| 1 | |
| 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 | |
| 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 |
|
| 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 |
|
| 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 |
|
| 823 | max = ffs(first | border) - 1;
|
| 824 |
|
| 825 |
|
| 826 | min = fls(len) - 1;
|
| 827 |
|
| 828 | if (max < min)
|
| 829 | min = max;
|
| 830 | chunk = 1 << min;
|
| 831 |
|
| 832 |
|
| 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 |
|
| 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 |
|
| 901 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
|
| 902 | ngroups = EXT4_SB(ac->ac_sb)->s_blockfile_groups;
|
| 903 |
|
| 904 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 | |
| 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 | |
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 1458 | continue;
|
| 1459 |
|
| 1460 | if (!buffer_verified(bh[group - first_group]))
|
| 1461 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 1510 | ext4_lock_group(sb, group);
|
| 1511 | memcpy(data, bitmap, blocksize);
|
| 1512 |
|
| 1513 |
|
| 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 | |
| 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 |
|
| 1573 | return 0;
|
| 1574 | }
|
| 1575 |
|
| 1576 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 1772 | if (folio_contains(folio, pnum)) {
|
| 1773 | folio_get(folio);
|
| 1774 | goto update_buddy;
|
| 1775 | }
|
| 1776 |
|
| 1777 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 | |
| 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 |
|
| 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 |
|
| 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 | |
| 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 | |
| 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 |
|
| 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 | |
| 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 |
|
| 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 | |
| 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 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 2181 | while (len) {
|
| 2182 | ord = mb_find_order_for_block(e4b, start);
|
| 2183 |
|
| 2184 | if (((start >> ord) << ord) == start && len >= (1 << ord)) {
|
| 2185 |
|
| 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 |
|
| 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 |
|
| 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 |
|
| 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 | |
| 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 |
|
| 2270 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
|
| 2271 | int hash = (unsigned int)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 | |
| 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 | |
| 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;
|
| 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 |
|
| 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 | |
| 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;
|
| 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 |
|
| 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;
|
| 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 |
|
| 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 |
|
| 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_info *grp = ext4_get_group_info(sb, group);
|
| 2865 |
|
| 2866 | |
| 2867 | * Prefetch block groups with free blocks; but don't
|
| 2868 | * bother if it is marked uninitialized on disk, since
|
| 2869 | * it won't require I/O to read. Also only try to
|
| 2870 | * prefetch once, so we avoid getblk() call, which can
|
| 2871 | * be expensive.
|
| 2872 |
|
| 2873 | if (grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
|
| 2874 | EXT4_MB_GRP_NEED_INIT(grp)) {
|
| 2875 | struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL);
|
| 2876 |
|
| 2877 | if (gdp && ext4_free_group_clusters(sb, gdp) > 0) {
|
| 2878 | bh = ext4_read_block_bitmap_nowait(sb, group, true);
|
| 2879 | if (!IS_ERR_OR_NULL(bh)) {
|
| 2880 | if (!buffer_uptodate(bh) && cnt)
|
| 2881 | (*cnt)++;
|
| 2882 | brelse(bh);
|
| 2883 | }
|
| 2884 | }
|
| 2885 | }
|
| 2886 | if (++group >= ngroups)
|
| 2887 | group = 0;
|
| 2888 | }
|
| 2889 | blk_finish_plug(&plug);
|
| 2890 | return group;
|
| 2891 | }
|
| 2892 |
|
| 2893 | |
| 2894 | * Batch reads of the block allocation bitmaps to get
|
| 2895 | * multiple READs in flight; limit prefetching at inexpensive
|
| 2896 | * CR, otherwise mballoc can spend a lot of time loading
|
| 2897 | * imperfect groups
|
| 2898 |
|
| 2899 | static void ext4_mb_might_prefetch(struct ext4_allocation_context *ac,
|
| 2900 | ext4_group_t group)
|
| 2901 | {
|
| 2902 | struct ext4_sb_info *sbi;
|
| 2903 |
|
| 2904 | if (ac->ac_prefetch_grp != group)
|
| 2905 | return;
|
| 2906 |
|
| 2907 | sbi = EXT4_SB(ac->ac_sb);
|
| 2908 | if (ext4_mb_cr_expensive(ac->ac_criteria) ||
|
| 2909 | ac->ac_prefetch_ios < sbi->s_mb_prefetch_limit) {
|
| 2910 | unsigned int nr = sbi->s_mb_prefetch;
|
| 2911 |
|
| 2912 | if (ext4_has_feature_flex_bg(ac->ac_sb)) {
|
| 2913 | nr = 1 << sbi->s_log_groups_per_flex;
|
| 2914 | nr -= group & (nr - 1);
|
| 2915 | nr = umin(nr, sbi->s_mb_prefetch);
|
| 2916 | }
|
| 2917 |
|
| 2918 | ac->ac_prefetch_nr = nr;
|
| 2919 | ac->ac_prefetch_grp = ext4_mb_prefetch(ac->ac_sb, group, nr,
|
| 2920 | &ac->ac_prefetch_ios);
|
| 2921 | }
|
| 2922 | }
|
| 2923 |
|
| 2924 | |
| 2925 | * Prefetching reads the block bitmap into the buffer cache; but we
|
| 2926 | * need to make sure that the buddy bitmap in the page cache has been
|
| 2927 | * initialized. Note that ext4_mb_init_group() will block if the I/O
|
| 2928 | * is not yet completed, or indeed if it was not initiated by
|
| 2929 | * ext4_mb_prefetch did not start the I/O.
|
| 2930 | *
|
| 2931 | * TODO: We should actually kick off the buddy bitmap setup in a work
|
| 2932 | * queue when the buffer I/O is completed, so that we don't block
|
| 2933 | * waiting for the block allocation bitmap read to finish when
|
| 2934 | * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
|
| 2935 |
|
| 2936 | void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
|
| 2937 | unsigned int nr)
|
| 2938 | {
|
| 2939 | struct ext4_group_desc *gdp;
|
| 2940 | struct ext4_group_info *grp;
|
| 2941 |
|
| 2942 | while (nr-- > 0) {
|
| 2943 | if (!group)
|
| 2944 | group = ext4_get_groups_count(sb);
|
| 2945 | group--;
|
| 2946 | gdp = ext4_get_group_desc(sb, group, NULL);
|
| 2947 | grp = ext4_get_group_info(sb, group);
|
| 2948 |
|
| 2949 | if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) &&
|
| 2950 | ext4_free_group_clusters(sb, gdp) > 0) {
|
| 2951 | if (ext4_mb_init_group(sb, group, GFP_NOFS))
|
| 2952 | break;
|
| 2953 | }
|
| 2954 | }
|
| 2955 | }
|
| 2956 |
|
| 2957 | static int ext4_mb_scan_group(struct ext4_allocation_context *ac,
|
| 2958 | ext4_group_t group)
|
| 2959 | {
|
| 2960 | int ret;
|
| 2961 | struct super_block *sb = ac->ac_sb;
|
| 2962 | enum criteria cr = ac->ac_criteria;
|
| 2963 |
|
| 2964 | ext4_mb_might_prefetch(ac, group);
|
| 2965 |
|
| 2966 |
|
| 2967 | if (cr < CR_ANY_FREE && spin_is_locked(ext4_group_lock_ptr(sb, group)))
|
| 2968 | return 0;
|
| 2969 |
|
| 2970 |
|
| 2971 | ret = ext4_mb_good_group_nolock(ac, group, cr);
|
| 2972 | if (ret <= 0) {
|
| 2973 | if (!ac->ac_first_err)
|
| 2974 | ac->ac_first_err = ret;
|
| 2975 | return 0;
|
| 2976 | }
|
| 2977 |
|
| 2978 | ret = ext4_mb_load_buddy(sb, group, ac->ac_e4b);
|
| 2979 | if (ret)
|
| 2980 | return ret;
|
| 2981 |
|
| 2982 |
|
| 2983 | if (cr >= CR_ANY_FREE)
|
| 2984 | ext4_lock_group(sb, group);
|
| 2985 | else if (!ext4_try_lock_group(sb, group))
|
| 2986 | goto out_unload;
|
| 2987 |
|
| 2988 |
|
| 2989 | if (unlikely(!ext4_mb_good_group(ac, group, cr)))
|
| 2990 | goto out_unlock;
|
| 2991 |
|
| 2992 | __ext4_mb_scan_group(ac);
|
| 2993 |
|
| 2994 | out_unlock:
|
| 2995 | ext4_unlock_group(sb, group);
|
| 2996 | out_unload:
|
| 2997 | ext4_mb_unload_buddy(ac->ac_e4b);
|
| 2998 | return ret;
|
| 2999 | }
|
| 3000 |
|
| 3001 | static noinline_for_stack int
|
| 3002 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
|
| 3003 | {
|
| 3004 | ext4_group_t i;
|
| 3005 | int err = 0;
|
| 3006 | struct super_block *sb = ac->ac_sb;
|
| 3007 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3008 | struct ext4_buddy e4b;
|
| 3009 |
|
| 3010 | BUG_ON(ac->ac_status == AC_STATUS_FOUND);
|
| 3011 |
|
| 3012 |
|
| 3013 | err = ext4_mb_find_by_goal(ac, &e4b);
|
| 3014 | if (err || ac->ac_status == AC_STATUS_FOUND)
|
| 3015 | goto out;
|
| 3016 |
|
| 3017 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
|
| 3018 | goto out;
|
| 3019 |
|
| 3020 | |
| 3021 | * ac->ac_2order is set only if the fe_len is a power of 2
|
| 3022 | * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED
|
| 3023 | * so that we try exact allocation using buddy.
|
| 3024 |
|
| 3025 | i = fls(ac->ac_g_ex.fe_len);
|
| 3026 | ac->ac_2order = 0;
|
| 3027 | |
| 3028 | * We search using buddy data only if the order of the request
|
| 3029 | * is greater than equal to the sbi_s_mb_order2_reqs
|
| 3030 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
|
| 3031 | * We also support searching for power-of-two requests only for
|
| 3032 | * requests upto maximum buddy size we have constructed.
