rbtree with |
| 5033 |
* pa_free == 0. |
| 5034 |
*/ |
| 5035 |
WARN_ON_ONCE(tmp_pa->pa_free == 0); |
| 5036 |
} |
| 5037 |
spin_unlock(&tmp_pa->pa_lock); |
| 5038 |
try_group_pa: |
| 5039 |
read_unlock(&ei->i_prealloc_lock); |
| 5040 |
|
| 5041 |
/* can we use group allocation? */ |
| 5042 |
if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) |
| 5043 |
return false; |
| 5044 |
|
| 5045 |
/* inode may have no locality group for some reason */ |
| 5046 |
lg = ac->ac_lg; |
| 5047 |
if (lg == NULL) |
| 5048 |
return false; |
| 5049 |
order = fls(ac->ac_o_ex.fe_len) - 1; |
| 5050 |
if (order > PREALLOC_TB_SIZE - 1) |
| 5051 |
/* The max size of hash table is PREALLOC_TB_SIZE */ |
| 5052 |
order = PREALLOC_TB_SIZE - 1; |
| 5053 |
|
| 5054 |
goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
| 5055 |
/* |
| 5056 |
* search for the prealloc space that is having |
| 5057 |
* minimal distance from the goal block. |
| 5058 |
*/ |
| 5059 |
for (i = order; i < PREALLOC_TB_SIZE; i++) { |
| 5060 |
rcu_read_lock(); |
| 5061 |
list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i], |
| 5062 |
pa_node.lg_list) { |
| 5063 |
spin_lock(&tmp_pa->pa_lock); |
| 5064 |
if (tmp_pa->pa_deleted == 0 && |
| 5065 |
tmp_pa->pa_free >= ac->ac_o_ex.fe_len) { |
| 5066 |
|
| 5067 |
cpa = ext4_mb_check_group_pa(goal_block, |
| 5068 |
tmp_pa, cpa); |
| 5069 |
} |
| 5070 |
spin_unlock(&tmp_pa->pa_lock); |
| 5071 |
} |
| 5072 |
rcu_read_unlock(); |
| 5073 |
} |
| 5074 |
if (cpa) { |
| 5075 |
ext4_mb_use_group_pa(ac, cpa); |
| 5076 |
return true; |
| 5077 |
} |
| 5078 |
return false; |
| 5079 |
} |
| 5080 |
|
| 5081 |
/* |
| 5082 |
* the function goes through all preallocation in this group and marks them |
| 5083 |
* used in in-core bitmap. buddy must be generated from this bitmap |
| 5084 |
* Need to be called with ext4 group lock held |
| 5085 |
*/ |
| 5086 |
static noinline_for_stack |
| 5087 |
void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
| 5088 |
ext4_group_t group) |
| 5089 |
{ |
| 5090 |
struct ext4_group_info *grp = ext4_get_group_info(sb, group); |
| 5091 |
struct ext4_prealloc_space *pa; |
| 5092 |
struct list_head *cur; |
| 5093 |
ext4_group_t groupnr; |
| 5094 |
ext4_grpblk_t start; |
| 5095 |
int preallocated = 0; |
| 5096 |
int len; |
| 5097 |
|
| 5098 |
if (!grp) |
| 5099 |
return; |
| 5100 |
|
| 5101 |
/* all form of preallocation discards first load group, |
| 5102 |
* so the only competing code is preallocation use. |
| 5103 |
* we don't need any locking here |
| 5104 |
* notice we do NOT ignore preallocations with pa_deleted |
| 5105 |
* otherwise we could leave used blocks available for |
| 5106 |
* allocation in buddy when concurrent ext4_mb_put_pa() |
| 5107 |
* is dropping preallocation |
| 5108 |
*/ |
| 5109 |
list_for_each(cur, &grp->bb_prealloc_list) { |
| 5110 |
pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
| 5111 |
spin_lock(&pa->pa_lock); |
| 5112 |
ext4_get_group_no_and_offset(sb, pa->pa_pstart, |
| 5113 |
&groupnr, &start); |
| 5114 |
len = pa->pa_len; |
| 5115 |
spin_unlock(&pa->pa_lock); |
| 5116 |
if (unlikely(len == 0)) |
| 5117 |
continue; |
| 5118 |
BUG_ON(groupnr != group); |
| 5119 |
mb_set_bits(bitmap, start, len); |
| 5120 |
preallocated += len; |
| 5121 |
} |
| 5122 |
mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); |
| 5123 |
} |
| 5124 |
|
| 5125 |
static void ext4_mb_mark_pa_deleted(struct super_block *sb, |
| 5126 |
struct ext4_prealloc_space *pa) |
| 5127 |
{ |
| 5128 |
struct ext4_inode_info *ei; |
| 5129 |
|
| 5130 |
if (pa->pa_deleted) { |
| 5131 |
ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", |
| 5132 |
pa->pa_type, pa->pa_pstart, pa->pa_lstart, |
| 5133 |
pa->pa_len); |
| 5134 |
return; |
| 5135 |
} |
| 5136 |
|
| 5137 |
pa->pa_deleted = 1; |
| 5138 |
|
| 5139 |
if (pa->pa_type == MB_INODE_PA) { |
| 5140 |
ei = EXT4_I(pa->pa_inode); |
| 5141 |
atomic_dec(&ei->i_prealloc_active); |
| 5142 |
} |
| 5143 |
} |
| 5144 |
|
| 5145 |
static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa) |
| 5146 |
{ |
| 5147 |
BUG_ON(!pa); |
| 5148 |
BUG_ON(atomic_read(&pa->pa_count)); |
| 5149 |
BUG_ON(pa->pa_deleted == 0); |
| 5150 |
kmem_cache_free(ext4_pspace_cachep, pa); |
| 5151 |
} |
| 5152 |
|
| 5153 |
static void ext4_mb_pa_callback(struct rcu_head *head) |
| 5154 |
{ |
| 5155 |
struct ext4_prealloc_space *pa; |
| 5156 |
|
| 5157 |
pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); |
| 5158 |
ext4_mb_pa_free(pa); |
| 5159 |
} |
| 5160 |
|
| 5161 |
/* |
| 5162 |
* drops a reference to preallocated space descriptor |
| 5163 |
* if this was the last reference and the space is consumed |
| 5164 |
*/ |
| 5165 |
static void ext4_mb_put_pa(struct ext4_allocation_context *ac, |
| 5166 |
struct super_block *sb, struct ext4_prealloc_space *pa) |
| 5167 |
{ |
| 5168 |
ext4_group_t grp; |
| 5169 |
ext4_fsblk_t grp_blk; |
| 5170 |
struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
| 5171 |
|
| 5172 |
/* in this short window concurrent discard can set pa_deleted */ |
| 5173 |
spin_lock(&pa->pa_lock); |
| 5174 |
if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
| 5175 |
spin_unlock(&pa->pa_lock); |
| 5176 |
return; |
| 5177 |
} |
| 5178 |
|
| 5179 |
if (pa->pa_deleted == 1) { |
| 5180 |
spin_unlock(&pa->pa_lock); |
| 5181 |
return; |
| 5182 |
} |
| 5183 |
|
| 5184 |
ext4_mb_mark_pa_deleted(sb, pa); |
| 5185 |
spin_unlock(&pa->pa_lock); |
| 5186 |
|
| 5187 |
grp_blk = pa->pa_pstart; |
| 5188 |
/* |
| 5189 |
* If doing group-based preallocation, pa_pstart may be in the |
| 5190 |
* next group when pa is used up |
| 5191 |
*/ |
| 5192 |
if (pa->pa_type == MB_GROUP_PA) |
| 5193 |
grp_blk--; |
| 5194 |
|
| 5195 |
grp = ext4_get_group_number(sb, grp_blk); |
| 5196 |
|
| 5197 |
/* |
| 5198 |
* possible race: |
| 5199 |
* |
| 5200 |
* P1 (buddy init) P2 (regular allocation) |
| 5201 |
* find block B in PA |
| 5202 |
* copy on-disk bitmap to buddy |
| 5203 |
* mark B in on-disk bitmap |
| 5204 |
* drop PA from group |
| 5205 |
* mark all PAs in buddy |
| 5206 |
* |
| 5207 |
* thus, P1 initializes buddy with B available. to prevent this |
| 5208 |
* we make "copy" and "mark all PAs" atomic and serialize "drop PA" |
| 5209 |
* against that pair |
| 5210 |
*/ |
| 5211 |
ext4_lock_group(sb, grp); |
| 5212 |
list_del(&pa->pa_group_list); |
| 5213 |
ext4_unlock_group(sb, grp); |
| 5214 |
|
| 5215 |
if (pa->pa_type == MB_INODE_PA) { |
| 5216 |
write_lock(pa->pa_node_lock.inode_lock); |
| 5217 |
rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); |
| 5218 |
write_unlock(pa->pa_node_lock.inode_lock); |
| 5219 |
ext4_mb_pa_free(pa); |
| 5220 |
} else { |
| 5221 |
spin_lock(pa->pa_node_lock.lg_lock); |
| 5222 |
list_del_rcu(&pa->pa_node.lg_list); |
| 5223 |
spin_unlock(pa->pa_node_lock.lg_lock); |
| 5224 |
call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
| 5225 |
} |
| 5226 |
} |
| 5227 |
|
| 5228 |
static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new) |
| 5229 |
{ |
| 5230 |
struct rb_node **iter = &root->rb_node, *parent = NULL; |
| 5231 |
struct ext4_prealloc_space *iter_pa, *new_pa; |
| 5232 |
ext4_lblk_t iter_start, new_start; |
| 5233 |
|
| 5234 |
while (*iter) { |
| 5235 |
iter_pa = rb_entry(*iter, struct ext4_prealloc_space, |
| 5236 |
pa_node.inode_node); |
| 5237 |
new_pa = rb_entry(new, struct ext4_prealloc_space, |
| 5238 |
pa_node.inode_node); |
| 5239 |
iter_start = iter_pa->pa_lstart; |
| 5240 |
new_start = new_pa->pa_lstart; |
| 5241 |
|
| 5242 |
parent = *iter; |
| 5243 |
if (new_start < iter_start) |
| 5244 |
iter = &((*iter)->rb_left); |
| 5245 |
else |
| 5246 |
iter = &((*iter)->rb_right); |
| 5247 |
} |
| 5248 |
|
| 5249 |
rb_link_node(new, parent, iter); |
| 5250 |
rb_insert_color(new, root); |
| 5251 |
} |
| 5252 |
|
| 5253 |
/* |
| 5254 |
* creates new preallocated space for given inode |
| 5255 |
*/ |
| 5256 |
static noinline_for_stack void |
| 5257 |
ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) |
| 5258 |
{ |
| 5259 |
struct super_block *sb = ac->ac_sb; |
| 5260 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 5261 |
struct ext4_prealloc_space *pa; |
| 5262 |
struct ext4_group_info *grp; |
| 5263 |
struct ext4_inode_info *ei; |
| 5264 |
|
| 5265 |
/* preallocate only when found space is larger then requested */ |
| 5266 |
BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); |
| 5267 |
BUG_ON(ac->ac_status != AC_STATUS_FOUND); |
| 5268 |
BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); |
| 5269 |
BUG_ON(ac->ac_pa == NULL); |
| 5270 |
|
| 5271 |
pa = ac->ac_pa; |
| 5272 |
|
| 5273 |
if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) { |
| 5274 |
struct ext4_free_extent ex = { |
| 5275 |
.fe_logical = ac->ac_g_ex.fe_logical, |
| 5276 |
.fe_len = ac->ac_orig_goal_len, |
| 5277 |
}; |
| 5278 |
loff_t orig_goal_end = extent_logical_end(sbi, &ex); |
| 5279 |
loff_t o_ex_end = extent_logical_end(sbi, &ac->ac_o_ex); |
| 5280 |
|
| 5281 |
/* |
| 5282 |
* We can't allocate as much as normalizer wants, so we try |
| 5283 |
* to get proper lstart to cover the original request, except |
| 5284 |
* when the goal doesn't cover the original request as below: |
| 5285 |
* |
| 5286 |
* orig_ex:2045/2055(10), isize:8417280 -> normalized:0/2048 |
| 5287 |
* best_ex:0/200(200) -> adjusted: 1848/2048(200) |
| 5288 |
*/ |
| 5289 |
BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); |
| 5290 |
BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); |
| 5291 |
|
| 5292 |
/* |
| 5293 |
* Use the below logic for adjusting best extent as it keeps |
| 5294 |
* fragmentation in check while ensuring logical range of best |
| 5295 |
* extent doesn't overflow out of goal extent: |
| 5296 |
* |
| 5297 |
* 1. Check if best ex can be kept at end of goal (before |
| 5298 |
* cr_best_avail trimmed it) and still cover original start |
| 5299 |
* 2. Else, check if best ex can be kept at start of goal and |
| 5300 |
* still cover original end |
| 5301 |
* 3. Else, keep the best ex at start of original request. |
| 5302 |
*/ |
| 5303 |
ex.fe_len = ac->ac_b_ex.fe_len; |
| 5304 |
|
| 5305 |
ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len); |
| 5306 |
if (ac->ac_o_ex.fe_logical >= ex.fe_logical) |
| 5307 |
goto adjust_bex; |
| 5308 |
|
| 5309 |
ex.fe_logical = ac->ac_g_ex.fe_logical; |
| 5310 |
if (o_ex_end <= extent_logical_end(sbi, &ex)) |
| 5311 |
goto adjust_bex; |
| 5312 |
|
| 5313 |
ex.fe_logical = ac->ac_o_ex.fe_logical; |
| 5314 |
adjust_bex: |
| 5315 |