|
| 3033 |
|
| 3034 | if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
|
| 3035 | if (is_power_of_2(ac->ac_g_ex.fe_len))
|
| 3036 | ac->ac_2order = array_index_nospec(i - 1,
|
| 3037 | MB_NUM_ORDERS(sb));
|
| 3038 | }
|
| 3039 |
|
| 3040 |
|
| 3041 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
|
| 3042 | int hash = (unsigned int)ac->ac_inode->i_ino % sbi->s_mb_nr_global_goals;
|
| 3043 |
|
| 3044 | ac->ac_g_ex.fe_group = READ_ONCE(sbi->s_mb_last_groups[hash]);
|
| 3045 | ac->ac_g_ex.fe_start = -1;
|
| 3046 | ac->ac_flags &= ~EXT4_MB_HINT_TRY_GOAL;
|
| 3047 | }
|
| 3048 |
|
| 3049 | |
| 3050 | * Let's just scan groups to find more-less suitable blocks We
|
| 3051 | * start with CR_GOAL_LEN_FAST, unless it is power of 2
|
| 3052 | * aligned, in which case let's do that faster approach first.
|
| 3053 |
|
| 3054 | ac->ac_criteria = CR_GOAL_LEN_FAST;
|
| 3055 | if (ac->ac_2order)
|
| 3056 | ac->ac_criteria = CR_POWER2_ALIGNED;
|
| 3057 |
|
| 3058 | ac->ac_e4b = &e4b;
|
| 3059 | ac->ac_prefetch_ios = 0;
|
| 3060 | ac->ac_first_err = 0;
|
| 3061 | repeat:
|
| 3062 | while (ac->ac_criteria < EXT4_MB_NUM_CRS) {
|
| 3063 | err = ext4_mb_scan_groups(ac);
|
| 3064 | if (err)
|
| 3065 | goto out;
|
| 3066 |
|
| 3067 | if (ac->ac_status != AC_STATUS_CONTINUE)
|
| 3068 | break;
|
| 3069 | }
|
| 3070 |
|
| 3071 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
|
| 3072 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
|
| 3073 | |
| 3074 | * We've been searching too long. Let's try to allocate
|
| 3075 | * the best chunk we've found so far
|
| 3076 |
|
| 3077 | ext4_mb_try_best_found(ac, &e4b);
|
| 3078 | if (ac->ac_status != AC_STATUS_FOUND) {
|
| 3079 | int lost;
|
| 3080 |
|
| 3081 | |
| 3082 | * Someone more lucky has already allocated it.
|
| 3083 | * The only thing we can do is just take first
|
| 3084 | * found block(s)
|
| 3085 |
|
| 3086 | lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
|
| 3087 | mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
|
| 3088 | ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
|
| 3089 | ac->ac_b_ex.fe_len, lost);
|
| 3090 |
|
| 3091 | ac->ac_b_ex.fe_group = 0;
|
| 3092 | ac->ac_b_ex.fe_start = 0;
|
| 3093 | ac->ac_b_ex.fe_len = 0;
|
| 3094 | ac->ac_status = AC_STATUS_CONTINUE;
|
| 3095 | ac->ac_flags |= EXT4_MB_HINT_FIRST;
|
| 3096 | ac->ac_criteria = CR_ANY_FREE;
|
| 3097 | goto repeat;
|
| 3098 | }
|
| 3099 | }
|
| 3100 |
|
| 3101 | if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) {
|
| 3102 | atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
|
| 3103 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC &&
|
| 3104 | ac->ac_b_ex.fe_group == ac->ac_g_ex.fe_group)
|
| 3105 | atomic_inc(&sbi->s_bal_stream_goals);
|
| 3106 | }
|
| 3107 | out:
|
| 3108 | if (!err && ac->ac_status != AC_STATUS_FOUND && ac->ac_first_err)
|
| 3109 | err = ac->ac_first_err;
|
| 3110 |
|
| 3111 | mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
|
| 3112 | ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
|
| 3113 | ac->ac_flags, ac->ac_criteria, err);
|
| 3114 |
|
| 3115 | if (ac->ac_prefetch_nr)
|
| 3116 | ext4_mb_prefetch_fini(sb, ac->ac_prefetch_grp, ac->ac_prefetch_nr);
|
| 3117 |
|
| 3118 | return err;
|
| 3119 | }
|
| 3120 |
|
| 3121 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
|
| 3122 | {
|
| 3123 | struct super_block *sb = pde_data(file_inode(seq->file));
|
| 3124 | ext4_group_t group;
|
| 3125 |
|
| 3126 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
|
| 3127 | return NULL;
|
| 3128 | group = *pos + 1;
|
| 3129 | return (void *) ((unsigned long) group);
|
| 3130 | }
|
| 3131 |
|
| 3132 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
|
| 3133 | {
|
| 3134 | struct super_block *sb = pde_data(file_inode(seq->file));
|
| 3135 | ext4_group_t group;
|
| 3136 |
|
| 3137 | ++*pos;
|
| 3138 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
|
| 3139 | return NULL;
|
| 3140 | group = *pos + 1;
|
| 3141 | return (void *) ((unsigned long) group);
|
| 3142 | }
|
| 3143 |
|
| 3144 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
|
| 3145 | {
|
| 3146 | struct super_block *sb = pde_data(file_inode(seq->file));
|
| 3147 | ext4_group_t group = (ext4_group_t) ((unsigned long) v);
|
| 3148 | int i, err;
|
| 3149 | char nbuf[16];
|
| 3150 | struct ext4_buddy e4b;
|
| 3151 | struct ext4_group_info *grinfo;
|
| 3152 | unsigned char blocksize_bits = min_t(unsigned char,
|
| 3153 | sb->s_blocksize_bits,
|
| 3154 | EXT4_MAX_BLOCK_LOG_SIZE);
|
| 3155 | DEFINE_RAW_FLEX(struct ext4_group_info, sg, bb_counters,
|
| 3156 | EXT4_MAX_BLOCK_LOG_SIZE + 2);
|
| 3157 |
|
| 3158 | group--;
|
| 3159 | if (group == 0)
|
| 3160 | seq_puts(seq, "#group: free frags first ["
|
| 3161 | " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
|
| 3162 | " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
|
| 3163 |
|
| 3164 | i = (blocksize_bits + 2) * sizeof(sg->bb_counters[0]) +
|
| 3165 | sizeof(struct ext4_group_info);
|
| 3166 |
|
| 3167 | grinfo = ext4_get_group_info(sb, group);
|
| 3168 | if (!grinfo)
|
| 3169 | return 0;
|
| 3170 |
|
| 3171 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
|
| 3172 | err = ext4_mb_load_buddy(sb, group, &e4b);
|
| 3173 | if (err) {
|
| 3174 | seq_printf(seq, "#%-5u: %s\n", group, ext4_decode_error(NULL, err, nbuf));
|
| 3175 | return 0;
|
| 3176 | }
|
| 3177 | ext4_mb_unload_buddy(&e4b);
|
| 3178 | }
|
| 3179 |
|
| 3180 | |
| 3181 | * We care only about free space counters in the group info and
|
| 3182 | * these are safe to access even after the buddy has been unloaded
|
| 3183 |
|
| 3184 | memcpy(sg, grinfo, i);
|
| 3185 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg->bb_free,
|
| 3186 | sg->bb_fragments, sg->bb_first_free);
|
| 3187 | for (i = 0; i <= 13; i++)
|
| 3188 | seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
|
| 3189 | sg->bb_counters[i] : 0);
|
| 3190 | seq_puts(seq, " ]");
|
| 3191 | if (EXT4_MB_GRP_BBITMAP_CORRUPT(sg))
|
| 3192 | seq_puts(seq, " Block bitmap corrupted!");
|
| 3193 | seq_putc(seq, '\n');
|
| 3194 | return 0;
|
| 3195 | }
|
| 3196 |
|
| 3197 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
|
| 3198 | {
|
| 3199 | }
|
| 3200 |
|
| 3201 | const struct seq_operations ext4_mb_seq_groups_ops = {
|
| 3202 | .start = ext4_mb_seq_groups_start,
|
| 3203 | .next = ext4_mb_seq_groups_next,
|
| 3204 | .stop = ext4_mb_seq_groups_stop,
|
| 3205 | .show = ext4_mb_seq_groups_show,
|
| 3206 | };
|
| 3207 |
|
| 3208 | int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
|
| 3209 | {
|
| 3210 | struct super_block *sb = seq->private;
|
| 3211 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3212 |
|
| 3213 | seq_puts(seq, "mballoc:\n");
|
| 3214 | if (!sbi->s_mb_stats) {
|
| 3215 | seq_puts(seq, "\tmb stats collection turned off.\n");
|
| 3216 | seq_puts(
|
| 3217 | seq,
|
| 3218 | "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
|
| 3219 | return 0;
|
| 3220 | }
|
| 3221 | seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
|
| 3222 | seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
|
| 3223 |
|
| 3224 | seq_printf(seq, "\tgroups_scanned: %u\n",
|
| 3225 | atomic_read(&sbi->s_bal_groups_scanned));
|
| 3226 |
|
| 3227 |
|
| 3228 | seq_puts(seq, "\tcr_p2_aligned_stats:\n");
|
| 3229 | seq_printf(seq, "\t\thits: %llu\n",
|
| 3230 | atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED]));
|
| 3231 | seq_printf(
|
| 3232 | seq, "\t\tgroups_considered: %llu\n",
|
| 3233 | atomic64_read(
|
| 3234 | &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]));
|
| 3235 | seq_printf(seq, "\t\textents_scanned: %u\n",
|
| 3236 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED]));
|
| 3237 | seq_printf(seq, "\t\tuseless_loops: %llu\n",
|
| 3238 | atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED]));
|
| 3239 |
|
| 3240 |
|
| 3241 | seq_puts(seq, "\tcr_goal_fast_stats:\n");
|
| 3242 | seq_printf(seq, "\t\thits: %llu\n",
|
| 3243 | atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST]));
|
| 3244 | seq_printf(seq, "\t\tgroups_considered: %llu\n",
|
| 3245 | atomic64_read(
|
| 3246 | &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST]));
|
| 3247 | seq_printf(seq, "\t\textents_scanned: %u\n",
|
| 3248 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST]));
|
| 3249 | seq_printf(seq, "\t\tuseless_loops: %llu\n",
|
| 3250 | atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST]));
|
| 3251 |
|
| 3252 |
|
| 3253 | seq_puts(seq, "\tcr_best_avail_stats:\n");
|
| 3254 | seq_printf(seq, "\t\thits: %llu\n",
|
| 3255 | atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN]));
|
| 3256 | seq_printf(
|
| 3257 | seq, "\t\tgroups_considered: %llu\n",
|
| 3258 | atomic64_read(
|
| 3259 | &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN]));
|
| 3260 | seq_printf(seq, "\t\textents_scanned: %u\n",
|
| 3261 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN]));
|
| 3262 | seq_printf(seq, "\t\tuseless_loops: %llu\n",
|
| 3263 | atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN]));
|
| 3264 |
|
| 3265 |
|
| 3266 | seq_puts(seq, "\tcr_goal_slow_stats:\n");
|
| 3267 | seq_printf(seq, "\t\thits: %llu\n",
|
| 3268 | atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW]));
|
| 3269 | seq_printf(seq, "\t\tgroups_considered: %llu\n",
|
| 3270 | atomic64_read(
|
| 3271 | &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW]));
|
| 3272 | seq_printf(seq, "\t\textents_scanned: %u\n",
|
| 3273 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW]));
|
| 3274 | seq_printf(seq, "\t\tuseless_loops: %llu\n",
|
| 3275 | atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW]));
|
| 3276 |
|
| 3277 |
|
| 3278 | seq_puts(seq, "\tcr_any_free_stats:\n");
|
| 3279 | seq_printf(seq, "\t\thits: %llu\n",
|
| 3280 | atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE]));
|
| 3281 | seq_printf(
|
| 3282 | seq, "\t\tgroups_considered: %llu\n",
|
| 3283 | atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE]));
|
| 3284 | seq_printf(seq, "\t\textents_scanned: %u\n",
|
| 3285 | atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE]));
|
| 3286 | seq_printf(seq, "\t\tuseless_loops: %llu\n",
|
| 3287 | atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE]));
|
| 3288 |
|
| 3289 |
|
| 3290 | seq_printf(seq, "\textents_scanned: %u\n",
|
| 3291 | atomic_read(&sbi->s_bal_ex_scanned));
|
| 3292 | seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
|
| 3293 | seq_printf(seq, "\t\tstream_goal_hits: %u\n",
|
| 3294 | atomic_read(&sbi->s_bal_stream_goals));
|
| 3295 | seq_printf(seq, "\t\tlen_goal_hits: %u\n",
|
| 3296 | atomic_read(&sbi->s_bal_len_goals));
|
| 3297 | seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
|
| 3298 | seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
|
| 3299 | seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
|
| 3300 | seq_printf(seq, "\tbuddies_generated: %u/%u\n",
|
| 3301 | atomic_read(&sbi->s_mb_buddies_generated),
|
| 3302 | ext4_get_groups_count(sb));
|
| 3303 | seq_printf(seq, "\tbuddies_time_used: %llu\n",
|
| 3304 | atomic64_read(&sbi->s_mb_generation_time));
|
| 3305 | seq_printf(seq, "\tpreallocated: %u\n",
|
| 3306 | atomic_read(&sbi->s_mb_preallocated));
|
| 3307 | seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded));
|
| 3308 | return 0;
|
| 3309 | }
|
| 3310 |
|
| 3311 | static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
|
| 3312 | {
|
| 3313 | struct super_block *sb = pde_data(file_inode(seq->file));
|
| 3314 | unsigned long position;
|
| 3315 |
|
| 3316 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
|
| 3317 | return NULL;
|
| 3318 | position = *pos + 1;
|
| 3319 | return (void *) ((unsigned long) position);
|
| 3320 | }
|
| 3321 |
|
| 3322 | static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
|
| 3323 | {
|
| 3324 | struct super_block *sb = pde_data(file_inode(seq->file));
|
| 3325 | unsigned long position;
|
| 3326 |
|
| 3327 | ++*pos;
|
| 3328 | if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
|
| 3329 | return NULL;
|
| 3330 | position = *pos + 1;
|
| 3331 | return (void *) ((unsigned long) position);
|
| 3332 | }
|
| 3333 |
|
| 3334 | static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
|
| 3335 | {
|
| 3336 | struct super_block *sb = pde_data(file_inode(seq->file));
|
| 3337 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3338 | unsigned long position = ((unsigned long) v);
|
| 3339 | struct ext4_group_info *grp;
|
| 3340 | unsigned int count;
|
| 3341 | unsigned long idx;
|
| 3342 |
|
| 3343 | position--;
|
| 3344 | if (position >= MB_NUM_ORDERS(sb)) {
|
| 3345 | position -= MB_NUM_ORDERS(sb);
|
| 3346 | if (position == 0)
|
| 3347 | seq_puts(seq, "avg_fragment_size_lists:\n");
|
| 3348 |
|
| 3349 | count = 0;
|
| 3350 | xa_for_each(&sbi->s_mb_avg_fragment_size[position], idx, grp)
|
| 3351 | count++;
|
| 3352 | seq_printf(seq, "\tlist_order_%u_groups: %u\n",
|
| 3353 | (unsigned int)position, count);
|
| 3354 | return 0;
|
| 3355 | }
|
| 3356 |
|
| 3357 | if (position == 0) {
|
| 3358 | seq_printf(seq, "optimize_scan: %d\n",
|
| 3359 | test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
|
| 3360 | seq_puts(seq, "max_free_order_lists:\n");
|
| 3361 | }
|
| 3362 | count = 0;
|
| 3363 | xa_for_each(&sbi->s_mb_largest_free_orders[position], idx, grp)
|
| 3364 | count++;
|
| 3365 | seq_printf(seq, "\tlist_order_%u_groups: %u\n",
|
| 3366 | (unsigned int)position, count);
|
| 3367 |
|
| 3368 | return 0;
|
| 3369 | }
|
| 3370 |
|
| 3371 | static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
|
| 3372 | {
|
| 3373 | }
|
| 3374 |
|
| 3375 | const struct seq_operations ext4_mb_seq_structs_summary_ops = {
|
| 3376 | .start = ext4_mb_seq_structs_summary_start,
|
| 3377 | .next = ext4_mb_seq_structs_summary_next,
|
| 3378 | .stop = ext4_mb_seq_structs_summary_stop,
|
| 3379 | .show = ext4_mb_seq_structs_summary_show,
|
| 3380 | };
|
| 3381 |
|
| 3382 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
|
| 3383 | {
|
| 3384 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
|
| 3385 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
|
| 3386 |
|
| 3387 | BUG_ON(!cachep);
|
| 3388 | return cachep;
|
| 3389 | }
|
| 3390 |
|
| 3391 | |
| 3392 | * Allocate the top-level s_group_info array for the specified number
|
| 3393 | * of groups
|
| 3394 |
|
| 3395 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
|
| 3396 | {
|
| 3397 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3398 | unsigned size;
|
| 3399 | struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
|
| 3400 |
|
| 3401 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
|
| 3402 | EXT4_DESC_PER_BLOCK_BITS(sb);
|
| 3403 | if (size <= sbi->s_group_info_size)
|
| 3404 | return 0;
|
| 3405 |
|
| 3406 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
|
| 3407 | new_groupinfo = kvzalloc(size, GFP_KERNEL);
|
| 3408 | if (!new_groupinfo) {
|
| 3409 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
|
| 3410 | return -ENOMEM;
|
| 3411 | }
|
| 3412 | rcu_read_lock();
|
| 3413 | old_groupinfo = rcu_dereference(sbi->s_group_info);
|
| 3414 | if (old_groupinfo)
|
| 3415 | memcpy(new_groupinfo, old_groupinfo,
|
| 3416 | sbi->s_group_info_size * sizeof(*sbi->s_group_info));
|
| 3417 | rcu_read_unlock();
|
| 3418 | rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
|
| 3419 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
|
| 3420 | if (old_groupinfo)
|
| 3421 | ext4_kvfree_array_rcu(old_groupinfo);
|
| 3422 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
|
| 3423 | sbi->s_group_info_size);
|
| 3424 | return 0;
|
| 3425 | }
|
| 3426 |
|
| 3427 |
|
| 3428 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
|
| 3429 | struct ext4_group_desc *desc)
|
| 3430 | {
|
| 3431 | int i;
|
| 3432 | int metalen = 0;
|
| 3433 | int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
|
| 3434 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3435 | struct ext4_group_info **meta_group_info;
|
| 3436 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
|
| 3437 |
|
| 3438 | |
| 3439 | * First check if this group is the first of a reserved block.