ac->ac_b_ex.fe_logical = ex.fe_logical; |
| 5316 |
|
| 5317 |
BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
| 5318 |
BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end); |
| 5319 |
} |
| 5320 |
|
| 5321 |
pa->pa_lstart = ac->ac_b_ex.fe_logical; |
| 5322 |
pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
| 5323 |
pa->pa_len = ac->ac_b_ex.fe_len; |
| 5324 |
pa->pa_free = pa->pa_len; |
| 5325 |
spin_lock_init(&pa->pa_lock); |
| 5326 |
INIT_LIST_HEAD(&pa->pa_group_list); |
| 5327 |
pa->pa_deleted = 0; |
| 5328 |
pa->pa_type = MB_INODE_PA; |
| 5329 |
|
| 5330 |
mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
| 5331 |
pa->pa_len, pa->pa_lstart); |
| 5332 |
trace_ext4_mb_new_inode_pa(ac, pa); |
| 5333 |
|
| 5334 |
atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
| 5335 |
ext4_mb_use_inode_pa(ac, pa); |
| 5336 |
|
| 5337 |
ei = EXT4_I(ac->ac_inode); |
| 5338 |
grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); |
| 5339 |
if (!grp) |
| 5340 |
return; |
| 5341 |
|
| 5342 |
pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock; |
| 5343 |
pa->pa_inode = ac->ac_inode; |
| 5344 |
|
| 5345 |
list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
| 5346 |
|
| 5347 |
write_lock(pa->pa_node_lock.inode_lock); |
| 5348 |
ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node); |
| 5349 |
write_unlock(pa->pa_node_lock.inode_lock); |
| 5350 |
atomic_inc(&ei->i_prealloc_active); |
| 5351 |
} |
| 5352 |
|
| 5353 |
/* |
| 5354 |
* creates new preallocated space for locality group inodes belongs to |
| 5355 |
*/ |
| 5356 |
static noinline_for_stack void |
| 5357 |
ext4_mb_new_group_pa(struct ext4_allocation_context *ac) |
| 5358 |
{ |
| 5359 |
struct super_block *sb = ac->ac_sb; |
| 5360 |
struct ext4_locality_group *lg; |
| 5361 |
struct ext4_prealloc_space *pa; |
| 5362 |
struct ext4_group_info *grp; |
| 5363 |
|
| 5364 |
/* preallocate only when found space is larger then requested */ |
| 5365 |
BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); |
| 5366 |
BUG_ON(ac->ac_status != AC_STATUS_FOUND); |
| 5367 |
BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); |
| 5368 |
BUG_ON(ac->ac_pa == NULL); |
| 5369 |
|
| 5370 |
pa = ac->ac_pa; |
| 5371 |
|
| 5372 |
pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
| 5373 |
pa->pa_lstart = pa->pa_pstart; |
| 5374 |
pa->pa_len = ac->ac_b_ex.fe_len; |
| 5375 |
pa->pa_free = pa->pa_len; |
| 5376 |
spin_lock_init(&pa->pa_lock); |
| 5377 |
INIT_LIST_HEAD(&pa->pa_node.lg_list); |
| 5378 |
INIT_LIST_HEAD(&pa->pa_group_list); |
| 5379 |
pa->pa_deleted = 0; |
| 5380 |
pa->pa_type = MB_GROUP_PA; |
| 5381 |
|
| 5382 |
mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, |
| 5383 |
pa->pa_len, pa->pa_lstart); |
| 5384 |
trace_ext4_mb_new_group_pa(ac, pa); |
| 5385 |
|
| 5386 |
ext4_mb_use_group_pa(ac, pa); |
| 5387 |
atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); |
| 5388 |
|
| 5389 |
grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); |
| 5390 |
if (!grp) |
| 5391 |
return; |
| 5392 |
lg = ac->ac_lg; |
| 5393 |
BUG_ON(lg == NULL); |
| 5394 |
|
| 5395 |
pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock; |
| 5396 |
pa->pa_inode = NULL; |
| 5397 |
|
| 5398 |
list_add(&pa->pa_group_list, &grp->bb_prealloc_list); |
| 5399 |
|
| 5400 |
/* |
| 5401 |
* We will later add the new pa to the right bucket |
| 5402 |
* after updating the pa_free in ext4_mb_release_context |
| 5403 |
*/ |
| 5404 |
} |
| 5405 |
|
| 5406 |
static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) |
| 5407 |
{ |
| 5408 |
if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
| 5409 |
ext4_mb_new_group_pa(ac); |
| 5410 |
else |
| 5411 |
ext4_mb_new_inode_pa(ac); |
| 5412 |
} |
| 5413 |
|
| 5414 |
/* |
| 5415 |
* finds all unused blocks in on-disk bitmap, frees them in |
| 5416 |
* in-core bitmap and buddy. |
| 5417 |
* @pa must be unlinked from inode and group lists, so that |
| 5418 |
* nobody else can find/use it. |
| 5419 |
* the caller MUST hold group/inode locks. |
| 5420 |
* TODO: optimize the case when there are no in-core structures yet |
| 5421 |
*/ |
| 5422 |
static noinline_for_stack void |
| 5423 |
ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, |
| 5424 |
struct ext4_prealloc_space *pa) |
| 5425 |
{ |
| 5426 |
struct super_block *sb = e4b->bd_sb; |
| 5427 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 5428 |
unsigned int end; |
| 5429 |
unsigned int next; |
| 5430 |
ext4_group_t group; |
| 5431 |
ext4_grpblk_t bit; |
| 5432 |
unsigned long long grp_blk_start; |
| 5433 |
int free = 0; |
| 5434 |
|
| 5435 |
BUG_ON(pa->pa_deleted == 0); |
| 5436 |
ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); |
| 5437 |
grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
| 5438 |
BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
| 5439 |
end = bit + pa->pa_len; |
| 5440 |
|
| 5441 |
while (bit < end) { |
| 5442 |
bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
| 5443 |
if (bit >= end) |
| 5444 |
break; |
| 5445 |
next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
| 5446 |
mb_debug(sb, "free preallocated %u/%u in group %u\n", |
| 5447 |
(unsigned) ext4_group_first_block_no(sb, group) + bit, |
| 5448 |
(unsigned) next - bit, (unsigned) group); |
| 5449 |
free += next - bit; |
| 5450 |
|
| 5451 |
trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
| 5452 |
trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
| 5453 |
EXT4_C2B(sbi, bit)), |
| 5454 |
next - bit); |
| 5455 |
mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
| 5456 |
bit = next + 1; |
| 5457 |
} |
| 5458 |
if (free != pa->pa_free) { |
| 5459 |
ext4_msg(e4b->bd_sb, KERN_CRIT, |
| 5460 |
"pa %p: logic %lu, phys. %lu, len %d", |
| 5461 |
pa, (unsigned long) pa->pa_lstart, |
| 5462 |
(unsigned long) pa->pa_pstart, |
| 5463 |
pa->pa_len); |
| 5464 |
ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
| 5465 |
free, pa->pa_free); |
| 5466 |
/* |
| 5467 |
* pa is already deleted so we use the value obtained |
| 5468 |
* from the bitmap and continue. |
| 5469 |
*/ |
| 5470 |
} |
| 5471 |
atomic_add(free, &sbi->s_mb_discarded); |
| 5472 |
} |
| 5473 |
|
| 5474 |
static noinline_for_stack void |
| 5475 |
ext4_mb_release_group_pa(struct ext4_buddy *e4b, |
| 5476 |
struct ext4_prealloc_space *pa) |
| 5477 |
{ |
| 5478 |
struct super_block *sb = e4b->bd_sb; |
| 5479 |
ext4_group_t group; |
| 5480 |
ext4_grpblk_t bit; |
| 5481 |
|
| 5482 |
trace_ext4_mb_release_group_pa(sb, pa); |
| 5483 |
BUG_ON(pa->pa_deleted == 0); |
| 5484 |
ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); |
| 5485 |
if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) { |
| 5486 |
ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu", |
| 5487 |
e4b->bd_group, group, pa->pa_pstart); |
| 5488 |
return; |
| 5489 |
} |
| 5490 |
mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); |
| 5491 |
atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); |
| 5492 |
trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
| 5493 |
} |
| 5494 |
|
| 5495 |
/* |
| 5496 |
* releases all preallocations in given group |
| 5497 |
* |
| 5498 |
* first, we need to decide discard policy: |
| 5499 |
* - when do we discard |
| 5500 |
* 1) ENOSPC |
| 5501 |
* - how many do we discard |
| 5502 |
* 1) how many requested |
| 5503 |
*/ |
| 5504 |
static noinline_for_stack int |
| 5505 |
ext4_mb_discard_group_preallocations(struct super_block *sb, |
| 5506 |
ext4_group_t group, int *busy) |
| 5507 |
{ |
| 5508 |
struct ext4_group_info *grp = ext4_get_group_info(sb, group); |
| 5509 |
struct buffer_head *bitmap_bh = NULL; |
| 5510 |
struct ext4_prealloc_space *pa, *tmp; |
| 5511 |
LIST_HEAD(list); |
| 5512 |
struct ext4_buddy e4b; |
| 5513 |
struct ext4_inode_info *ei; |
| 5514 |
int err; |
| 5515 |
int free = 0; |
| 5516 |
|
| 5517 |
if (!grp) |
| 5518 |
return 0; |
| 5519 |
mb_debug(sb, "discard preallocation for group %u\n", group); |
| 5520 |
if (list_empty(&grp->bb_prealloc_list)) |
| 5521 |
goto out_dbg; |
| 5522 |
|
| 5523 |
bitmap_bh = ext4_read_block_bitmap(sb, group); |
| 5524 |
if (IS_ERR(bitmap_bh)) { |
| 5525 |
err = PTR_ERR(bitmap_bh); |
| 5526 |
ext4_error_err(sb, -err, |
| 5527 |
"Error %d reading block bitmap for %u", |
| 5528 |
err, group); |
| 5529 |
goto out_dbg; |
| 5530 |
} |
| 5531 |
|
| 5532 |
err = ext4_mb_load_buddy(sb, group, &e4b); |
| 5533 |
if (err) { |
| 5534 |
ext4_warning(sb, "Error %d loading buddy information for %u", |
| 5535 |
err, group); |
| 5536 |
put_bh(bitmap_bh); |
| 5537 |
goto out_dbg; |
| 5538 |
} |
| 5539 |
|
| 5540 |
ext4_lock_group(sb, group); |
| 5541 |
list_for_each_entry_safe(pa, tmp, |
| 5542 |
&grp->bb_prealloc_list, pa_group_list) { |
| 5543 |
spin_lock(&pa->pa_lock); |
| 5544 |
if (atomic_read(&pa->pa_count)) { |
| 5545 |
spin_unlock(&pa->pa_lock); |
| 5546 |
*busy = 1; |
| 5547 |
continue; |
| 5548 |
} |
| 5549 |
if (pa->pa_deleted) { |
| 5550 |
spin_unlock(&pa->pa_lock); |
| 5551 |
continue; |
| 5552 |
} |
| 5553 |
|
| 5554 |
/* seems this one can be freed ... */ |
| 5555 |
ext4_mb_mark_pa_deleted(sb, pa); |
| 5556 |
|
| 5557 |
if (!free) |
| 5558 |
this_cpu_inc(discard_pa_seq); |
| 5559 |
|
| 5560 |
/* we can trust pa_free ... */ |
| 5561 |
free += pa->pa_free; |
| 5562 |
|
| 5563 |
spin_unlock(&pa->pa_lock); |
| 5564 |
|
| 5565 |
list_del(&pa->pa_group_list); |
| 5566 |
list_add(&pa->u.pa_tmp_list, &list); |
| 5567 |
} |
| 5568 |
|
| 5569 |
/* now free all selected PAs */ |
| 5570 |
list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { |
| 5571 |
|
| 5572 |
/* remove from object (inode or locality group) */ |
| 5573 |
if (pa->pa_type == MB_GROUP_PA) { |
| 5574 |
spin_lock(pa->pa_node_lock.lg_lock); |
| 5575 |
list_del_rcu(&pa->pa_node.lg_list); |
| 5576 |
spin_unlock(pa->pa_node_lock.lg_lock); |
| 5577 |
} else { |
| 5578 |
write_lock(pa->pa_node_lock.inode_lock); |
| 5579 |
ei = EXT4_I(pa->pa_inode); |
| 5580 |
rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); |
| 5581 |
write_unlock(pa->pa_node_lock.inode_lock); |
| 5582 |
} |
| 5583 |
|
| 5584 |
list_del(&pa->u.pa_tmp_list); |
| 5585 |
|
| 5586 |
if (pa->pa_type == MB_GROUP_PA) { |
| 5587 |
ext4_mb_release_group_pa(&e4b, pa); |
| 5588 |
call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
| 5589 |
} else { |
| 5590 |
ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
| 5591 |
ext4_mb_pa_free(pa); |
| 5592 |
} |
| 5593 |
} |
| 5594 |
|
| 5595 |
ext4_unlock_group(sb, group); |
| 5596 |
ext4_mb_unload_buddy(&e4b); |
| 5597 |