|
| 3440 | * If it's true, we have to allocate a new table of pointers
|
| 3441 | * to ext4_group_info structures
|
| 3442 |
|
| 3443 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
|
| 3444 | metalen = sizeof(*meta_group_info) <<
|
| 3445 | EXT4_DESC_PER_BLOCK_BITS(sb);
|
| 3446 | meta_group_info = kmalloc(metalen, GFP_NOFS);
|
| 3447 | if (meta_group_info == NULL) {
|
| 3448 | ext4_msg(sb, KERN_ERR, "can't allocate mem "
|
| 3449 | "for a buddy group");
|
| 3450 | return -ENOMEM;
|
| 3451 | }
|
| 3452 | rcu_read_lock();
|
| 3453 | rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
|
| 3454 | rcu_read_unlock();
|
| 3455 | }
|
| 3456 |
|
| 3457 | meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
|
| 3458 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
|
| 3459 |
|
| 3460 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
|
| 3461 | if (meta_group_info[i] == NULL) {
|
| 3462 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
|
| 3463 | goto exit_group_info;
|
| 3464 | }
|
| 3465 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
|
| 3466 | &(meta_group_info[i]->bb_state));
|
| 3467 |
|
| 3468 | |
| 3469 | * initialize bb_free to be able to skip
|
| 3470 | * empty groups without initialization
|
| 3471 |
|
| 3472 | if (ext4_has_group_desc_csum(sb) &&
|
| 3473 | (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
|
| 3474 | meta_group_info[i]->bb_free =
|
| 3475 | ext4_free_clusters_after_init(sb, group, desc);
|
| 3476 | } else {
|
| 3477 | meta_group_info[i]->bb_free =
|
| 3478 | ext4_free_group_clusters(sb, desc);
|
| 3479 | }
|
| 3480 |
|
| 3481 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
|
| 3482 | init_rwsem(&meta_group_info[i]->alloc_sem);
|
| 3483 | meta_group_info[i]->bb_free_root = RB_ROOT;
|
| 3484 | meta_group_info[i]->bb_largest_free_order = -1;
|
| 3485 | meta_group_info[i]->bb_avg_fragment_size_order = -1;
|
| 3486 | meta_group_info[i]->bb_group = group;
|
| 3487 |
|
| 3488 | mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
|
| 3489 | return 0;
|
| 3490 |
|
| 3491 | exit_group_info:
|
| 3492 |
|
| 3493 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
|
| 3494 | struct ext4_group_info ***group_info;
|
| 3495 |
|
| 3496 | rcu_read_lock();
|
| 3497 | group_info = rcu_dereference(sbi->s_group_info);
|
| 3498 | kfree(group_info[idx]);
|
| 3499 | group_info[idx] = NULL;
|
| 3500 | rcu_read_unlock();
|
| 3501 | }
|
| 3502 | return -ENOMEM;
|
| 3503 | }
|
| 3504 |
|
| 3505 | static int ext4_mb_init_backend(struct super_block *sb)
|
| 3506 | {
|
| 3507 | ext4_group_t ngroups = ext4_get_groups_count(sb);
|
| 3508 | ext4_group_t i;
|
| 3509 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3510 | int err;
|
| 3511 | struct ext4_group_desc *desc;
|
| 3512 | struct ext4_group_info ***group_info;
|
| 3513 | struct kmem_cache *cachep;
|
| 3514 |
|
| 3515 | err = ext4_mb_alloc_groupinfo(sb, ngroups);
|
| 3516 | if (err)
|
| 3517 | return err;
|
| 3518 |
|
| 3519 | sbi->s_buddy_cache = new_inode(sb);
|
| 3520 | if (sbi->s_buddy_cache == NULL) {
|
| 3521 | ext4_msg(sb, KERN_ERR, "can't get new inode");
|
| 3522 | goto err_freesgi;
|
| 3523 | }
|
| 3524 | |
| 3525 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is
|
| 3526 | * not in the inode hash, so it should never be found by iget(), but
|
| 3527 | * this will avoid confusion if it ever shows up during debugging.
|
| 3528 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
|
| 3529 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
|
| 3530 | ext4_set_inode_mapping_order(sbi->s_buddy_cache);
|
| 3531 |
|
| 3532 | for (i = 0; i < ngroups; i++) {
|
| 3533 | cond_resched();
|
| 3534 | desc = ext4_get_group_desc(sb, i, NULL);
|
| 3535 | if (desc == NULL) {
|
| 3536 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
|
| 3537 | goto err_freebuddy;
|
| 3538 | }
|
| 3539 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
|
| 3540 | goto err_freebuddy;
|
| 3541 | }
|
| 3542 |
|
| 3543 | if (ext4_has_feature_flex_bg(sb)) {
|
| 3544 | |
| 3545 | * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
|
| 3546 | * unsigned integer, so the maximum shift is 32.
|
| 3547 |
|
| 3548 | if (sbi->s_es->s_log_groups_per_flex >= 32) {
|
| 3549 | ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
|
| 3550 | goto err_freebuddy;
|
| 3551 | }
|
| 3552 | sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
|
| 3553 | BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
|
| 3554 | sbi->s_mb_prefetch *= 8;
|
| 3555 | } else {
|
| 3556 | sbi->s_mb_prefetch = 32;
|
| 3557 | }
|
| 3558 | if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
|
| 3559 | sbi->s_mb_prefetch = ext4_get_groups_count(sb);
|
| 3560 | |
| 3561 | * now many real IOs to prefetch within a single allocation at
|
| 3562 | * CR_POWER2_ALIGNED. Given CR_POWER2_ALIGNED is an CPU-related
|
| 3563 | * optimization we shouldn't try to load too many groups, at some point
|
| 3564 | * we should start to use what we've got in memory.
|
| 3565 | * with an average random access time 5ms, it'd take a second to get
|
| 3566 | * 200 groups (* N with flex_bg), so let's make this limit 4
|
| 3567 |
|
| 3568 | sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
|
| 3569 | if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
|
| 3570 | sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
|
| 3571 |
|
| 3572 | return 0;
|
| 3573 |
|
| 3574 | err_freebuddy:
|
| 3575 | cachep = get_groupinfo_cache(sb->s_blocksize_bits);
|
| 3576 | while (i-- > 0) {
|
| 3577 | struct ext4_group_info *grp = ext4_get_group_info(sb, i);
|
| 3578 |
|
| 3579 | if (grp)
|
| 3580 | kmem_cache_free(cachep, grp);
|
| 3581 | }
|
| 3582 | i = sbi->s_group_info_size;
|
| 3583 | rcu_read_lock();
|
| 3584 | group_info = rcu_dereference(sbi->s_group_info);
|
| 3585 | while (i-- > 0)
|
| 3586 | kfree(group_info[i]);
|
| 3587 | rcu_read_unlock();
|
| 3588 | iput(sbi->s_buddy_cache);
|
| 3589 | err_freesgi:
|
| 3590 | kvfree(rcu_access_pointer(sbi->s_group_info));
|
| 3591 | return -ENOMEM;
|
| 3592 | }
|
| 3593 |
|
| 3594 | static void ext4_groupinfo_destroy_slabs(void)
|
| 3595 | {
|
| 3596 | int i;
|
| 3597 |
|
| 3598 | for (i = 0; i < NR_GRPINFO_CACHES; i++) {
|
| 3599 | kmem_cache_destroy(ext4_groupinfo_caches[i]);
|
| 3600 | ext4_groupinfo_caches[i] = NULL;
|
| 3601 | }
|
| 3602 | }
|
| 3603 |
|
| 3604 | static int ext4_groupinfo_create_slab(size_t size)
|
| 3605 | {
|
| 3606 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
|
| 3607 | int slab_size;
|
| 3608 | int blocksize_bits = order_base_2(size);
|
| 3609 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
|
| 3610 | struct kmem_cache *cachep;
|
| 3611 |
|
| 3612 | if (cache_index >= NR_GRPINFO_CACHES)
|
| 3613 | return -EINVAL;
|
| 3614 |
|
| 3615 | if (unlikely(cache_index < 0))
|
| 3616 | cache_index = 0;
|
| 3617 |
|
| 3618 | mutex_lock(&ext4_grpinfo_slab_create_mutex);
|
| 3619 | if (ext4_groupinfo_caches[cache_index]) {
|
| 3620 | mutex_unlock(&ext4_grpinfo_slab_create_mutex);
|
| 3621 | return 0;
|
| 3622 | }
|
| 3623 |
|
| 3624 | slab_size = offsetof(struct ext4_group_info,
|
| 3625 | bb_counters[blocksize_bits + 2]);
|
| 3626 |
|
| 3627 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
|
| 3628 | slab_size, 0, SLAB_RECLAIM_ACCOUNT,
|
| 3629 | NULL);
|
| 3630 |
|
| 3631 | ext4_groupinfo_caches[cache_index] = cachep;
|
| 3632 |
|
| 3633 | mutex_unlock(&ext4_grpinfo_slab_create_mutex);
|
| 3634 | if (!cachep) {
|
| 3635 | printk(KERN_EMERG
|
| 3636 | "EXT4-fs: no memory for groupinfo slab cache\n");
|
| 3637 | return -ENOMEM;
|
| 3638 | }
|
| 3639 |
|
| 3640 | return 0;
|
| 3641 | }
|
| 3642 |
|
| 3643 | static void ext4_discard_work(struct work_struct *work)
|
| 3644 | {
|
| 3645 | struct ext4_sb_info *sbi = container_of(work,
|
| 3646 | struct ext4_sb_info, s_discard_work);
|
| 3647 | struct super_block *sb = sbi->s_sb;
|
| 3648 | struct ext4_free_data *fd, *nfd;
|