put_bh(bitmap_bh); |
| 5598 |
out_dbg: |
| 5599 |
mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", |
| 5600 |
free, group, grp->bb_free); |
| 5601 |
return free; |
| 5602 |
} |
| 5603 |
|
| 5604 |
/* |
| 5605 |
* releases all non-used preallocated blocks for given inode |
| 5606 |
* |
| 5607 |
* It's important to discard preallocations under i_data_sem |
| 5608 |
* We don't want another block to be served from the prealloc |
| 5609 |
* space when we are discarding the inode prealloc space. |
| 5610 |
* |
| 5611 |
* FIXME!! Make sure it is valid at all the call sites |
| 5612 |
*/ |
| 5613 |
void ext4_discard_preallocations(struct inode *inode) |
| 5614 |
{ |
| 5615 |
struct ext4_inode_info *ei = EXT4_I(inode); |
| 5616 |
struct super_block *sb = inode->i_sb; |
| 5617 |
struct buffer_head *bitmap_bh = NULL; |
| 5618 |
struct ext4_prealloc_space *pa, *tmp; |
| 5619 |
ext4_group_t group = 0; |
| 5620 |
LIST_HEAD(list); |
| 5621 |
struct ext4_buddy e4b; |
| 5622 |
struct rb_node *iter; |
| 5623 |
int err; |
| 5624 |
|
| 5625 |
if (!S_ISREG(inode->i_mode)) |
| 5626 |
return; |
| 5627 |
|
| 5628 |
if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) |
| 5629 |
return; |
| 5630 |
|
| 5631 |
mb_debug(sb, "discard preallocation for inode %lu\n", |
| 5632 |
inode->i_ino); |
| 5633 |
trace_ext4_discard_preallocations(inode, |
| 5634 |
atomic_read(&ei->i_prealloc_active)); |
| 5635 |
|
| 5636 |
repeat: |
| 5637 |
/* first, collect all pa's in the inode */ |
| 5638 |
write_lock(&ei->i_prealloc_lock); |
| 5639 |
for (iter = rb_first(&ei->i_prealloc_node); iter; |
| 5640 |
iter = rb_next(iter)) { |
| 5641 |
pa = rb_entry(iter, struct ext4_prealloc_space, |
| 5642 |
pa_node.inode_node); |
| 5643 |
BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock); |
| 5644 |
|
| 5645 |
spin_lock(&pa->pa_lock); |
| 5646 |
if (atomic_read(&pa->pa_count)) { |
| 5647 |
/* this shouldn't happen often - nobody should |
| 5648 |
* use preallocation while we're discarding it */ |
| 5649 |
spin_unlock(&pa->pa_lock); |
| 5650 |
write_unlock(&ei->i_prealloc_lock); |
| 5651 |
ext4_msg(sb, KERN_ERR, |
| 5652 |
"uh-oh! used pa while discarding"); |
| 5653 |
WARN_ON(1); |
| 5654 |
schedule_timeout_uninterruptible(HZ); |
| 5655 |
goto repeat; |
| 5656 |
|
| 5657 |
} |
| 5658 |
if (pa->pa_deleted == 0) { |
| 5659 |
ext4_mb_mark_pa_deleted(sb, pa); |
| 5660 |
spin_unlock(&pa->pa_lock); |
| 5661 |
rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); |
| 5662 |
list_add(&pa->u.pa_tmp_list, &list); |
| 5663 |
continue; |
| 5664 |
} |
| 5665 |
|
| 5666 |
/* someone is deleting pa right now */ |
| 5667 |
spin_unlock(&pa->pa_lock); |
| 5668 |
write_unlock(&ei->i_prealloc_lock); |
| 5669 |
|
| 5670 |
/* we have to wait here because pa_deleted |
| 5671 |
* doesn't mean pa is already unlinked from |
| 5672 |
* the list. as we might be called from |
| 5673 |
* ->clear_inode() the inode will get freed |
| 5674 |
* and concurrent thread which is unlinking |
| 5675 |
* pa from inode's list may access already |
| 5676 |
* freed memory, bad-bad-bad */ |
| 5677 |
|
| 5678 |
/* XXX: if this happens too often, we can |
| 5679 |
* add a flag to force wait only in case |
| 5680 |
* of ->clear_inode(), but not in case of |
| 5681 |
* regular truncate */ |
| 5682 |
schedule_timeout_uninterruptible(HZ); |
| 5683 |
goto repeat; |
| 5684 |
} |
| 5685 |
write_unlock(&ei->i_prealloc_lock); |
| 5686 |
|
| 5687 |
list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { |
| 5688 |
BUG_ON(pa->pa_type != MB_INODE_PA); |
| 5689 |
group = ext4_get_group_number(sb, pa->pa_pstart); |
| 5690 |
|
| 5691 |
err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
| 5692 |
GFP_NOFS|__GFP_NOFAIL); |
| 5693 |
if (err) { |
| 5694 |
ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
| 5695 |
err, group); |
| 5696 |
continue; |
| 5697 |
} |
| 5698 |
|
| 5699 |
bitmap_bh = ext4_read_block_bitmap(sb, group); |
| 5700 |
if (IS_ERR(bitmap_bh)) { |
| 5701 |
err = PTR_ERR(bitmap_bh); |
| 5702 |
ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", |
| 5703 |
err, group); |
| 5704 |
ext4_mb_unload_buddy(&e4b); |
| 5705 |
continue; |
| 5706 |
} |
| 5707 |
|
| 5708 |
ext4_lock_group(sb, group); |
| 5709 |
list_del(&pa->pa_group_list); |
| 5710 |
ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
| 5711 |
ext4_unlock_group(sb, group); |
| 5712 |
|
| 5713 |
ext4_mb_unload_buddy(&e4b); |
| 5714 |
put_bh(bitmap_bh); |
| 5715 |
|
| 5716 |
list_del(&pa->u.pa_tmp_list); |
| 5717 |
ext4_mb_pa_free(pa); |
| 5718 |
} |
| 5719 |
} |
| 5720 |
|
| 5721 |
static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) |
| 5722 |
{ |
| 5723 |
struct ext4_prealloc_space *pa; |
| 5724 |
|
| 5725 |
BUG_ON(ext4_pspace_cachep == NULL); |
| 5726 |
pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); |
| 5727 |
if (!pa) |
| 5728 |
return -ENOMEM; |
| 5729 |
atomic_set(&pa->pa_count, 1); |
| 5730 |
ac->ac_pa = pa; |
| 5731 |
return 0; |
| 5732 |
} |
| 5733 |
|
| 5734 |
static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac) |
| 5735 |
{ |
| 5736 |
struct ext4_prealloc_space *pa = ac->ac_pa; |
| 5737 |
|
| 5738 |
BUG_ON(!pa); |
| 5739 |
ac->ac_pa = NULL; |
| 5740 |
WARN_ON(!atomic_dec_and_test(&pa->pa_count)); |
| 5741 |
/* |
| 5742 |
* current function is only called due to an error or due to |
| 5743 |
* len of found blocks < len of requested blocks hence the PA has not |
| 5744 |
* been added to grp->bb_prealloc_list. So we don't need to lock it |
| 5745 |
*/ |
| 5746 |
pa->pa_deleted = 1; |
| 5747 |
ext4_mb_pa_free(pa); |
| 5748 |
} |
| 5749 |
|
| 5750 |
#ifdef CONFIG_EXT4_DEBUG |
| 5751 |
static inline void ext4_mb_show_pa(struct super_block *sb) |
| 5752 |
{ |
| 5753 |
ext4_group_t i, ngroups; |
| 5754 |
|
| 5755 |
if (ext4_emergency_state(sb)) |
| 5756 |
return; |
| 5757 |
|
| 5758 |
ngroups = ext4_get_groups_count(sb); |
| 5759 |
mb_debug(sb, "groups: "); |
| 5760 |
for (i = 0; i < ngroups; i++) { |
| 5761 |
struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
| 5762 |
struct ext4_prealloc_space *pa; |
| 5763 |
ext4_grpblk_t start; |
| 5764 |
struct list_head *cur; |
| 5765 |
|
| 5766 |
if (!grp) |
| 5767 |
continue; |
| 5768 |
ext4_lock_group(sb, i); |
| 5769 |
list_for_each(cur, &grp->bb_prealloc_list) { |
| 5770 |
pa = list_entry(cur, struct ext4_prealloc_space, |
| 5771 |
pa_group_list); |
| 5772 |
spin_lock(&pa->pa_lock); |
| 5773 |
ext4_get_group_no_and_offset(sb, pa->pa_pstart, |
| 5774 |
NULL, &start); |
| 5775 |
spin_unlock(&pa->pa_lock); |
| 5776 |
mb_debug(sb, "PA:%u:%d:%d\n", i, start, |
| 5777 |
pa->pa_len); |
| 5778 |
} |
| 5779 |
ext4_unlock_group(sb, i); |
| 5780 |
mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, |
| 5781 |
grp->bb_fragments); |
| 5782 |
} |
| 5783 |
} |
| 5784 |
|
| 5785 |
static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
| 5786 |
{ |
| 5787 |
struct super_block *sb = ac->ac_sb; |
| 5788 |
|
| 5789 |
if (ext4_emergency_state(sb)) |
| 5790 |
return; |
| 5791 |
|
| 5792 |
mb_debug(sb, "Can't allocate:" |
| 5793 |
" Allocation context details:"); |
| 5794 |
mb_debug(sb, "status %u flags 0x%x", |
| 5795 |
ac->ac_status, ac->ac_flags); |
| 5796 |
mb_debug(sb, "orig %lu/%lu/%lu@%lu, " |
| 5797 |
"goal %lu/%lu/%lu@%lu, " |
| 5798 |
"best %lu/%lu/%lu@%lu cr %d", |
| 5799 |
(unsigned long)ac->ac_o_ex.fe_group, |
| 5800 |
(unsigned long)ac->ac_o_ex.fe_start, |
| 5801 |
(unsigned long)ac->ac_o_ex.fe_len, |
| 5802 |
(unsigned long)ac->ac_o_ex.fe_logical, |
| 5803 |
(unsigned long)ac->ac_g_ex.fe_group, |
| 5804 |
(unsigned long)ac->ac_g_ex.fe_start, |
| 5805 |
(unsigned long)ac->ac_g_ex.fe_len, |
| 5806 |
(unsigned long)ac->ac_g_ex.fe_logical, |
| 5807 |
(unsigned long)ac->ac_b_ex.fe_group, |
| 5808 |
(unsigned long)ac->ac_b_ex.fe_start, |
| 5809 |
(unsigned long)ac->ac_b_ex.fe_len, |
| 5810 |
(unsigned long)ac->ac_b_ex.fe_logical, |
| 5811 |
(int)ac->ac_criteria); |
| 5812 |
mb_debug(sb, "%u found", ac->ac_found); |
| 5813 |
mb_debug(sb, "used pa: %s, ", str_yes_no(ac->ac_pa)); |
| 5814 |
if (ac->ac_pa) |
| 5815 |
mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ? |
| 5816 |
"group pa" : "inode pa"); |
| 5817 |
ext4_mb_show_pa(sb); |
| 5818 |
} |
| 5819 |
#else |
| 5820 |
static inline void ext4_mb_show_pa(struct super_block *sb) |
| 5821 |
{ |
| 5822 |
} |
| 5823 |
static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
| 5824 |
{ |
| 5825 |
ext4_mb_show_pa(ac->ac_sb); |
| 5826 |
} |
| 5827 |
#endif |
| 5828 |
|
| 5829 |
/* |
| 5830 |
* We use locality group preallocation for small size file. The size of the |
| 5831 |
* file is determined by the current size or the resulting size after |
| 5832 |
* allocation which ever is larger |
| 5833 |
* |
| 5834 |
* One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
| 5835 |
*/ |
| 5836 |
static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) |
| 5837 |
{ |
| 5838 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 5839 |
int bsbits = ac->ac_sb->s_blocksize_bits; |
| 5840 |
loff_t size, isize; |
| 5841 |
bool inode_pa_eligible, group_pa_eligible; |
| 5842 |
|
| 5843 |
if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) |
| 5844 |
return; |
| 5845 |
|
| 5846 |
if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
| 5847 |
return; |
| 5848 |
|
| 5849 |
group_pa_eligible = sbi->s_mb_group_prealloc > 0; |
| 5850 |
inode_pa_eligible = true; |
| 5851 |
size = extent_logical_end(sbi, &ac->ac_o_ex); |
| 5852 |
isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
| 5853 |
>> bsbits; |
| 5854 |
|
| 5855 |
/* No point in using inode preallocation for closed files */ |
| 5856 |
if ((size == isize) && !ext4_fs_is_busy(sbi) && |
| 5857 |
!inode_is_open_for_write(ac->ac_inode)) |
| 5858 |
inode_pa_eligible = false; |
| 5859 |
|
| 5860 |
size = max(size, isize); |
| 5861 |
/* Don't use group allocation for large files */ |
| 5862 |
if (size > sbi->s_mb_stream_request) |
| 5863 |
group_pa_eligible = false; |
| 5864 |
|
| 5865 |
if (!group_pa_eligible) { |
| 5866 |
if (inode_pa_eligible) |
| 5867 |
ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
| 5868 |
else |
| 5869 |
ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; |
| 5870 |
return; |
| 5871 |
} |
| 5872 |
|
| 5873 |
BUG_ON(ac->ac_lg != NULL); |
| 5874 |
/* |
| 5875 |
* locality group prealloc space are per cpu. The reason for having |
| 5876 |
* per cpu locality group is to reduce the contention between block |
| 5877 |
* request from multiple CPUs. |
| 5878 |
*/ |
| 5879 |
ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); |
| 5880 |
|
| 5881 |
/* we're going to use group allocation */ |
| 5882 |
ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; |
| 5883 |
|
| 5884 |
/* serialize all allocations in the group */ |
| 5885 |
mutex_lock(&ac->ac_lg->lg_mutex); |
| 5886 |
} |
| 5887 |
|
| 5888 |
static noinline_for_stack void |
| 5889 |
ext4_mb_initialize_context(struct ext4_allocation_context *ac, |
| 5890 |
struct ext4_allocation_request *ar) |
| 5891 |
{ |
| 5892 |
struct super_block *sb = ar->inode->i_sb; |
| 5893 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 5894 |
struct ext4_super_block *es = sbi->s_es; |
| 5895 |
ext4_group_t group; |
| 5896 |
unsigned int len; |
| 5897 |
ext4_fsblk_t goal; |
| 5898 |
ext4_grpblk_t block; |
| 5899 |
|
| 5900 |
/* we can't allocate > group size */ |
| 5901 |
len = ar->len; |
| 5902 |
|
| 5903 |
/* just a dirty hack to filter too big requests */ |
| 5904 |
if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
| 5905 |
len = EXT4_CLUSTERS_PER_GROUP(sb); |
| 5906 |
|
| 5907 |
/* start searching from the goal */ |
| 5908 |
goal = ar->goal; |
| 5909 |
if (goal < le32_to_cpu(es->s_first_data_block) || |
| 5910 |
goal >= ext4_blocks_count(es)) |
| 5911 |
goal = le32_to_cpu(es->s_first_data_block); |
| 5912 |
ext4_get_group_no_and_offset(sb, goal, &group, &block); |
| 5913 |
|
| 5914 |
/* set up allocation goals */ |
| 5915 |
ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
| 5916 |
ac->ac_status = AC_STATUS_CONTINUE; |
| 5917 |
ac->ac_sb = sb; |
| 5918 |
ac->ac_inode = ar->inode; |
| 5919 |
ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
| 5920 |
ac->ac_o_ex.fe_group = group; |
| 5921 |
ac->ac_o_ex.fe_start = block; |
| 5922 |
ac->ac_o_ex.fe_len = len; |
| 5923 |
ac->ac_g_ex = ac->ac_o_ex; |
| 5924 |
ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; |
| 5925 |
ac->ac_flags = ar->flags; |
| 5926 |
|
| 5927 |
/* we have to define context: we'll work with a file or |
| 5928 |
* locality group. this is a policy, actually */ |
| 5929 |
ext4_mb_group_or_file(ac); |
| 5930 |
|
| 5931 |
mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " |
| 5932 |
"left: %u/%u, right %u/%u to %swritable\n", |
| 5933 |
(unsigned) ar->len, (unsigned) ar->logical, |
| 5934 |
(unsigned) ar->goal, ac->ac_flags, ac->ac_2order, |
| 5935 |
(unsigned) ar->lleft, (unsigned) ar->pleft, |
| 5936 |
(unsigned) ar->lright, (unsigned) ar->pright, |
| 5937 |
inode_is_open_for_write(ar->inode) ? "" : "non-"); |
| 5938 |
} |
| 5939 |
|
| 5940 |
static noinline_for_stack void |
| 5941 |
ext4_mb_discard_lg_preallocations(struct super_block *sb, |
| 5942 |
struct ext4_locality_group *lg, |
| 5943 |
int order, int total_entries) |
| 5944 |
{ |
| 5945 |
ext4_group_t group = 0; |
| 5946 |
struct ext4_buddy e4b; |
| 5947 |
LIST_HEAD(discard_list); |
| 5948 |
struct ext4_prealloc_space *pa, *tmp; |
| 5949 |
|
| 5950 |
mb_debug(sb, "discard locality group preallocation\n"); |
| 5951 |
|
| 5952 |
spin_lock(&lg->lg_prealloc_lock); |
| 5953 |
list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], |
| 5954 |
pa_node.lg_list, |
| 5955 |
lockdep_is_held(&lg->lg_prealloc_lock)) { |
| 5956 |
spin_lock(&pa->pa_lock); |
| 5957 |
if (atomic_read(&pa->pa_count)) { |
| 5958 |
/* |
| 5959 |
* This is the pa that we just used |
| 5960 |
* for block allocation. So don't |
| 5961 |
* free that |
| 5962 |
*/ |
| 5963 |
spin_unlock(&pa->pa_lock); |
| 5964 |
continue; |
| 5965 |
} |
| 5966 |
if (pa->pa_deleted) { |
| 5967 |
spin_unlock(&pa->pa_lock); |
| 5968 |
continue; |
| 5969 |
} |
| 5970 |
/* only lg prealloc space */ |
| 5971 |
BUG_ON(pa->pa_type != MB_GROUP_PA); |
| 5972 |
|
| 5973 |
/* seems this one can be freed ... */ |
| 5974 |
ext4_mb_mark_pa_deleted(sb, pa); |
| 5975 |
spin_unlock(&pa->pa_lock); |
| 5976 |
|
| 5977 |
list_del_rcu(&pa->pa_node.lg_list); |
| 5978 |
list_add(&pa->u.pa_tmp_list, &discard_list); |
| 5979 |
|
| 5980 |
total_entries--; |
| 5981 |
if (total_entries <= 5) { |
| 5982 |
/* |
| 5983 |
* we want to keep only 5 entries |
| 5984 |
* allowing it to grow to 8. This |
| 5985 |
* mak sure we don't call discard |
| 5986 |
* soon for this list. |
| 5987 |
*/ |
| 5988 |
break; |
| 5989 |
} |
| 5990 |
} |
| 5991 |
spin_unlock(&lg->lg_prealloc_lock); |
| 5992 |
|
| 5993 |
list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { |
| 5994 |
int err; |
| 5995 |
|
| 5996 |
group = ext4_get_group_number(sb, pa->pa_pstart); |
| 5997 |
err = ext4_mb_load_buddy_gfp(sb, group, &e4b, |
| 5998 |
GFP_NOFS|__GFP_NOFAIL); |
| 5999 |
if (err) { |
| 6000 |
ext4_error_err(sb, -err, "Error %d loading buddy information for %u", |
| 6001 |
err, group); |
| 6002 |
continue; |
| 6003 |
} |
| 6004 |
ext4_lock_group(sb, group); |
| 6005 |
list_del(&pa->pa_group_list); |
| 6006 |
ext4_mb_release_group_pa(&e4b, pa); |
| 6007 |
ext4_unlock_group(sb, group); |
| 6008 |
|
| 6009 |
ext4_mb_unload_buddy(&e4b); |
| 6010 |
list_del(&pa->u.pa_tmp_list); |
| 6011 |
call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); |
| 6012 |
} |
| 6013 |
} |
| 6014 |
|
| 6015 |
/* |
| 6016 |
* We have incremented pa_count. So it cannot be freed at this |
| 6017 |
* point. Also we hold lg_mutex. So no parallel allocation is |
| 6018 |
* possible from this lg. That means pa_free cannot be updated. |
| 6019 |
* |
| 6020 |
* A parallel ext4_mb_discard_group_preallocations is possible. |
| 6021 |
* which can cause the lg_prealloc_list to be updated. |
| 6022 |
*/ |
| 6023 |
|
| 6024 |
static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) |
| 6025 |
{ |
| 6026 |
int order, added = 0, lg_prealloc_count = 1; |
| 6027 |
struct super_block *sb = ac->ac_sb; |
| 6028 |
struct ext4_locality_group *lg = ac->ac_lg; |
| 6029 |
struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; |
| 6030 |
|
| 6031 |
order = fls(pa->pa_free) - 1; |
| 6032 |
if (order > PREALLOC_TB_SIZE - 1) |
| 6033 |
/* The max size of hash table is PREALLOC_TB_SIZE */ |
| 6034 |
order = PREALLOC_TB_SIZE - 1; |
| 6035 |
/* Add the prealloc space to lg */ |
| 6036 |
spin_lock(&lg->lg_prealloc_lock); |
| 6037 |
list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
| 6038 |
pa_node.lg_list, |
| 6039 |
lockdep_is_held(&lg->lg_prealloc_lock)) { |
| 6040 |
spin_lock(&tmp_pa->pa_lock); |
| 6041 |
if (tmp_pa->pa_deleted) { |
| 6042 |
spin_unlock(&tmp_pa->pa_lock); |
| 6043 |
continue; |
| 6044 |
} |
| 6045 |
if (!added && pa->pa_free < tmp_pa->pa_free) { |
| 6046 |
/* Add to the tail of the previous entry */ |
| 6047 |
list_add_tail_rcu(&pa->pa_node.lg_list, |
| 6048 |
&tmp_pa->pa_node.lg_list); |
| 6049 |
added = 1; |
| 6050 |
/* |
| 6051 |
* we want to count the total |
| 6052 |
* number of entries in the list |
| 6053 |
*/ |
| 6054 |
} |
| 6055 |
spin_unlock(&tmp_pa->pa_lock); |
| 6056 |
lg_prealloc_count++; |
| 6057 |
} |
| 6058 |
if (!added) |
| 6059 |
list_add_tail_rcu(&pa->pa_node.lg_list, |
| 6060 |
&lg->lg_prealloc_list[order]); |
| 6061 |
spin_unlock(&lg->lg_prealloc_lock); |
| 6062 |
|
| 6063 |
/* Now trim the list to be not more than 8 elements */ |
| 6064 |
if (lg_prealloc_count > 8) |
| 6065 |
ext4_mb_discard_lg_preallocations(sb, lg, |
| 6066 |
order, lg_prealloc_count); |
| 6067 |
} |
| 6068 |
|
| 6069 |
/* |
| 6070 |
* release all resource we used in allocation |
| 6071 |
*/ |
| 6072 |
static void ext4_mb_release_context(struct ext4_allocation_context *ac) |
| 6073 |
{ |
| 6074 |
struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
| 6075 |
struct ext4_prealloc_space *pa = ac->ac_pa; |
| 6076 |
if (pa) { |
| 6077 |
if (pa->pa_type == MB_GROUP_PA) { |
| 6078 |
/* see comment in ext4_mb_use_group_pa() */ |
| 6079 |
spin_lock(&pa->pa_lock); |
| 6080 |
pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
| 6081 |
pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
| 6082 |
pa->pa_free -= ac->ac_b_ex.fe_len; |
| 6083 |
pa->pa_len -= ac->ac_b_ex.fe_len; |
| 6084 |
spin_unlock(&pa->pa_lock); |
| 6085 |
|
| 6086 |
/* |
| 6087 |
* We want to add the pa to the right bucket. |
| 6088 |
* Remove it from the list and while adding |
| 6089 |
* make sure the list to which we are adding |
| 6090 |
* doesn't grow big. |
| 6091 |
*/ |
| 6092 |
if (likely(pa->pa_free)) { |
| 6093 |
spin_lock(pa->pa_node_lock.lg_lock); |
| 6094 |
list_del_rcu(&pa->pa_node.lg_list); |
| 6095 |
spin_unlock(pa->pa_node_lock.lg_lock); |
| 6096 |
ext4_mb_add_n_trim(ac); |
| 6097 |
} |
| 6098 |
} |
| 6099 |
|
| 6100 |
ext4_mb_put_pa(ac, ac->ac_sb, pa); |
| 6101 |
} |
| 6102 |
if (ac->ac_bitmap_folio) |
| 6103 |
folio_put(ac->ac_bitmap_folio); |
| 6104 |
if (ac->ac_buddy_folio) |
| 6105 |
folio_put(ac->ac_buddy_folio); |
| 6106 |
if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) |
| 6107 |
mutex_unlock(&ac->ac_lg->lg_mutex); |
| 6108 |
ext4_mb_collect_stats(ac); |
| 6109 |
} |
| 6110 |
|
| 6111 |
static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) |
| 6112 |
{ |
| 6113 |
ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
| 6114 |
int ret; |
| 6115 |
int freed = 0, busy = 0; |
| 6116 |
int retry = 0; |
| 6117 |
|
| 6118 |
trace_ext4_mb_discard_preallocations(sb, needed); |
| 6119 |
|
| 6120 |
if (needed == 0) |
| 6121 |
needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
| 6122 |
repeat: |
| 6123 |
for (i = 0; i < ngroups && needed > 0; i++) { |
| 6124 |
ret = ext4_mb_discard_group_preallocations(sb, i, &busy); |
| 6125 |
freed += ret; |
| 6126 |
needed -= ret; |
| 6127 |
cond_resched(); |
| 6128 |
} |
| 6129 |
|
| 6130 |
if (needed > 0 && busy && ++retry < 3) { |
| 6131 |
busy = 0; |
| 6132 |
goto repeat; |
| 6133 |
} |
| 6134 |
|
| 6135 |
return freed; |
| 6136 |
} |
| 6137 |
|
| 6138 |
static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, |
| 6139 |
struct ext4_allocation_context *ac, u64 *seq) |
| 6140 |
{ |
| 6141 |
int freed; |
| 6142 |
u64 seq_retry = 0; |
| 6143 |
bool ret = false; |
| 6144 |
|
| 6145 |
freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
| 6146 |
if (freed) { |
| 6147 |
ret = true; |
| 6148 |
goto out_dbg; |
| 6149 |
} |
| 6150 |
seq_retry = ext4_get_discard_pa_seq_sum(); |
| 6151 |
if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { |
| 6152 |
ac->ac_flags |= EXT4_MB_STRICT_CHECK; |
| 6153 |
*seq = seq_retry; |
| 6154 |
ret = true; |
| 6155 |
} |
| 6156 |
|
| 6157 |
out_dbg: |
| 6158 |
mb_debug(sb, "freed %d, retry ? %s\n", freed, str_yes_no(ret)); |
| 6159 |
return ret; |
| 6160 |
} |
| 6161 |
|
| 6162 |
/* |
| 6163 |