| 3649 | struct ext4_buddy e4b;
|
| 3650 | LIST_HEAD(discard_list);
|
| 3651 | ext4_group_t grp, load_grp;
|
| 3652 | int err = 0;
|
| 3653 |
|
| 3654 | spin_lock(&sbi->s_md_lock);
|
| 3655 | list_splice_init(&sbi->s_discard_list, &discard_list);
|
| 3656 | spin_unlock(&sbi->s_md_lock);
|
| 3657 |
|
| 3658 | load_grp = UINT_MAX;
|
| 3659 | list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
|
| 3660 | |
| 3661 | * If filesystem is umounting or no memory or suffering
|
| 3662 | * from no space, give up the discard
|
| 3663 |
|
| 3664 | if ((sb->s_flags & SB_ACTIVE) && !err &&
|
| 3665 | !atomic_read(&sbi->s_retry_alloc_pending)) {
|
| 3666 | grp = fd->efd_group;
|
| 3667 | if (grp != load_grp) {
|
| 3668 | if (load_grp != UINT_MAX)
|
| 3669 | ext4_mb_unload_buddy(&e4b);
|
| 3670 |
|
| 3671 | err = ext4_mb_load_buddy(sb, grp, &e4b);
|
| 3672 | if (err) {
|
| 3673 | kmem_cache_free(ext4_free_data_cachep, fd);
|
| 3674 | load_grp = UINT_MAX;
|
| 3675 | continue;
|
| 3676 | } else {
|
| 3677 | load_grp = grp;
|
| 3678 | }
|
| 3679 | }
|
| 3680 |
|
| 3681 | ext4_lock_group(sb, grp);
|
| 3682 | ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
|
| 3683 | fd->efd_start_cluster + fd->efd_count - 1, 1);
|
| 3684 | ext4_unlock_group(sb, grp);
|
| 3685 | }
|
| 3686 | kmem_cache_free(ext4_free_data_cachep, fd);
|
| 3687 | }
|
| 3688 |
|
| 3689 | if (load_grp != UINT_MAX)
|
| 3690 | ext4_mb_unload_buddy(&e4b);
|
| 3691 | }
|
| 3692 |
|
| 3693 | static inline void ext4_mb_avg_fragment_size_destroy(struct ext4_sb_info *sbi)
|
| 3694 | {
|
| 3695 | if (!sbi->s_mb_avg_fragment_size)
|
| 3696 | return;
|
| 3697 |
|
| 3698 | for (int i = 0; i < MB_NUM_ORDERS(sbi->s_sb); i++)
|
| 3699 | xa_destroy(&sbi->s_mb_avg_fragment_size[i]);
|
| 3700 |
|
| 3701 | kfree(sbi->s_mb_avg_fragment_size);
|
| 3702 | sbi->s_mb_avg_fragment_size = NULL;
|
| 3703 | }
|
| 3704 |
|
| 3705 | static inline void ext4_mb_largest_free_orders_destroy(struct ext4_sb_info *sbi)
|
| 3706 | {
|
| 3707 | if (!sbi->s_mb_largest_free_orders)
|
| 3708 | return;
|
| 3709 |
|
| 3710 | for (int i = 0; i < MB_NUM_ORDERS(sbi->s_sb); i++)
|
| 3711 | xa_destroy(&sbi->s_mb_largest_free_orders[i]);
|
| 3712 |
|
| 3713 | kfree(sbi->s_mb_largest_free_orders);
|
| 3714 | sbi->s_mb_largest_free_orders = NULL;
|
| 3715 | }
|
| 3716 |
|
| 3717 | int ext4_mb_init(struct super_block *sb)
|
| 3718 | {
|
| 3719 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3720 | unsigned i, j;
|
| 3721 | unsigned offset, offset_incr;
|
| 3722 | unsigned max;
|
| 3723 | int ret;
|
| 3724 |
|
| 3725 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
|
| 3726 |
|
| 3727 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
|
| 3728 | if (sbi->s_mb_offsets == NULL) {
|
| 3729 | ret = -ENOMEM;
|
| 3730 | goto out;
|
| 3731 | }
|
| 3732 |
|
| 3733 | i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
|
| 3734 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
|
| 3735 | if (sbi->s_mb_maxs == NULL) {
|
| 3736 | ret = -ENOMEM;
|
| 3737 | goto out;
|
| 3738 | }
|
| 3739 |
|
| 3740 | ret = ext4_groupinfo_create_slab(sb->s_blocksize);
|
| 3741 | if (ret < 0)
|
| 3742 | goto out;
|
| 3743 |
|
| 3744 |
|
| 3745 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
|
| 3746 | sbi->s_mb_offsets[0] = 0;
|
| 3747 |
|
| 3748 | i = 1;
|
| 3749 | offset = 0;
|
| 3750 | offset_incr = 1 << (sb->s_blocksize_bits - 1);
|
| 3751 | max = sb->s_blocksize << 2;
|
| 3752 | do {
|
| 3753 | sbi->s_mb_offsets[i] = offset;
|
| 3754 | sbi->s_mb_maxs[i] = max;
|
| 3755 | offset += offset_incr;
|
| 3756 | offset_incr = offset_incr >> 1;
|
| 3757 | max = max >> 1;
|
| 3758 | i++;
|
| 3759 | } while (i < MB_NUM_ORDERS(sb));
|
| 3760 |
|
| 3761 | sbi->s_mb_avg_fragment_size =
|
| 3762 | kmalloc_objs(struct xarray, MB_NUM_ORDERS(sb));
|
| 3763 | if (!sbi->s_mb_avg_fragment_size) {
|
| 3764 | ret = -ENOMEM;
|
| 3765 | goto out;
|
| 3766 | }
|
| 3767 | for (i = 0; i < MB_NUM_ORDERS(sb); i++)
|
| 3768 | xa_init(&sbi->s_mb_avg_fragment_size[i]);
|
| 3769 |
|
| 3770 | sbi->s_mb_largest_free_orders =
|
| 3771 | kmalloc_objs(struct xarray, MB_NUM_ORDERS(sb));
|
| 3772 | if (!sbi->s_mb_largest_free_orders) {
|
| 3773 | ret = -ENOMEM;
|
| 3774 | goto out;
|
| 3775 | }
|
| 3776 | for (i = 0; i < MB_NUM_ORDERS(sb); i++)
|
| 3777 | xa_init(&sbi->s_mb_largest_free_orders[i]);
|
| 3778 |
|
| 3779 | spin_lock_init(&sbi->s_md_lock);
|
| 3780 | atomic_set(&sbi->s_mb_free_pending, 0);
|
| 3781 | INIT_LIST_HEAD(&sbi->s_freed_data_list[0]);
|
| 3782 | INIT_LIST_HEAD(&sbi->s_freed_data_list[1]);
|
| 3783 | INIT_LIST_HEAD(&sbi->s_discard_list);
|
| 3784 | INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
|
| 3785 | atomic_set(&sbi->s_retry_alloc_pending, 0);
|
| 3786 |
|
| 3787 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
|
| 3788 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
|
| 3789 | sbi->s_mb_stats = MB_DEFAULT_STATS;
|
| 3790 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
|
| 3791 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
|
| 3792 | sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER;
|
| 3793 |
|
| 3794 | |
| 3795 | * The default group preallocation is 512, which for 4k block
|
| 3796 | * sizes translates to 2 megabytes. However for bigalloc file
|
| 3797 | * systems, this is probably too big (i.e, if the cluster size
|
| 3798 | * is 1 megabyte, then group preallocation size becomes half a
|
| 3799 | * gigabyte!). As a default, we will keep a two megabyte
|
| 3800 | * group pralloc size for cluster sizes up to 64k, and after
|
| 3801 | * that, we will force a minimum group preallocation size of
|
| 3802 | * 32 clusters. This translates to 8 megs when the cluster
|
| 3803 | * size is 256k, and 32 megs when the cluster size is 1 meg,
|
| 3804 | * which seems reasonable as a default.
|
| 3805 |
|
| 3806 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
|
| 3807 | sbi->s_cluster_bits, 32);
|
| 3808 | |
| 3809 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
|
| 3810 | * to the lowest multiple of s_stripe which is bigger than
|
| 3811 | * the s_mb_group_prealloc as determined above. We want
|
| 3812 | * the preallocation size to be an exact multiple of the
|
| 3813 | * RAID stripe size so that preallocations don't fragment
|
| 3814 | * the stripes.
|
| 3815 |
|
| 3816 | if (sbi->s_stripe > 1) {
|
| 3817 | sbi->s_mb_group_prealloc = roundup(
|
| 3818 | sbi->s_mb_group_prealloc, EXT4_NUM_B2C(sbi, sbi->s_stripe));
|
| 3819 | }
|
| 3820 |
|
| 3821 | sbi->s_mb_nr_global_goals = umin(num_possible_cpus(),
|
| 3822 | DIV_ROUND_UP(sbi->s_groups_count, 4));
|
| 3823 | sbi->s_mb_last_groups = kzalloc_objs(ext4_group_t,
|
| 3824 | sbi->s_mb_nr_global_goals);
|
| 3825 | if (sbi->s_mb_last_groups == NULL) {
|
| 3826 | ret = -ENOMEM;
|
| 3827 | goto out;
|
| 3828 | }
|
| 3829 |
|
| 3830 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
|
| 3831 | if (sbi->s_locality_groups == NULL) {
|
| 3832 | ret = -ENOMEM;
|
| 3833 | goto out_free_last_groups;
|
| 3834 | }
|
| 3835 | for_each_possible_cpu(i) {
|
| 3836 | struct ext4_locality_group *lg;
|
| 3837 | lg = per_cpu_ptr(sbi->s_locality_groups, i);
|
| 3838 | mutex_init(&lg->lg_mutex);
|
| 3839 | for (j = 0; j < PREALLOC_TB_SIZE; j++)
|
| 3840 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
|
| 3841 | spin_lock_init(&lg->lg_prealloc_lock);
|
| 3842 | }
|
| 3843 |
|
| 3844 | if (!bdev_rot(sb->s_bdev))
|
| 3845 | sbi->s_mb_max_linear_groups = 0;
|
| 3846 | else
|
| 3847 | sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
|
| 3848 |
|
| 3849 | ret = ext4_mb_init_backend(sb);
|
| 3850 | if (ret != 0)
|
| 3851 | goto out_free_locality_groups;
|
| 3852 |
|
| 3853 | return 0;
|
| 3854 |
|
| 3855 | out_free_locality_groups:
|
| 3856 | free_percpu(sbi->s_locality_groups);
|
| 3857 | sbi->s_locality_groups = NULL;
|
| 3858 | out_free_last_groups:
|
| 3859 | kfree(sbi->s_mb_last_groups);
|
| 3860 | sbi->s_mb_last_groups = NULL;