* Simple allocator for Ext4 fast commit replay path. It searches for blocks |
| 6164 |
* linearly starting at the goal block and also excludes the blocks which |
| 6165 |
* are going to be in use after fast commit replay. |
| 6166 |
*/ |
| 6167 |
static ext4_fsblk_t |
| 6168 |
ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp) |
| 6169 |
{ |
| 6170 |
struct buffer_head *bitmap_bh; |
| 6171 |
struct super_block *sb = ar->inode->i_sb; |
| 6172 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 6173 |
ext4_group_t group, nr; |
| 6174 |
ext4_grpblk_t blkoff; |
| 6175 |
ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
| 6176 |
ext4_grpblk_t i = 0; |
| 6177 |
ext4_fsblk_t goal, block; |
| 6178 |
struct ext4_super_block *es = sbi->s_es; |
| 6179 |
|
| 6180 |
goal = ar->goal; |
| 6181 |
if (goal < le32_to_cpu(es->s_first_data_block) || |
| 6182 |
goal >= ext4_blocks_count(es)) |
| 6183 |
goal = le32_to_cpu(es->s_first_data_block); |
| 6184 |
|
| 6185 |
ar->len = 0; |
| 6186 |
ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); |
| 6187 |
for (nr = ext4_get_groups_count(sb); nr > 0; nr--) { |
| 6188 |
bitmap_bh = ext4_read_block_bitmap(sb, group); |
| 6189 |
if (IS_ERR(bitmap_bh)) { |
| 6190 |
*errp = PTR_ERR(bitmap_bh); |
| 6191 |
pr_warn("Failed to read block bitmap\n"); |
| 6192 |
return 0; |
| 6193 |
} |
| 6194 |
|
| 6195 |
while (1) { |
| 6196 |
i = mb_find_next_zero_bit(bitmap_bh->b_data, max, |
| 6197 |
blkoff); |
| 6198 |
if (i >= max) |
| 6199 |
break; |
| 6200 |
if (ext4_fc_replay_check_excluded(sb, |
| 6201 |
ext4_group_first_block_no(sb, group) + |
| 6202 |
EXT4_C2B(sbi, i))) { |
| 6203 |
blkoff = i + 1; |
| 6204 |
} else |
| 6205 |
break; |
| 6206 |
} |
| 6207 |
brelse(bitmap_bh); |
| 6208 |
if (i < max) |
| 6209 |
break; |
| 6210 |
|
| 6211 |
if (++group >= ext4_get_groups_count(sb)) |
| 6212 |
group = 0; |
| 6213 |
|
| 6214 |
blkoff = 0; |
| 6215 |
} |
| 6216 |
|
| 6217 |
if (i >= max) { |
| 6218 |
*errp = -ENOSPC; |
| 6219 |
return 0; |
| 6220 |
} |
| 6221 |
|
| 6222 |
block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i); |
| 6223 |
ext4_mb_mark_bb(sb, block, 1, true); |
| 6224 |
ar->len = 1; |
| 6225 |
|
| 6226 |
*errp = 0; |
| 6227 |
return block; |
| 6228 |
} |
| 6229 |
|
| 6230 |
/* |
| 6231 |
* Main entry point into mballoc to allocate blocks |
| 6232 |
* it tries to use preallocation first, then falls back |
| 6233 |
* to usual allocation |
| 6234 |
*/ |
| 6235 |
ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, |
| 6236 |
struct ext4_allocation_request *ar, int *errp) |
| 6237 |
{ |
| 6238 |
struct ext4_allocation_context *ac = NULL; |
| 6239 |
struct ext4_sb_info *sbi; |
| 6240 |
struct super_block *sb; |
| 6241 |
ext4_fsblk_t block = 0; |
| 6242 |
unsigned int inquota = 0; |
| 6243 |
unsigned int reserv_clstrs = 0; |
| 6244 |
int retries = 0; |
| 6245 |
u64 seq; |
| 6246 |
|
| 6247 |
might_sleep(); |
| 6248 |
sb = ar->inode->i_sb; |
| 6249 |
sbi = EXT4_SB(sb); |
| 6250 |
|
| 6251 |
trace_ext4_request_blocks(ar); |
| 6252 |
if (sbi->s_mount_state & EXT4_FC_REPLAY) |
| 6253 |
return ext4_mb_new_blocks_simple(ar, errp); |
| 6254 |
|
| 6255 |
/* Allow to use superuser reservation for quota file */ |
| 6256 |
if (ext4_is_quota_file(ar->inode)) |
| 6257 |
ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; |
| 6258 |
|
| 6259 |
if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { |
| 6260 |
/* Without delayed allocation we need to verify |
| 6261 |
* there is enough free blocks to do block allocation |
| 6262 |
* and verify allocation doesn't exceed the quota limits. |
| 6263 |
*/ |
| 6264 |
while (ar->len && |
| 6265 |
ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
| 6266 |
|
| 6267 |
/* let others to free the space */ |
| 6268 |
cond_resched(); |
| 6269 |
ar->len = ar->len >> 1; |
| 6270 |
} |
| 6271 |
if (!ar->len) { |
| 6272 |
ext4_mb_show_pa(sb); |
| 6273 |
*errp = -ENOSPC; |
| 6274 |
return 0; |
| 6275 |
} |
| 6276 |
reserv_clstrs = ar->len; |
| 6277 |
if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
| 6278 |
dquot_alloc_block_nofail(ar->inode, |
| 6279 |
EXT4_C2B(sbi, ar->len)); |
| 6280 |
} else { |
| 6281 |
while (ar->len && |
| 6282 |
dquot_alloc_block(ar->inode, |
| 6283 |
EXT4_C2B(sbi, ar->len))) { |
| 6284 |
|
| 6285 |
ar->flags |= EXT4_MB_HINT_NOPREALLOC; |
| 6286 |
ar->len--; |
| 6287 |
} |
| 6288 |
} |
| 6289 |
inquota = ar->len; |
| 6290 |
if (ar->len == 0) { |
| 6291 |
*errp = -EDQUOT; |
| 6292 |
goto out; |
| 6293 |
} |
| 6294 |
} |
| 6295 |
|
| 6296 |
ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
| 6297 |
if (!ac) { |
| 6298 |
ar->len = 0; |
| 6299 |
*errp = -ENOMEM; |
| 6300 |
goto out; |
| 6301 |
} |
| 6302 |
|
| 6303 |
ext4_mb_initialize_context(ac, ar); |
| 6304 |
|
| 6305 |
ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
| 6306 |
seq = this_cpu_read(discard_pa_seq); |
| 6307 |
if (!ext4_mb_use_preallocated(ac)) { |
| 6308 |
ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
| 6309 |
ext4_mb_normalize_request(ac, ar); |
| 6310 |
|
| 6311 |
*errp = ext4_mb_pa_alloc(ac); |
| 6312 |
if (*errp) |
| 6313 |
goto errout; |
| 6314 |
repeat: |
| 6315 |
/* allocate space in core */ |
| 6316 |
*errp = ext4_mb_regular_allocator(ac); |
| 6317 |
/* |
| 6318 |
* pa allocated above is added to grp->bb_prealloc_list only |
| 6319 |
* when we were able to allocate some block i.e. when |
| 6320 |
* ac->ac_status == AC_STATUS_FOUND. |
| 6321 |
* And error from above mean ac->ac_status != AC_STATUS_FOUND |
| 6322 |
* So we have to free this pa here itself. |
| 6323 |
*/ |
| 6324 |
if (*errp) { |
| 6325 |
ext4_mb_pa_put_free(ac); |
| 6326 |
ext4_discard_allocated_blocks(ac); |
| 6327 |
goto errout; |
| 6328 |
} |
| 6329 |
if (ac->ac_status == AC_STATUS_FOUND && |
| 6330 |
ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) |
| 6331 |
ext4_mb_pa_put_free(ac); |
| 6332 |
} |
| 6333 |
if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
| 6334 |
*errp = ext4_mb_mark_diskspace_used(ac, handle); |
| 6335 |
if (*errp) { |
| 6336 |
ext4_discard_allocated_blocks(ac); |
| 6337 |
goto errout; |
| 6338 |
} else { |
| 6339 |
block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
| 6340 |
ar->len = ac->ac_b_ex.fe_len; |
| 6341 |
} |
| 6342 |
} else { |
| 6343 |
if (++retries < 3 && |
| 6344 |
ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) |
| 6345 |
goto repeat; |
| 6346 |
/* |
| 6347 |
* If block allocation fails then the pa allocated above |
| 6348 |
* needs to be freed here itself. |
| 6349 |
*/ |
| 6350 |
ext4_mb_pa_put_free(ac); |
| 6351 |
*errp = -ENOSPC; |
| 6352 |
} |
| 6353 |
|
| 6354 |
if (*errp) { |
| 6355 |
errout: |
| 6356 |
ac->ac_b_ex.fe_len = 0; |
| 6357 |
ar->len = 0; |
| 6358 |
ext4_mb_show_ac(ac); |
| 6359 |
} |
| 6360 |
ext4_mb_release_context(ac); |
| 6361 |
kmem_cache_free(ext4_ac_cachep, ac); |
| 6362 |
out: |
| 6363 |
if (inquota && ar->len < inquota) |
| 6364 |
dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
| 6365 |
/* release any reserved blocks */ |
| 6366 |
if (reserv_clstrs) |
| 6367 |
percpu_counter_sub(&sbi->s_dirtyclusters_counter, reserv_clstrs); |
| 6368 |
|
| 6369 |
trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
| 6370 |
|
| 6371 |
return block; |
| 6372 |
} |
| 6373 |
|
| 6374 |
/* |
| 6375 |
* We can merge two free data extents only if the physical blocks |
| 6376 |
* are contiguous, AND the extents were freed by the same transaction, |
| 6377 |
* AND the blocks are associated with the same group. |
| 6378 |
*/ |
| 6379 |
static inline bool |
| 6380 |
ext4_freed_extents_can_be_merged(struct ext4_free_data *entry1, |
| 6381 |
struct ext4_free_data *entry2) |
| 6382 |
{ |
| 6383 |
if (entry1->efd_tid != entry2->efd_tid) |
| 6384 |
return false; |
| 6385 |
if (entry1->efd_start_cluster + entry1->efd_count != |
| 6386 |
entry2->efd_start_cluster) |
| 6387 |
return false; |
| 6388 |
if (WARN_ON_ONCE(entry1->efd_group != entry2->efd_group)) |
| 6389 |
return false; |
| 6390 |
return true; |
| 6391 |
} |
| 6392 |
|
| 6393 |
static inline void |
| 6394 |
ext4_merge_freed_extents(struct ext4_sb_info *sbi, struct rb_root *root, |
| 6395 |
struct ext4_free_data *entry1, |
| 6396 |
struct ext4_free_data *entry2) |
| 6397 |
{ |
| 6398 |
entry1->efd_count += entry2->efd_count; |
| 6399 |
spin_lock(&sbi->s_md_lock); |
| 6400 |
list_del(&entry2->efd_list); |
| 6401 |
spin_unlock(&sbi->s_md_lock); |
| 6402 |
rb_erase(&entry2->efd_node, root); |
| 6403 |
kmem_cache_free(ext4_free_data_cachep, entry2); |
| 6404 |
} |
| 6405 |
|
| 6406 |
static inline void |
| 6407 |
ext4_try_merge_freed_extent_prev(struct ext4_sb_info *sbi, struct rb_root *root, |
| 6408 |
struct ext4_free_data *entry) |
| 6409 |
{ |
| 6410 |
struct ext4_free_data *prev; |
| 6411 |
struct rb_node *node; |
| 6412 |
|
| 6413 |
node = rb_prev(&entry->efd_node); |
| 6414 |
if (!node) |
| 6415 |
return; |
| 6416 |
|
| 6417 |
prev = rb_entry(node, struct ext4_free_data, efd_node); |
| 6418 |
if (ext4_freed_extents_can_be_merged(prev, entry)) |
| 6419 |
ext4_merge_freed_extents(sbi, root, prev, entry); |
| 6420 |
} |
| 6421 |
|
| 6422 |
static inline void |
| 6423 |
ext4_try_merge_freed_extent_next(struct ext4_sb_info *sbi, struct rb_root *root, |
| 6424 |
struct ext4_free_data *entry) |
| 6425 |
{ |
| 6426 |
struct ext4_free_data *next; |
| 6427 |
struct rb_node *node; |
| 6428 |
|
| 6429 |
node = rb_next(&entry->efd_node); |
| 6430 |
if (!node) |
| 6431 |
return; |
| 6432 |
|
| 6433 |
next = rb_entry(node, struct ext4_free_data, efd_node); |
| 6434 |
if (ext4_freed_extents_can_be_merged(entry, next)) |
| 6435 |
ext4_merge_freed_extents(sbi, root, entry, next); |
| 6436 |
} |
| 6437 |
|
| 6438 |
static noinline_for_stack void |
| 6439 |
ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, |
| 6440 |
struct ext4_free_data *new_entry) |
| 6441 |
{ |
| 6442 |
ext4_group_t group = e4b->bd_group; |
| 6443 |
ext4_grpblk_t cluster; |
| 6444 |
ext4_grpblk_t clusters = new_entry->efd_count; |
| 6445 |
struct ext4_free_data *entry = NULL; |
| 6446 |
struct ext4_group_info *db = e4b->bd_info; |
| 6447 |
struct super_block *sb = e4b->bd_sb; |
| 6448 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 6449 |
struct rb_root *root = &db->bb_free_root; |
| 6450 |
struct rb_node **n = &root->rb_node; |
| 6451 |
struct rb_node *parent = NULL, *new_node; |
| 6452 |
|
| 6453 |
BUG_ON(!ext4_handle_valid(handle)); |
| 6454 |
BUG_ON(e4b->bd_bitmap_folio == NULL); |
| 6455 |
BUG_ON(e4b->bd_buddy_folio == NULL); |
| 6456 |
|
| 6457 |