|
| 3861 | out:
|
| 3862 | ext4_mb_avg_fragment_size_destroy(sbi);
|
| 3863 | ext4_mb_largest_free_orders_destroy(sbi);
|
| 3864 | kfree(sbi->s_mb_offsets);
|
| 3865 | sbi->s_mb_offsets = NULL;
|
| 3866 | kfree(sbi->s_mb_maxs);
|
| 3867 | sbi->s_mb_maxs = NULL;
|
| 3868 | return ret;
|
| 3869 | }
|
| 3870 |
|
| 3871 |
|
| 3872 | static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
|
| 3873 | {
|
| 3874 | struct ext4_prealloc_space *pa;
|
| 3875 | struct list_head *cur, *tmp;
|
| 3876 | int count = 0;
|
| 3877 |
|
| 3878 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
|
| 3879 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
|
| 3880 | list_del(&pa->pa_group_list);
|
| 3881 | count++;
|
| 3882 | kmem_cache_free(ext4_pspace_cachep, pa);
|
| 3883 | }
|
| 3884 | return count;
|
| 3885 | }
|
| 3886 |
|
| 3887 | void ext4_mb_release(struct super_block *sb)
|
| 3888 | {
|
| 3889 | ext4_group_t ngroups = ext4_get_groups_count(sb);
|
| 3890 | ext4_group_t i;
|
| 3891 | int num_meta_group_infos;
|
| 3892 | struct ext4_group_info *grinfo, ***group_info;
|
| 3893 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 3894 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
|
| 3895 | int count;
|
| 3896 |
|
| 3897 | |
| 3898 | * wait the discard work to drain all of ext4_free_data
|
| 3899 |
|
| 3900 | flush_work(&sbi->s_discard_work);
|
| 3901 | WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
|
| 3902 |
|
| 3903 | group_info = rcu_access_pointer(sbi->s_group_info);
|
| 3904 | if (group_info) {
|
| 3905 | for (i = 0; i < ngroups; i++) {
|
| 3906 | cond_resched();
|
| 3907 | grinfo = ext4_get_group_info(sb, i);
|
| 3908 | if (!grinfo)
|
| 3909 | continue;
|
| 3910 | mb_group_bb_bitmap_free(grinfo);
|
| 3911 | ext4_lock_group(sb, i);
|
| 3912 | count = ext4_mb_cleanup_pa(grinfo);
|
| 3913 | if (count)
|
| 3914 | mb_debug(sb, "mballoc: %d PAs left\n",
|
| 3915 | count);
|
| 3916 | ext4_unlock_group(sb, i);
|
| 3917 | kmem_cache_free(cachep, grinfo);
|
| 3918 | }
|
| 3919 | num_meta_group_infos = (ngroups +
|
| 3920 | EXT4_DESC_PER_BLOCK(sb) - 1) >>
|
| 3921 | EXT4_DESC_PER_BLOCK_BITS(sb);
|
| 3922 | for (i = 0; i < num_meta_group_infos; i++)
|
| 3923 | kfree(group_info[i]);
|
| 3924 | kvfree(group_info);
|
| 3925 | }
|
| 3926 | ext4_mb_avg_fragment_size_destroy(sbi);
|
| 3927 | ext4_mb_largest_free_orders_destroy(sbi);
|
| 3928 | kfree(sbi->s_mb_offsets);
|
| 3929 | kfree(sbi->s_mb_maxs);
|
| 3930 | iput(sbi->s_buddy_cache);
|
| 3931 | if (sbi->s_mb_stats) {
|
| 3932 | ext4_msg(sb, KERN_INFO,
|
| 3933 | "mballoc: %u blocks %u reqs (%u success)",
|
| 3934 | atomic_read(&sbi->s_bal_allocated),
|
| 3935 | atomic_read(&sbi->s_bal_reqs),
|
| 3936 | atomic_read(&sbi->s_bal_success));
|
| 3937 | ext4_msg(sb, KERN_INFO,
|
| 3938 | "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
|
| 3939 | "%u 2^N hits, %u breaks, %u lost",
|
| 3940 | atomic_read(&sbi->s_bal_ex_scanned),
|
| 3941 | atomic_read(&sbi->s_bal_groups_scanned),
|
| 3942 | atomic_read(&sbi->s_bal_goals),
|
| 3943 | atomic_read(&sbi->s_bal_2orders),
|
| 3944 | atomic_read(&sbi->s_bal_breaks),
|
| 3945 | atomic_read(&sbi->s_mb_lost_chunks));
|
| 3946 | ext4_msg(sb, KERN_INFO,
|
| 3947 | "mballoc: %u generated and it took %llu",
|
| 3948 | atomic_read(&sbi->s_mb_buddies_generated),
|
| 3949 | atomic64_read(&sbi->s_mb_generation_time));
|
| 3950 | ext4_msg(sb, KERN_INFO,
|
| 3951 | "mballoc: %u preallocated, %u discarded",
|
| 3952 | atomic_read(&sbi->s_mb_preallocated),
|
| 3953 | atomic_read(&sbi->s_mb_discarded));
|
| 3954 | }
|
| 3955 |
|
| 3956 | free_percpu(sbi->s_locality_groups);
|
| 3957 | kfree(sbi->s_mb_last_groups);
|
| 3958 | }
|
| 3959 |
|
| 3960 | static inline int ext4_issue_discard(struct super_block *sb,
|
| 3961 | ext4_group_t block_group, ext4_grpblk_t cluster, int count)
|
| 3962 | {
|
| 3963 | ext4_fsblk_t discard_block;
|
| 3964 |
|
| 3965 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
|
| 3966 | ext4_group_first_block_no(sb, block_group));
|
| 3967 | count = EXT4_C2B(EXT4_SB(sb), count);
|
| 3968 | trace_ext4_discard_blocks(sb,
|
| 3969 | (unsigned long long) discard_block, count);
|
| 3970 |
|
| 3971 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
|
| 3972 | }
|
| 3973 |
|
| 3974 | static void ext4_free_data_in_buddy(struct super_block *sb,
|
| 3975 | struct ext4_free_data *entry)
|
| 3976 | {
|
| 3977 | struct ext4_buddy e4b;
|
| 3978 | struct ext4_group_info *db;
|
| 3979 | int err, count = 0;
|
| 3980 |
|
| 3981 | mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
|
| 3982 | entry->efd_count, entry->efd_group, entry);
|
| 3983 |
|
| 3984 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
|
| 3985 |
|
| 3986 | BUG_ON(err != 0);
|
| 3987 |
|
| 3988 | atomic_sub(entry->efd_count, &EXT4_SB(sb)->s_mb_free_pending);
|
| 3989 | db = e4b.bd_info;
|
| 3990 |
|
| 3991 | count += entry->efd_count;
|
| 3992 | ext4_lock_group(sb, entry->efd_group);
|
| 3993 |
|
| 3994 | rb_erase(&entry->efd_node, &(db->bb_free_root));
|
| 3995 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
|
| 3996 |
|
| 3997 | |
| 3998 | * Clear the trimmed flag for the group so that the next
|
| 3999 | * ext4_trim_fs can trim it.
|
| 4000 |
|
| 4001 | EXT4_MB_GRP_CLEAR_TRIMMED(db);
|
| 4002 |
|
| 4003 | if (!db->bb_free_root.rb_node) {
|
| 4004 | |
| 4005 | * balance refcounts from ext4_mb_free_metadata()
|
| 4006 |
|
| 4007 | folio_put(e4b.bd_buddy_folio);
|
| 4008 | folio_put(e4b.bd_bitmap_folio);
|
| 4009 | }
|
| 4010 | ext4_unlock_group(sb, entry->efd_group);
|
| 4011 | ext4_mb_unload_buddy(&e4b);
|
| 4012 |
|
| 4013 | mb_debug(sb, "freed %d blocks in 1 structures\n", count);
|
| 4014 | }
|
| 4015 |
|
| 4016 | |
| 4017 | * This function is called by the jbd2 layer once the commit has finished,
|
| 4018 | * so we know we can free the blocks that were released with that commit.
|
| 4019 |
|
| 4020 | void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
|
| 4021 | {
|
| 4022 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 4023 | struct ext4_free_data *entry, *tmp;
|
| 4024 | LIST_HEAD(freed_data_list);
|
| 4025 | struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1];
|
| 4026 | bool wake;
|
| 4027 |
|
| 4028 | list_replace_init(s_freed_head, &freed_data_list);
|
| 4029 |
|
| 4030 | list_for_each_entry(entry, &freed_data_list, efd_list)
|
| 4031 | ext4_free_data_in_buddy(sb, entry);
|
| 4032 |
|
| 4033 | if (test_opt(sb, DISCARD)) {
|
| 4034 | spin_lock(&sbi->s_md_lock);
|
| 4035 | wake = list_empty(&sbi->s_discard_list);
|
| 4036 | list_splice_tail(&freed_data_list, &sbi->s_discard_list);
|
| 4037 | spin_unlock(&sbi->s_md_lock);
|
| 4038 | if (wake)
|
| 4039 | queue_work(system_dfl_wq, &sbi->s_discard_work);
|
| 4040 | } else {
|
| 4041 | list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
|
| 4042 | kmem_cache_free(ext4_free_data_cachep, entry);
|
| 4043 | }
|
| 4044 | }
|
| 4045 |
|
| 4046 | int __init ext4_init_mballoc(void)
|
| 4047 | {
|
| 4048 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
|
| 4049 | SLAB_RECLAIM_ACCOUNT);
|
| 4050 | if (ext4_pspace_cachep == NULL)
|
| 4051 | goto out;
|
| 4052 |
|
| 4053 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
|
| 4054 | SLAB_RECLAIM_ACCOUNT);
|
| 4055 | if (ext4_ac_cachep == NULL)
|
| 4056 | goto out_pa_free;
|
| 4057 |
|
| 4058 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
|
| 4059 | SLAB_RECLAIM_ACCOUNT);
|
| 4060 | if (ext4_free_data_cachep == NULL)
|
| 4061 | goto out_ac_free;
|
| 4062 |
|
| 4063 | return 0;
|
| 4064 |
|
| 4065 | out_ac_free:
|
| 4066 | kmem_cache_destroy(ext4_ac_cachep);
|
| 4067 | out_pa_free:
|
| 4068 | kmem_cache_destroy(ext4_pspace_cachep);
|
| 4069 | out:
|
| 4070 | return -ENOMEM;
|
| 4071 | }
|
| 4072 |
|
| 4073 | void ext4_exit_mballoc(void)
|
| 4074 | {
|
| 4075 | |
| 4076 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep
|
| 4077 | * before destroying the slab cache.