new_node = &new_entry->efd_node; |
| 6458 |
cluster = new_entry->efd_start_cluster; |
| 6459 |
|
| 6460 |
if (!*n) { |
| 6461 |
/* first free block exent. We need to |
| 6462 |
protect buddy cache from being freed, |
| 6463 |
* otherwise we'll refresh it from |
| 6464 |
* on-disk bitmap and lose not-yet-available |
| 6465 |
* blocks */ |
| 6466 |
folio_get(e4b->bd_buddy_folio); |
| 6467 |
folio_get(e4b->bd_bitmap_folio); |
| 6468 |
} |
| 6469 |
while (*n) { |
| 6470 |
parent = *n; |
| 6471 |
entry = rb_entry(parent, struct ext4_free_data, efd_node); |
| 6472 |
if (cluster < entry->efd_start_cluster) |
| 6473 |
n = &(*n)->rb_left; |
| 6474 |
else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
| 6475 |
n = &(*n)->rb_right; |
| 6476 |
else { |
| 6477 |
ext4_grp_locked_error(sb, group, 0, |
| 6478 |
ext4_group_first_block_no(sb, group) + |
| 6479 |
EXT4_C2B(sbi, cluster), |
| 6480 |
"Block already on to-be-freed list"); |
| 6481 |
kmem_cache_free(ext4_free_data_cachep, new_entry); |
| 6482 |
return; |
| 6483 |
} |
| 6484 |
} |
| 6485 |
|
| 6486 |
atomic_add(clusters, &sbi->s_mb_free_pending); |
| 6487 |
if (!entry) |
| 6488 |
goto insert; |
| 6489 |
|
| 6490 |
/* Now try to see the extent can be merged to prev and next */ |
| 6491 |
if (ext4_freed_extents_can_be_merged(new_entry, entry)) { |
| 6492 |
entry->efd_start_cluster = cluster; |
| 6493 |
entry->efd_count += new_entry->efd_count; |
| 6494 |
kmem_cache_free(ext4_free_data_cachep, new_entry); |
| 6495 |
ext4_try_merge_freed_extent_prev(sbi, root, entry); |
| 6496 |
return; |
| 6497 |
} |
| 6498 |
if (ext4_freed_extents_can_be_merged(entry, new_entry)) { |
| 6499 |
entry->efd_count += new_entry->efd_count; |
| 6500 |
kmem_cache_free(ext4_free_data_cachep, new_entry); |
| 6501 |
ext4_try_merge_freed_extent_next(sbi, root, entry); |
| 6502 |
return; |
| 6503 |
} |
| 6504 |
insert: |
| 6505 |
rb_link_node(new_node, parent, n); |
| 6506 |
rb_insert_color(new_node, root); |
| 6507 |
|
| 6508 |
spin_lock(&sbi->s_md_lock); |
| 6509 |
list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]); |
| 6510 |
spin_unlock(&sbi->s_md_lock); |
| 6511 |
} |
| 6512 |
|
| 6513 |
static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, |
| 6514 |
unsigned long count) |
| 6515 |
{ |
| 6516 |
struct super_block *sb = inode->i_sb; |
| 6517 |
ext4_group_t group; |
| 6518 |
ext4_grpblk_t blkoff; |
| 6519 |
|
| 6520 |
ext4_get_group_no_and_offset(sb, block, &group, &blkoff); |
| 6521 |
ext4_mb_mark_context(NULL, sb, false, group, blkoff, count, |
| 6522 |
EXT4_MB_BITMAP_MARKED_CHECK | |
| 6523 |
EXT4_MB_SYNC_UPDATE, |
| 6524 |
NULL); |
| 6525 |
} |
| 6526 |
|
| 6527 |
/** |
| 6528 |
* ext4_mb_clear_bb() -- helper function for freeing blocks. |
| 6529 |
* Used by ext4_free_blocks() |
| 6530 |
* @handle: handle for this transaction |
| 6531 |
* @inode: inode |
| 6532 |
* @block: starting physical block to be freed |
| 6533 |
* @count: number of blocks to be freed |
| 6534 |
* @flags: flags used by ext4_free_blocks |
| 6535 |
*/ |
| 6536 |
static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, |
| 6537 |
ext4_fsblk_t block, unsigned long count, |
| 6538 |
int flags) |
| 6539 |
{ |
| 6540 |
struct super_block *sb = inode->i_sb; |
| 6541 |
struct ext4_group_info *grp; |
| 6542 |
unsigned int overflow; |
| 6543 |
ext4_grpblk_t bit; |
| 6544 |
ext4_group_t block_group; |
| 6545 |
struct ext4_sb_info *sbi; |
| 6546 |
struct ext4_buddy e4b; |
| 6547 |
unsigned int count_clusters; |
| 6548 |
int err = 0; |
| 6549 |
int mark_flags = 0; |
| 6550 |
ext4_grpblk_t changed; |
| 6551 |
|
| 6552 |
sbi = EXT4_SB(sb); |
| 6553 |
|
| 6554 |
if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
| 6555 |
!ext4_inode_block_valid(inode, block, count)) { |
| 6556 |
ext4_error(sb, "Freeing blocks in system zone - " |
| 6557 |
"Block = %llu, count = %lu", block, count); |
| 6558 |
/* err = 0. ext4_std_error should be a no op */ |
| 6559 |
goto error_out; |
| 6560 |
} |
| 6561 |
flags |= EXT4_FREE_BLOCKS_VALIDATED; |
| 6562 |
|
| 6563 |
do_more: |
| 6564 |
overflow = 0; |
| 6565 |
ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
| 6566 |
|
| 6567 |
grp = ext4_get_group_info(sb, block_group); |
| 6568 |
if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
| 6569 |
return; |
| 6570 |
|
| 6571 |
/* |
| 6572 |
* Check to see if we are freeing blocks across a group |
| 6573 |
* boundary. |
| 6574 |
*/ |
| 6575 |
if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
| 6576 |
overflow = EXT4_C2B(sbi, bit) + count - |
| 6577 |
EXT4_BLOCKS_PER_GROUP(sb); |
| 6578 |
count -= overflow; |
| 6579 |
/* The range changed so it's no longer validated */ |
| 6580 |
flags &= ~EXT4_FREE_BLOCKS_VALIDATED; |
| 6581 |
} |
| 6582 |
count_clusters = EXT4_NUM_B2C(sbi, count); |
| 6583 |
trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
| 6584 |
|
| 6585 |
/* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ |
| 6586 |
err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, |
| 6587 |
GFP_NOFS|__GFP_NOFAIL); |
| 6588 |
if (err) |
| 6589 |
goto error_out; |
| 6590 |
|
| 6591 |
if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
| 6592 |
!ext4_inode_block_valid(inode, block, count)) { |
| 6593 |
ext4_error(sb, "Freeing blocks in system zone - " |
| 6594 |
"Block = %llu, count = %lu", block, count); |
| 6595 |
/* err = 0. ext4_std_error should be a no op */ |
| 6596 |
goto error_clean; |
| 6597 |
} |
| 6598 |
|
| 6599 |
#ifdef AGGRESSIVE_CHECK |
| 6600 |
mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK; |
| 6601 |
#endif |
| 6602 |
err = ext4_mb_mark_context(handle, sb, false, block_group, bit, |
| 6603 |
count_clusters, mark_flags, &changed); |
| 6604 |
|
| 6605 |
|
| 6606 |
if (err && changed == 0) |
| 6607 |
goto error_clean; |
| 6608 |
|
| 6609 |
#ifdef AGGRESSIVE_CHECK |
| 6610 |
BUG_ON(changed != count_clusters); |
| 6611 |
#endif |
| 6612 |
|
| 6613 |
/* |
| 6614 |
* We need to make sure we don't reuse the freed block until after the |
| 6615 |
* transaction is committed. We make an exception if the inode is to be |
| 6616 |
* written in writeback mode since writeback mode has weak data |
| 6617 |
* consistency guarantees. |
| 6618 |
*/ |
| 6619 |
if (ext4_handle_valid(handle) && |
| 6620 |
((flags & EXT4_FREE_BLOCKS_METADATA) || |
| 6621 |
!ext4_should_writeback_data(inode))) { |
| 6622 |
struct ext4_free_data *new_entry; |
| 6623 |
/* |
| 6624 |
* We use __GFP_NOFAIL because ext4_free_blocks() is not allowed |
| 6625 |
* to fail. |
| 6626 |
*/ |
| 6627 |
new_entry = kmem_cache_alloc(ext4_free_data_cachep, |
| 6628 |
GFP_NOFS|__GFP_NOFAIL); |
| 6629 |
new_entry->efd_start_cluster = bit; |
| 6630 |
new_entry->efd_group = block_group; |
| 6631 |
new_entry->efd_count = count_clusters; |
| 6632 |
new_entry->efd_tid = handle->h_transaction->t_tid; |
| 6633 |
|
| 6634 |
ext4_lock_group(sb, block_group); |
| 6635 |
ext4_mb_free_metadata(handle, &e4b, new_entry); |
| 6636 |
} else { |
| 6637 |
if (test_opt(sb, DISCARD)) { |
| 6638 |
err = ext4_issue_discard(sb, block_group, bit, |
| 6639 |
count_clusters); |
| 6640 |
/* |
| 6641 |
* Ignore EOPNOTSUPP error. This is consistent with |
| 6642 |
* what happens when using journal. |
| 6643 |
*/ |
| 6644 |
if (err == -EOPNOTSUPP) |
| 6645 |
err = 0; |
| 6646 |
if (err) |
| 6647 |
ext4_msg(sb, KERN_WARNING, "discard request in" |
| 6648 |
" group:%u block:%d count:%lu failed" |
| 6649 |
" with %d", block_group, bit, count, |
| 6650 |
err); |
| 6651 |
} |
| 6652 |
|
| 6653 |
EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); |
| 6654 |
|
| 6655 |
ext4_lock_group(sb, block_group); |
| 6656 |
mb_free_blocks(inode, &e4b, bit, count_clusters); |
| 6657 |
} |
| 6658 |
|
| 6659 |
ext4_unlock_group(sb, block_group); |
| 6660 |
|
| 6661 |
/* |
| 6662 |
* on a bigalloc file system, defer the s_freeclusters_counter |
| 6663 |
* update to the caller (ext4_remove_space and friends) so they |
| 6664 |
* can determine if a cluster freed here should be rereserved |
| 6665 |
*/ |
| 6666 |
if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { |
| 6667 |
if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) |
| 6668 |
dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); |
| 6669 |
percpu_counter_add(&sbi->s_freeclusters_counter, |
| 6670 |
count_clusters); |
| 6671 |
} |
| 6672 |
|
| 6673 |
if (overflow && !err) { |
| 6674 |
block += count; |
| 6675 |
count = overflow; |
| 6676 |
ext4_mb_unload_buddy(&e4b); |
| 6677 |
/* The range changed so it's no longer validated */ |
| 6678 |
flags &= ~EXT4_FREE_BLOCKS_VALIDATED; |
| 6679 |
goto do_more; |
| 6680 |
} |
| 6681 |
|
| 6682 |
error_clean: |
| 6683 |
ext4_mb_unload_buddy(&e4b); |
| 6684 |
error_out: |
| 6685 |
ext4_std_error(sb, err); |
| 6686 |
} |
| 6687 |
|
| 6688 |
/** |
| 6689 |
* ext4_free_blocks() -- Free given blocks and update quota |
| 6690 |
* @handle: handle for this transaction |
| 6691 |
* @inode: inode |
| 6692 |
* @bh: optional buffer of the block to be freed |
| 6693 |
* @block: starting physical block to be freed |
| 6694 |
* @count: number of blocks to be freed |
| 6695 |
* @flags: flags used by ext4_free_blocks |
| 6696 |
*/ |
| 6697 |
void ext4_free_blocks(handle_t *handle, struct inode *inode, |
| 6698 |
struct buffer_head *bh, ext4_fsblk_t block, |
| 6699 |
unsigned long count, int flags) |
| 6700 |
{ |
| 6701 |
struct super_block *sb = inode->i_sb; |
| 6702 |
unsigned int overflow; |
| 6703 |
struct ext4_sb_info *sbi; |
| 6704 |
|
| 6705 |
sbi = EXT4_SB(sb); |
| 6706 |
|
| 6707 |
if (bh) { |
| 6708 |
if (block) |
| 6709 |
BUG_ON(block != bh->b_blocknr); |
| 6710 |
else |
| 6711 |
block = bh->b_blocknr; |
| 6712 |
} |
| 6713 |
|
| 6714 |
if (sbi->s_mount_state & EXT4_FC_REPLAY) { |
| 6715 |
ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count)); |
| 6716 |
return; |
| 6717 |
} |
| 6718 |
|
| 6719 |
might_sleep(); |
| 6720 |
|
| 6721 |
if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
| 6722 |
!ext4_inode_block_valid(inode, block, count)) { |
| 6723 |
ext4_error(sb, "Freeing blocks not in datazone - " |
| 6724 |
"block = %llu, count = %lu", block, count); |
| 6725 |
return; |
| 6726 |
} |
| 6727 |
flags |= EXT4_FREE_BLOCKS_VALIDATED; |
| 6728 |
|
| 6729 |
ext4_debug("freeing block %llu\n", block); |
| 6730 |
trace_ext4_free_blocks(inode, block, count, flags); |
| 6731 |
|
| 6732 |
if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
| 6733 |
BUG_ON(count > 1); |
| 6734 |
|
| 6735 |
ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
| 6736 |