|
| 4078 |
|
| 4079 | rcu_barrier();
|
| 4080 | kmem_cache_destroy(ext4_pspace_cachep);
|
| 4081 | kmem_cache_destroy(ext4_ac_cachep);
|
| 4082 | kmem_cache_destroy(ext4_free_data_cachep);
|
| 4083 | ext4_groupinfo_destroy_slabs();
|
| 4084 | }
|
| 4085 |
|
| 4086 | #define EXT4_MB_BITMAP_MARKED_CHECK 0x0001
|
| 4087 | #define EXT4_MB_SYNC_UPDATE 0x0002
|
| 4088 | int
|
| 4089 | ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state,
|
| 4090 | ext4_group_t group, ext4_grpblk_t blkoff,
|
| 4091 | ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed)
|
| 4092 | {
|
| 4093 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 4094 | struct buffer_head *bitmap_bh = NULL;
|
| 4095 | struct ext4_group_desc *gdp;
|
| 4096 | struct buffer_head *gdp_bh;
|
| 4097 | int err;
|
| 4098 | unsigned int i, already, changed = len;
|
| 4099 |
|
| 4100 | KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context,
|
| 4101 | handle, sb, state, group, blkoff, len,
|
| 4102 | flags, ret_changed);
|
| 4103 |
|
| 4104 | if (ret_changed)
|
| 4105 | *ret_changed = 0;
|
| 4106 | bitmap_bh = ext4_read_block_bitmap(sb, group);
|
| 4107 | if (IS_ERR(bitmap_bh))
|
| 4108 | return PTR_ERR(bitmap_bh);
|
| 4109 |
|
| 4110 | if (handle) {
|
| 4111 | BUFFER_TRACE(bitmap_bh, "getting write access");
|
| 4112 | err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
|
| 4113 | EXT4_JTR_NONE);
|
| 4114 | if (err)
|
| 4115 | goto out_err;
|
| 4116 | }
|
| 4117 |
|
| 4118 | err = -EIO;
|
| 4119 | gdp = ext4_get_group_desc(sb, group, &gdp_bh);
|
| 4120 | if (!gdp)
|
| 4121 | goto out_err;
|
| 4122 |
|
| 4123 | if (handle) {
|
| 4124 | BUFFER_TRACE(gdp_bh, "get_write_access");
|
| 4125 | err = ext4_journal_get_write_access(handle, sb, gdp_bh,
|
| 4126 | EXT4_JTR_NONE);
|
| 4127 | if (err)
|
| 4128 | goto out_err;
|
| 4129 | }
|
| 4130 |
|
| 4131 | ext4_lock_group(sb, group);
|
| 4132 | if (ext4_has_group_desc_csum(sb) &&
|
| 4133 | (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
|
| 4134 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
|
| 4135 | ext4_free_group_clusters_set(sb, gdp,
|
| 4136 | ext4_free_clusters_after_init(sb, group, gdp));
|
| 4137 | }
|
| 4138 |
|
| 4139 | if (flags & EXT4_MB_BITMAP_MARKED_CHECK) {
|
| 4140 | already = 0;
|
| 4141 | for (i = 0; i < len; i++)
|
| 4142 | if (mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
|
| 4143 | state)
|
| 4144 | already++;
|
| 4145 | changed = len - already;
|
| 4146 | }
|
| 4147 |
|
| 4148 | if (state) {
|
| 4149 | mb_set_bits(bitmap_bh->b_data, blkoff, len);
|
| 4150 | ext4_free_group_clusters_set(sb, gdp,
|
| 4151 | ext4_free_group_clusters(sb, gdp) - changed);
|
| 4152 | } else {
|
| 4153 | mb_clear_bits(bitmap_bh->b_data, blkoff, len);
|
| 4154 | ext4_free_group_clusters_set(sb, gdp,
|
| 4155 | ext4_free_group_clusters(sb, gdp) + changed);
|
| 4156 | }
|
| 4157 |
|
| 4158 | ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh);
|
| 4159 | ext4_group_desc_csum_set(sb, group, gdp);
|
| 4160 | ext4_unlock_group(sb, group);
|
| 4161 | if (ret_changed)
|
| 4162 | *ret_changed = changed;
|
| 4163 |
|
| 4164 | if (sbi->s_log_groups_per_flex) {
|
| 4165 | ext4_group_t flex_group = ext4_flex_group(sbi, group);
|
| 4166 | struct flex_groups *fg = sbi_array_rcu_deref(sbi,
|
| 4167 | s_flex_groups, flex_group);
|
| 4168 |
|
| 4169 | if (state)
|
| 4170 | atomic64_sub(changed, &fg->free_clusters);
|
| 4171 | else
|
| 4172 | atomic64_add(changed, &fg->free_clusters);
|
| 4173 | }
|
| 4174 |
|
| 4175 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
|
| 4176 | if (err)
|
| 4177 | goto out_err;
|
| 4178 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
|
| 4179 | if (err)
|
| 4180 | goto out_err;
|
| 4181 |
|
| 4182 | if (flags & EXT4_MB_SYNC_UPDATE) {
|
| 4183 | sync_dirty_buffer(bitmap_bh);
|
| 4184 | sync_dirty_buffer(gdp_bh);
|
| 4185 | }
|
| 4186 |
|
| 4187 | out_err:
|
| 4188 | brelse(bitmap_bh);
|
| 4189 | return err;
|
| 4190 | }
|
| 4191 |
|
| 4192 | |
| 4193 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
|
| 4194 | * Returns 0 if success or error code
|
| 4195 |
|
| 4196 | static noinline_for_stack int
|
| 4197 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, handle_t *handle)
|
| 4198 | {
|
| 4199 | struct ext4_group_desc *gdp;
|
| 4200 | struct ext4_sb_info *sbi;
|
| 4201 | struct super_block *sb;
|
| 4202 | ext4_fsblk_t block;
|
| 4203 | int err, len;
|
| 4204 | int flags = 0;
|
| 4205 | ext4_grpblk_t changed;
|
| 4206 |
|
| 4207 | BUG_ON(ac->ac_status != AC_STATUS_FOUND);
|
| 4208 | BUG_ON(ac->ac_b_ex.fe_len <= 0);
|
| 4209 |
|
| 4210 | sb = ac->ac_sb;
|
| 4211 | sbi = EXT4_SB(sb);
|
| 4212 |
|
| 4213 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL);
|
| 4214 | if (!gdp)
|
| 4215 | return -EIO;
|
| 4216 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
|
| 4217 | ext4_free_group_clusters(sb, gdp));
|
| 4218 |
|
| 4219 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
|
| 4220 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
|
| 4221 | if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
|
| 4222 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
|
| 4223 | "fs metadata", block, block+len);
|
| 4224 | |
| 4225 | * Fix the bitmap and return EFSCORRUPTED
|
| 4226 | * We leak some of the blocks here.
|
| 4227 |
|
| 4228 | err = ext4_mb_mark_context(handle, sb, true,
|
| 4229 | ac->ac_b_ex.fe_group,
|
| 4230 | ac->ac_b_ex.fe_start,
|
| 4231 | ac->ac_b_ex.fe_len,
|
| 4232 | 0, NULL);
|
| 4233 | if (!err)
|
| 4234 | err = -EFSCORRUPTED;
|
| 4235 | return err;
|
| 4236 | }
|
| 4237 |
|
| 4238 | #ifdef AGGRESSIVE_CHECK
|
| 4239 | flags |= EXT4_MB_BITMAP_MARKED_CHECK;
|
| 4240 | #endif
|
| 4241 | err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group,
|
| 4242 | ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len,
|
| 4243 | flags, &changed);
|
| 4244 |
|
| 4245 | if (err && changed == 0)
|
| 4246 | return err;
|
| 4247 |
|
| 4248 | #ifdef AGGRESSIVE_CHECK
|
| 4249 | BUG_ON(changed != ac->ac_b_ex.fe_len);
|
| 4250 | #endif
|
| 4251 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
|
| 4252 |
|
| 4253 | return err;
|
| 4254 | }
|
| 4255 |
|
| 4256 | |
| 4257 | * Idempotent helper for Ext4 fast commit replay path to set the state of
|
| 4258 | * blocks in bitmaps and update counters.
|
| 4259 |
|
| 4260 | void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
|
| 4261 | int len, bool state)
|
| 4262 | {
|
| 4263 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 4264 | ext4_group_t group;
|
| 4265 | ext4_grpblk_t blkoff;
|
| 4266 | int err = 0;
|
| 4267 | unsigned int clen, thisgrp_len;
|
| 4268 |
|
| 4269 | while (len > 0) {
|
| 4270 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
|
| 4271 |
|
| 4272 | |
| 4273 | * Check to see if we are freeing blocks across a group
|
| 4274 | * boundary.
|
| 4275 | * In case of flex_bg, this can happen that (block, len) may
|
| 4276 | * span across more than one group. In that case we need to
|
| 4277 | * get the corresponding group metadata to work with.
|
| 4278 | * For this we have goto again loop.
|
| 4279 |
|
| 4280 | thisgrp_len = min(len, EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
|
| 4281 | clen = EXT4_NUM_B2C(sbi, thisgrp_len);
|
| 4282 |
|
| 4283 | if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
|
| 4284 | ext4_error(sb, "Marking blocks in system zone - "
|
| 4285 | "Block = %llu, len = %u",
|
| 4286 | block, thisgrp_len);
|
| 4287 | break;
|
| 4288 | }
|
| 4289 |
|
| 4290 | err = ext4_mb_mark_context(NULL, sb, state,
|
| 4291 | group, blkoff, clen,
|
| 4292 | EXT4_MB_BITMAP_MARKED_CHECK |
|
| 4293 | EXT4_MB_SYNC_UPDATE,
|
| 4294 | NULL);
|
| 4295 | if (err)
|
| 4296 | break;
|
| 4297 |
|
| 4298 | block += thisgrp_len;
|
| 4299 | len -= thisgrp_len;
|
| 4300 | BUG_ON(len < 0);
|
| 4301 | }
|
| 4302 | }
|
| 4303 |
|
| 4304 | |
| 4305 | * here we normalize request for locality group
|
| 4306 | * Group request are normalized to s_mb_group_prealloc, which goes to
|
| 4307 | * s_strip if we set the same via mount option.