inode, bh, block); |
| 6737 |
} |
| 6738 |
|
| 6739 |
/* |
| 6740 |
* If the extent to be freed does not begin on a cluster |
| 6741 |
* boundary, we need to deal with partial clusters at the |
| 6742 |
* beginning and end of the extent. Normally we will free |
| 6743 |
* blocks at the beginning or the end unless we are explicitly |
| 6744 |
* requested to avoid doing so. |
| 6745 |
*/ |
| 6746 |
overflow = EXT4_PBLK_COFF(sbi, block); |
| 6747 |
if (overflow) { |
| 6748 |
if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { |
| 6749 |
overflow = sbi->s_cluster_ratio - overflow; |
| 6750 |
block += overflow; |
| 6751 |
if (count > overflow) |
| 6752 |
count -= overflow; |
| 6753 |
else |
| 6754 |
return; |
| 6755 |
} else { |
| 6756 |
block -= overflow; |
| 6757 |
count += overflow; |
| 6758 |
} |
| 6759 |
/* The range changed so it's no longer validated */ |
| 6760 |
flags &= ~EXT4_FREE_BLOCKS_VALIDATED; |
| 6761 |
} |
| 6762 |
overflow = EXT4_LBLK_COFF(sbi, count); |
| 6763 |
if (overflow) { |
| 6764 |
if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { |
| 6765 |
if (count > overflow) |
| 6766 |
count -= overflow; |
| 6767 |
else |
| 6768 |
return; |
| 6769 |
} else |
| 6770 |
count += sbi->s_cluster_ratio - overflow; |
| 6771 |
/* The range changed so it's no longer validated */ |
| 6772 |
flags &= ~EXT4_FREE_BLOCKS_VALIDATED; |
| 6773 |
} |
| 6774 |
|
| 6775 |
if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { |
| 6776 |
int i; |
| 6777 |
int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; |
| 6778 |
|
| 6779 |
for (i = 0; i < count; i++) { |
| 6780 |
cond_resched(); |
| 6781 |
if (is_metadata) |
| 6782 |
bh = sb_find_get_block_nonatomic(inode->i_sb, |
| 6783 |
block + i); |
| 6784 |
ext4_forget(handle, is_metadata, inode, bh, block + i); |
| 6785 |
} |
| 6786 |
} |
| 6787 |
|
| 6788 |
ext4_mb_clear_bb(handle, inode, block, count, flags); |
| 6789 |
} |
| 6790 |
|
| 6791 |
/** |
| 6792 |
* ext4_group_add_blocks() -- Add given blocks to an existing group |
| 6793 |
* @handle: handle to this transaction |
| 6794 |
* @sb: super block |
| 6795 |
* @block: start physical block to add to the block group |
| 6796 |
* @count: number of blocks to free |
| 6797 |
* |
| 6798 |
* This marks the blocks as free in the bitmap and buddy. |
| 6799 |
*/ |
| 6800 |
int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
| 6801 |
ext4_fsblk_t block, unsigned long count) |
| 6802 |
{ |
| 6803 |
ext4_group_t block_group; |
| 6804 |
ext4_grpblk_t bit; |
| 6805 |
struct ext4_sb_info *sbi = EXT4_SB(sb); |
| 6806 |
struct ext4_buddy e4b; |
| 6807 |
int err = 0; |
| 6808 |
ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); |
| 6809 |
ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); |
| 6810 |
unsigned long cluster_count = last_cluster - first_cluster + 1; |
| 6811 |
ext4_grpblk_t changed; |
| 6812 |
|
| 6813 |
ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); |
| 6814 |
|
| 6815 |
if (cluster_count == 0) |
| 6816 |
return 0; |
| 6817 |
|
| 6818 |
ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
| 6819 |
/* |
| 6820 |
* Check to see if we are freeing blocks across a group |
| 6821 |
* boundary. |
| 6822 |
*/ |
| 6823 |
if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { |
| 6824 |
ext4_warning(sb, "too many blocks added to group %u", |
| 6825 |
block_group); |
| 6826 |
err = -EINVAL; |
| 6827 |
goto error_out; |
| 6828 |
} |
| 6829 |
|
| 6830 |
err = ext4_mb_load_buddy(sb, block_group, &e4b); |
| 6831 |
if (err) |
| 6832 |
goto error_out; |
| 6833 |
|
| 6834 |
if (!ext4_sb_block_valid(sb, NULL, block, count)) { |
| 6835 |
ext4_error(sb, "Adding blocks in system zones - " |
| 6836 |
"Block = %llu, count = %lu", |
| 6837 |
block, count); |
| 6838 |
err = -EINVAL; |
| 6839 |
goto error_clean; |
| 6840 |
} |
| 6841 |
|
| 6842 |
err = ext4_mb_mark_context(handle, sb, false, block_group, bit, |
| 6843 |
cluster_count, EXT4_MB_BITMAP_MARKED_CHECK, |
| 6844 |
&changed); |
| 6845 |
if (err && changed == 0) |
| 6846 |
goto error_clean; |
| 6847 |
|
| 6848 |
if (changed != cluster_count) |
| 6849 |
ext4_error(sb, "bit already cleared in group %u", block_group); |
| 6850 |
|
| 6851 |
ext4_lock_group(sb, block_group); |
| 6852 |
mb_free_blocks(NULL, &e4b, bit, cluster_count); |
| 6853 |
ext4_unlock_group(sb, block_group); |
| 6854 |
percpu_counter_add(&sbi->s_freeclusters_counter, |
| 6855 |
changed); |
| 6856 |
|
| 6857 |
error_clean: |
| 6858 |
ext4_mb_unload_buddy(&e4b); |
| 6859 |
error_out: |
| 6860 |
ext4_std_error(sb, err); |
| 6861 |
return err; |
| 6862 |
} |
| 6863 |
|
| 6864 |
/** |
| 6865 |
* ext4_trim_extent -- function to TRIM one single free extent in the group |
| 6866 |
* @sb: super block for the file system |
| 6867 |
* @start: starting block of the free extent in the alloc. group |
| 6868 |
* @count: number of blocks to TRIM |
| 6869 |
* @e4b: ext4 buddy for the group |
| 6870 |
* |
| 6871 |
* Trim "count" blocks starting at "start" in the "group". To assure that no |
| 6872 |
* one will allocate those blocks, mark it as used in buddy bitmap. This must |
| 6873 |
* be called with under the group lock. |
| 6874 |
*/ |
| 6875 |
static int ext4_trim_extent(struct super_block *sb, |
| 6876 |
int start, int count, struct ext4_buddy *e4b) |
| 6877 |
__releases(bitlock) |
| 6878 |
__acquires(bitlock) |
| 6879 |
{ |
| 6880 |
struct ext4_free_extent ex; |
| 6881 |
ext4_group_t group = e4b->bd_group; |
| 6882 |
int ret = 0; |
| 6883 |
|
| 6884 |
trace_ext4_trim_extent(sb, group, start, count); |
| 6885 |
|
| 6886 |
assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
| 6887 |
|
| 6888 |
ex.fe_start = start; |
| 6889 |
ex.fe_group = group; |
| 6890 |
ex.fe_len = count; |
| 6891 |
|
| 6892 |
/* |
| 6893 |
* Mark blocks used, so no one can reuse them while |
| 6894 |
* being trimmed. |
| 6895 |
*/ |
| 6896 |
mb_mark_used(e4b, &ex); |
| 6897 |
ext4_unlock_group(sb, group); |
| 6898 |
ret = ext4_issue_discard(sb, group, start, count); |
| 6899 |
ext4_lock_group(sb, group); |
| 6900 |
mb_free_blocks(NULL, e4b, start, ex.fe_len); |
| 6901 |
return ret; |
| 6902 |
} |
| 6903 |
|
| 6904 |
static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb, |
| 6905 |
ext4_group_t grp) |
| 6906 |
{ |
| 6907 |
unsigned long nr_clusters_in_group; |
| 6908 |
|
| 6909 |
if (grp < (ext4_get_groups_count(sb) - 1)) |
| 6910 |
nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb); |
| 6911 |
else |
| 6912 |
nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) - |
| 6913 |
ext4_group_first_block_no(sb, grp)) |
| 6914 |
>> EXT4_CLUSTER_BITS(sb); |
| 6915 |
|
| 6916 |
return nr_clusters_in_group - 1; |
| 6917 |
} |
| 6918 |
|
| 6919 |
static bool ext4_trim_interrupted(void) |
| 6920 |
{ |
| 6921 |
return fatal_signal_pending(current) || freezing(current); |
| 6922 |
} |
| 6923 |
|
| 6924 |
static int ext4_try_to_trim_range(struct super_block *sb, |
| 6925 |
struct ext4_buddy *e4b, ext4_grpblk_t start, |
| 6926 |
ext4_grpblk_t max, ext4_grpblk_t minblocks) |
| 6927 |
__acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) |
| 6928 |
__releases(ext4_group_lock_ptr(sb, e4b->bd_group)) |
| 6929 |
{ |
| 6930 |
ext4_grpblk_t next, count, free_count, last, origin_start; |
| 6931 |
bool set_trimmed = false; |
| 6932 |
void *bitmap; |
| 6933 |
|
| 6934 |
if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) |
| 6935 |
return 0; |
| 6936 |
|
| 6937 |
last = ext4_last_grp_cluster(sb, e4b->bd_group); |
| 6938 |
bitmap = e4b->bd_bitmap; |
| 6939 |
if (start == 0 && max >= last) |
| 6940 |
set_trimmed = true; |
| 6941 |
origin_start = start; |
| 6942 |
start = max(e4b->bd_info->bb_first_free, start); |
| 6943 |
count = 0; |
| 6944 |
free_count = 0; |
| 6945 |
|
| 6946 |
while (start <= max) { |
| 6947 |
start = mb_find_next_zero_bit(bitmap, max + 1, start); |
| 6948 |
if (start > max) |
| 6949 |
break; |
| 6950 |
|
| 6951 |
next = mb_find_next_bit(bitmap, last + 1, start); |
| 6952 |
if (origin_start == 0 && next >= last) |
| 6953 |
set_trimmed = true; |
| 6954 |
|
| 6955 |
if ((next - start) >= minblocks) { |
| 6956 |
int ret = ext4_trim_extent(sb, start, next - start, e4b); |
| 6957 |
|
| 6958 |
if (ret && ret != -EOPNOTSUPP) |
| 6959 |
return count; |
| 6960 |
count += next - start; |
| 6961 |
} |
| 6962 |
free_count += next - start; |
| 6963 |
start = next + 1; |
| 6964 |
|
| 6965 |
if (ext4_trim_interrupted()) |
| 6966 |
return count; |
| 6967 |
|
| 6968 |
if (need_resched()) { |
| 6969 |
ext4_unlock_group(sb, e4b->bd_group); |
| 6970 |
cond_resched(); |
| 6971 |
ext4_lock_group(sb, e4b->bd_group); |
| 6972 |
} |
| 6973 |
|
| 6974 |
if ((e4b->bd_info->bb_free - free_count) < minblocks) |
| 6975 |
break; |
| 6976 |
} |
| 6977 |
|
| 6978 |
if (set_trimmed) |
| 6979 |
EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info); |
| 6980 |
|
| 6981 |
return count; |
| 6982 |
} |
| 6983 |
|
| 6984 |
/** |
| 6985 |
* ext4_trim_all_free -- function to trim all free space in alloc. group |
| 6986 |
* @sb: super block for file system |
| 6987 |
* @group: group to be trimmed |
| 6988 |
* @start: first group block to examine |
| 6989 |
* @max: last group block to examine |
| 6990 |
* @minblocks: minimum extent block count |
| 6991 |
* |
| 6992 |
* ext4_trim_all_free walks through group's block bitmap searching for free |
| 6993 |
* extents. When the free extent is found, mark it as used in group buddy |
| 6994 |
* bitmap. Then issue a TRIM command on this extent and free the extent in |
| 6995 |
* the group buddy bitmap. |
| 6996 |
*/ |
| 6997 |
static ext4_grpblk_t |
| 6998 |
ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
| 6999 |
ext4_grpblk_t start, ext4_grpblk_t max, |
| 7000 |
ext4_grpblk_t minblocks) |
| 7001 |
{ |
| 7002 |
struct ext4_buddy e4b; |
| 7003 |
int ret; |
| 7004 |
|
| 7005 |
trace_ext4_trim_all_free(sb, group, start, max); |
| 7006 |
|
| 7007 |
ret = ext4_mb_load_buddy(sb, group, &e4b); |
| 7008 |
if (ret) { |
| 7009 |
ext4_warning(sb, "Error %d loading buddy information for %u", |
| 7010 |
ret, group); |
| 7011 |
return ret; |
| 7012 |
} |
| 7013 |
|
| 7014 |
ext4_lock_group(sb, group); |
| 7015 |
|
| 7016 |
if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || |
| 7017 |
minblocks < EXT4_SB(sb)->s_last_trim_minblks) |
| 7018 |
ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); |
| 7019 |
else |
| 7020 |
ret = 0; |
| 7021 |
|
| 7022 |
ext4_unlock_group(sb, group); |
| 7023 |
ext4_mb_unload_buddy(&e4b); |
| 7024 |
|
| 7025 |
ext4_debug("trimmed %d blocks in the group %d\n", |
| 7026 |
ret, group); |
| 7027 |
|
| 7028 |
return ret; |
| 7029 |
} |
| 7030 |
|
| 7031 |
/** |
| 7032 |
* ext4_trim_fs() -- trim ioctl handle function |
| 7033 |
* @sb: superblock for filesystem |
| 7034 |
* @range: fstrim_range structure |
| 7035 |
* |
| 7036 |
* start: First Byte to trim |
| 7037 |
* len: number of Bytes to trim from start |
| 7038 |
* minlen: minimum extent length in Bytes |
| 7039 |
* ext4_trim_fs goes through all allocation groups containing Bytes from |
| 7040 |
* start to start+len. For each such a group ext4_trim_all_free function |
| 7041 |
* is invoked to trim all free space. |
| 7042 |
*/ |
| 7043 |
int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) |
| 7044 |
{ |
| 7045 |
unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); |
| 7046 |
struct ext4_group_info *grp; |
| 7047 |
ext4_group_t group, first_group, last_group; |
| 7048 |
ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
| 7049 |
uint64_t start, end, minlen, trimmed = 0; |
| 7050 |
ext4_fsblk_t first_data_blk = |
| 7051 |
le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); |
| 7052 |
ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
| 7053 |
int ret = 0; |
| 7054 |
|
| 7055 |
start = range->start >> sb->s_blocksize_bits; |
| 7056 |
end = start + (range->len >> sb->s_blocksize_bits) - 1; |
| 7057 |
minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
| 7058 |
range->minlen >> sb->s_blocksize_bits); |
| 7059 |
|
| 7060 |
if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
| 7061 |
start >= max_blks || |
| 7062 |
range->len < sb->s_blocksize) |
| 7063 |
return -EINVAL; |
| 7064 |
/* No point to try to trim less than discard granularity */ |
| 7065 |
if (range->minlen < discard_granularity) { |
| 7066 |
minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
| 7067 |
discard_granularity >> sb->s_blocksize_bits); |
| 7068 |
if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) |
| 7069 |
goto out; |
| 7070 |
} |
| 7071 |
if (end >= max_blks - 1) |
| 7072 |
end = max_blks - 1; |
| 7073 |
if (end <= first_data_blk) |
| 7074 |
goto out; |
| 7075 |
if (start < first_data_blk) |
| 7076 |
start = first_data_blk; |
| 7077 |
|
| 7078 |
/* Determine first and last group to examine based on start and end */ |
| 7079 |
ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
| 7080 |
&first_group, &first_cluster); |
| 7081 |
ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
| 7082 |
&last_group, &last_cluster); |
| 7083 |
|
| 7084 |
/* end now represents the last cluster to discard in this group */ |
| 7085 |
end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; |
| 7086 |
|
| 7087 |
for (group = first_group; group <= last_group; group++) { |
| 7088 |
if (ext4_trim_interrupted()) |
| 7089 |
break; |
| 7090 |
grp = ext4_get_group_info(sb, group); |
| 7091 |
if (!grp) |
| 7092 |
continue; |
| 7093 |
/* We only do this if the grp has never been initialized */ |
| 7094 |
if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
| 7095 |
ret = ext4_mb_init_group(sb, group, GFP_NOFS); |
| 7096 |
if (ret) |
| 7097 |
break; |
| 7098 |
} |
| 7099 |
|
| 7100 |
/* |
| 7101 |
* For all the groups except the last one, last cluster will |
| 7102 |
* always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to |
| 7103 |
* change it for the last group, note that last_cluster is |
| 7104 |
* already computed earlier by ext4_get_group_no_and_offset() |
| 7105 |
*/ |
| 7106 |
if (group == last_group) |
| 7107 |
end = last_cluster; |
| 7108 |
if (grp->bb_free >= minlen) { |
| 7109 |
cnt = ext4_trim_all_free(sb, group, first_cluster, |
| 7110 |
end, minlen); |
| 7111 |
if (cnt < 0) { |
| 7112 |
ret = cnt; |
| 7113 |
break; |
| 7114 |
} |
| 7115 |
trimmed += cnt; |
| 7116 |
} |
| 7117 |
|
| 7118 |
/* |
| 7119 |
* For every group except the first one, we are sure |
| 7120 |
* that the first cluster to discard will be cluster #0. |
| 7121 |
*/ |
| 7122 |
first_cluster = 0; |
| 7123 |
} |
| 7124 |
|
| 7125 |
if (!ret) |
| 7126 |
EXT4_SB(sb)->s_last_trim_minblks = minlen; |
| 7127 |
|
| 7128 |
out: |
| 7129 |
range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
| 7130 |
return ret; |
| 7131 |
} |
| 7132 |
|
| 7133 |
/* Iterate all the free extents in the group. */ |
| 7134 |
int |
| 7135 |
ext4_mballoc_query_range( |
| 7136 |
struct super_block *sb, |
| 7137 |
ext4_group_t group, |
| 7138 |
ext4_grpblk_t first, |
| 7139 |
ext4_grpblk_t end, |
| 7140 |
ext4_mballoc_query_range_fn meta_formatter, |
| 7141 |
ext4_mballoc_query_range_fn formatter, |
| 7142 |
void *priv) |
| 7143 |
{ |
| 7144 |
void *bitmap; |
| 7145 |
ext4_grpblk_t start, next; |
| 7146 |
struct ext4_buddy e4b; |
| 7147 |
int error; |
| 7148 |
|
| 7149 |
error = ext4_mb_load_buddy(sb, group, &e4b); |
| 7150 |
if (error) |
| 7151 |
return error; |
| 7152 |
bitmap = e4b.bd_bitmap; |
| 7153 |
|
| 7154 |
ext4_lock_group(sb, group); |
| 7155 |
|
| 7156 |
start = max(e4b.bd_info->bb_first_free, first); |
| 7157 |
if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) |
| 7158 |
end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; |
| 7159 |
if (meta_formatter && start != first) { |
| 7160 |
if (start > end) |
| 7161 |
start = end; |
| 7162 |
ext4_unlock_group(sb, group); |
| 7163 |
error = meta_formatter(sb, group, first, start - first, |
| 7164 |
priv); |
| 7165 |
if (error) |
| 7166 |
goto out_unload; |
| 7167 |
ext4_lock_group(sb, group); |
| 7168 |
} |
| 7169 |
while (start <= end) { |
| 7170 |
start = mb_find_next_zero_bit(bitmap, end + 1, start); |
| 7171 |
if (start > end) |
| 7172 |
break; |
| 7173 |
next = mb_find_next_bit(bitmap, end + 1, start); |
| 7174 |
|
| 7175 |
ext4_unlock_group(sb, group); |
| 7176 |
error = formatter(sb, group, start, next - start, priv); |
| 7177 |
if (error) |
| 7178 |
goto out_unload; |
| 7179 |
ext4_lock_group(sb, group); |
| 7180 |
|
| 7181 |
start = next + 1; |
| 7182 |
} |
| 7183 |
|
| 7184 |
ext4_unlock_group(sb, group); |
| 7185 |
out_unload: |
| 7186 |
ext4_mb_unload_buddy(&e4b); |
| 7187 |
|
| 7188 |
return error; |
| 7189 |
} |
| 7190 |
|
| 7191 |
#if IS_ENABLED(CONFIG_EXT4_KUNIT_TESTS) |
| 7192 |
void mb_clear_bits_test(void *bm, int cur, int len) |
| 7193 |
{ |
| 7194 |
mb_clear_bits(bm, cur, len); |
| 7195 |
} |
| 7196 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_clear_bits_test); |
| 7197 |
|
| 7198 |
ext4_fsblk_t |
| 7199 |
ext4_mb_new_blocks_simple_test(struct ext4_allocation_request *ar, |
| 7200 |
int *errp) |
| 7201 |
{ |
| 7202 |
return ext4_mb_new_blocks_simple(ar, errp); |
| 7203 |
} |
| 7204 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_new_blocks_simple_test); |
| 7205 |
|
| 7206 |
int mb_find_next_zero_bit_test(void *addr, int max, int start) |
| 7207 |
{ |
| 7208 |
return mb_find_next_zero_bit(addr, max, start); |
| 7209 |
} |
| 7210 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_find_next_zero_bit_test); |
| 7211 |
|
| 7212 |
int mb_find_next_bit_test(void *addr, int max, int start) |
| 7213 |
{ |
| 7214 |
return mb_find_next_bit(addr, max, start); |
| 7215 |
} |
| 7216 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_find_next_bit_test); |
| 7217 |
|
| 7218 |
void mb_clear_bit_test(int bit, void *addr) |
| 7219 |
{ |
| 7220 |
mb_clear_bit(bit, addr); |
| 7221 |
} |
| 7222 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_clear_bit_test); |
| 7223 |
|
| 7224 |
int mb_test_bit_test(int bit, void *addr) |
| 7225 |
{ |
| 7226 |
return mb_test_bit(bit, addr); |
| 7227 |
} |
| 7228 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_test_bit_test); |
| 7229 |
|
| 7230 |
int ext4_mb_mark_diskspace_used_test(struct ext4_allocation_context *ac, |
| 7231 |
handle_t *handle) |
| 7232 |
{ |
| 7233 |
return ext4_mb_mark_diskspace_used(ac, handle); |
| 7234 |
} |
| 7235 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_mark_diskspace_used_test); |
| 7236 |
|
| 7237 |
int mb_mark_used_test(struct ext4_buddy *e4b, struct ext4_free_extent *ex) |
| 7238 |
{ |
| 7239 |
return mb_mark_used(e4b, ex); |
| 7240 |
} |
| 7241 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_mark_used_test); |
| 7242 |
|
| 7243 |
void ext4_mb_generate_buddy_test(struct super_block *sb, void *buddy, |
| 7244 |
void *bitmap, ext4_group_t group, |
| 7245 |
struct ext4_group_info *grp) |
| 7246 |
{ |
| 7247 |
ext4_mb_generate_buddy(sb, buddy, bitmap, group, grp); |
| 7248 |
} |
| 7249 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_generate_buddy_test); |
| 7250 |
|
| 7251 |
int ext4_mb_load_buddy_test(struct super_block *sb, ext4_group_t group, |
| 7252 |
struct ext4_buddy *e4b) |
| 7253 |
{ |
| 7254 |
return ext4_mb_load_buddy(sb, group, e4b); |
| 7255 |
} |
| 7256 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_load_buddy_test); |
| 7257 |
|
| 7258 |
void ext4_mb_unload_buddy_test(struct ext4_buddy *e4b) |
| 7259 |
{ |
| 7260 |
ext4_mb_unload_buddy(e4b); |
| 7261 |
} |
| 7262 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_unload_buddy_test); |
| 7263 |
|
| 7264 |
void mb_free_blocks_test(struct inode *inode, struct ext4_buddy *e4b, |
| 7265 |
int first, int count) |
| 7266 |
{ |
| 7267 |
mb_free_blocks(inode, e4b, first, count); |
| 7268 |
} |
| 7269 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_free_blocks_test); |
| 7270 |
|
| 7271 |
void ext4_free_blocks_simple_test(struct inode *inode, ext4_fsblk_t block, |
| 7272 |
unsigned long count) |
| 7273 |
{ |
| 7274 |
return ext4_free_blocks_simple(inode, block, count); |
| 7275 |
} |
| 7276 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_free_blocks_simple_test); |
| 7277 |
|
| 7278 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_wait_block_bitmap); |
| 7279 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_init); |
| 7280 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_get_group_desc); |
| 7281 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_count_free_clusters); |
| 7282 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_get_group_info); |
| 7283 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_free_group_clusters_set); |
| 7284 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_release); |
| 7285 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_read_block_bitmap_nowait); |
| 7286 |
EXPORT_SYMBOL_FOR_EXT4_TEST(mb_set_bits); |
| 7287 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_fc_init_inode); |
| 7288 |
EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_mark_context); |
| 7289 |
#endif |