|
| 4308 | * s_mb_group_prealloc can be configured via
|
| 4309 | * /sys/fs/ext4/<partition>/mb_group_prealloc
|
| 4310 | *
|
| 4311 | * XXX: should we try to preallocate more than the group has now?
|
| 4312 |
|
| 4313 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
|
| 4314 | {
|
| 4315 | struct super_block *sb = ac->ac_sb;
|
| 4316 | struct ext4_locality_group *lg = ac->ac_lg;
|
| 4317 |
|
| 4318 | BUG_ON(lg == NULL);
|
| 4319 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
|
| 4320 | mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
|
| 4321 | }
|
| 4322 |
|
| 4323 | |
| 4324 | * This function returns the next element to look at during inode
|
| 4325 | * PA rbtree walk. We assume that we have held the inode PA rbtree lock
|
| 4326 | * (ei->i_prealloc_lock)
|
| 4327 | *
|
| 4328 | * new_start The start of the range we want to compare
|
| 4329 | * cur_start The existing start that we are comparing against
|
| 4330 | * node The node of the rb_tree
|
| 4331 |
|
| 4332 | static inline struct rb_node*
|
| 4333 | ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node)
|
| 4334 | {
|
| 4335 | if (new_start < cur_start)
|
| 4336 | return node->rb_left;
|
| 4337 | else
|
| 4338 | return node->rb_right;
|
| 4339 | }
|
| 4340 |
|
| 4341 | static inline void
|
| 4342 | ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
|
| 4343 | ext4_lblk_t start, loff_t end)
|
| 4344 | {
|
| 4345 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4346 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
|
| 4347 | struct ext4_prealloc_space *tmp_pa;
|
| 4348 | ext4_lblk_t tmp_pa_start;
|
| 4349 | loff_t tmp_pa_end;
|
| 4350 | struct rb_node *iter;
|
| 4351 |
|
| 4352 | read_lock(&ei->i_prealloc_lock);
|
| 4353 | for (iter = ei->i_prealloc_node.rb_node; iter;
|
| 4354 | iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) {
|
| 4355 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 4356 | pa_node.inode_node);
|
| 4357 | tmp_pa_start = tmp_pa->pa_lstart;
|
| 4358 | tmp_pa_end = pa_logical_end(sbi, tmp_pa);
|
| 4359 |
|
| 4360 | spin_lock(&tmp_pa->pa_lock);
|
| 4361 | if (tmp_pa->pa_deleted == 0)
|
| 4362 | BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
|
| 4363 | spin_unlock(&tmp_pa->pa_lock);
|
| 4364 | }
|
| 4365 | read_unlock(&ei->i_prealloc_lock);
|
| 4366 | }
|
| 4367 |
|
| 4368 | |
| 4369 | * Given an allocation context "ac" and a range "start", "end", check
|
| 4370 | * and adjust boundaries if the range overlaps with any of the existing
|
| 4371 | * preallocatoins stored in the corresponding inode of the allocation context.
|
| 4372 | *
|
| 4373 | * Parameters:
|
| 4374 | * ac allocation context
|
| 4375 | * start start of the new range
|
| 4376 | * end end of the new range
|
| 4377 |
|
| 4378 | static inline void
|
| 4379 | ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
|
| 4380 | ext4_lblk_t *start, loff_t *end)
|
| 4381 | {
|
| 4382 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
|
| 4383 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4384 | struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL;
|
| 4385 | struct rb_node *iter;
|
| 4386 | ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1;
|
| 4387 | loff_t new_end, tmp_pa_end, left_pa_end = -1;
|
| 4388 |
|
| 4389 | new_start = *start;
|
| 4390 | new_end = *end;
|
| 4391 |
|
| 4392 | |
| 4393 | * Adjust the normalized range so that it doesn't overlap with any
|
| 4394 | * existing preallocated blocks(PAs). Make sure to hold the rbtree lock
|
| 4395 | * so it doesn't change underneath us.
|
| 4396 |
|
| 4397 | read_lock(&ei->i_prealloc_lock);
|
| 4398 |
|
| 4399 |
|
| 4400 | for (iter = ei->i_prealloc_node.rb_node; iter;
|
| 4401 | iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
|
| 4402 | tmp_pa_start, iter)) {
|
| 4403 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 4404 | pa_node.inode_node);
|
| 4405 | tmp_pa_start = tmp_pa->pa_lstart;
|
| 4406 | tmp_pa_end = pa_logical_end(sbi, tmp_pa);
|
| 4407 |
|
| 4408 |
|
| 4409 | spin_lock(&tmp_pa->pa_lock);
|
| 4410 | if (tmp_pa->pa_deleted == 0)
|
| 4411 | BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
|
| 4412 | ac->ac_o_ex.fe_logical < tmp_pa_start));
|
| 4413 | spin_unlock(&tmp_pa->pa_lock);
|
| 4414 | }
|
| 4415 |
|
| 4416 | |
| 4417 | * Step 2: check if the found PA is left or right neighbor and
|
| 4418 | * get the other neighbor
|
| 4419 |
|
| 4420 | if (tmp_pa) {
|
| 4421 | if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) {
|
| 4422 | struct rb_node *tmp;
|
| 4423 |
|
| 4424 | left_pa = tmp_pa;
|
| 4425 | tmp = rb_next(&left_pa->pa_node.inode_node);
|
| 4426 | if (tmp) {
|
| 4427 | right_pa = rb_entry(tmp,
|
| 4428 | struct ext4_prealloc_space,
|
| 4429 | pa_node.inode_node);
|
| 4430 | }
|
| 4431 | } else {
|
| 4432 | struct rb_node *tmp;
|
| 4433 |
|
| 4434 | right_pa = tmp_pa;
|
| 4435 | tmp = rb_prev(&right_pa->pa_node.inode_node);
|
| 4436 | if (tmp) {
|
| 4437 | left_pa = rb_entry(tmp,
|
| 4438 | struct ext4_prealloc_space,
|
| 4439 | pa_node.inode_node);
|
| 4440 | }
|
| 4441 | }
|
| 4442 | }
|
| 4443 |
|
| 4444 |
|
| 4445 | if (left_pa) {
|
| 4446 | for (iter = &left_pa->pa_node.inode_node;;
|
| 4447 | iter = rb_prev(iter)) {
|
| 4448 | if (!iter) {
|
| 4449 | left_pa = NULL;
|
| 4450 | break;
|
| 4451 | }
|
| 4452 |
|
| 4453 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 4454 | pa_node.inode_node);
|
| 4455 | left_pa = tmp_pa;
|
| 4456 | spin_lock(&tmp_pa->pa_lock);
|
| 4457 | if (tmp_pa->pa_deleted == 0) {
|
| 4458 | spin_unlock(&tmp_pa->pa_lock);
|
| 4459 | break;
|
| 4460 | }
|
| 4461 | spin_unlock(&tmp_pa->pa_lock);
|
| 4462 | }
|
| 4463 | }
|
| 4464 |
|
| 4465 | if (right_pa) {
|
| 4466 | for (iter = &right_pa->pa_node.inode_node;;
|
| 4467 | iter = rb_next(iter)) {
|
| 4468 | if (!iter) {
|
| 4469 | right_pa = NULL;
|
| 4470 | break;
|
| 4471 | }
|
| 4472 |
|
| 4473 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 4474 | pa_node.inode_node);
|
| 4475 | right_pa = tmp_pa;
|
| 4476 | spin_lock(&tmp_pa->pa_lock);
|
| 4477 | if (tmp_pa->pa_deleted == 0) {
|
| 4478 | spin_unlock(&tmp_pa->pa_lock);
|
| 4479 | break;
|
| 4480 | }
|
| 4481 | spin_unlock(&tmp_pa->pa_lock);
|
| 4482 | }
|
| 4483 | }
|
| 4484 |
|
| 4485 | if (left_pa) {
|
| 4486 | left_pa_end = pa_logical_end(sbi, left_pa);
|
| 4487 | BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical);
|
| 4488 | }
|
| 4489 |
|
| 4490 | if (right_pa) {
|
| 4491 | right_pa_start = right_pa->pa_lstart;
|
| 4492 | BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical);
|
| 4493 | }
|
| 4494 |
|
| 4495 |
|
| 4496 | if (left_pa) {
|
| 4497 | if (left_pa_end > new_start)
|
| 4498 | new_start = left_pa_end;
|
| 4499 | }
|
| 4500 |
|
| 4501 | if (right_pa) {
|
| 4502 | if (right_pa_start < new_end)
|
| 4503 | new_end = right_pa_start;
|
| 4504 | }
|
| 4505 | read_unlock(&ei->i_prealloc_lock);
|
| 4506 |
|
| 4507 |
|
| 4508 | ext4_mb_pa_assert_overlap(ac, new_start, new_end);
|
| 4509 |
|
| 4510 | *start = new_start;
|
| 4511 | *end = new_end;
|
| 4512 | }
|
| 4513 |
|
| 4514 | |
| 4515 | * Normalization means making request better in terms of
|
| 4516 | * size and alignment
|
| 4517 |
|
| 4518 | static noinline_for_stack void
|
| 4519 | ext4_mb_normalize_request(struct ext4_allocation_context *ac,
|
| 4520 | struct ext4_allocation_request *ar)
|
| 4521 | {
|
| 4522 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4523 | struct ext4_super_block *es = sbi->s_es;
|
| 4524 | int bsbits, max;
|
| 4525 | loff_t size, start_off, end;
|
| 4526 | loff_t orig_size __maybe_unused;
|
| 4527 | ext4_lblk_t start;
|
| 4528 |
|
| 4529 | |
| 4530 | do not need preallocation
|
| 4531 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
|
| 4532 | return;
|
| 4533 |
|
| 4534 |
|
| 4535 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
|
| 4536 | return;
|
| 4537 |
|
| 4538 | |
| 4539 | * required (it's a tail, for example)
|
| 4540 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
|
| 4541 | return;
|
| 4542 |
|
| 4543 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
|
| 4544 | ext4_mb_normalize_group_request(ac);
|
| 4545 | return ;
|
| 4546 | }
|
| 4547 |
|
| 4548 | bsbits = ac->ac_sb->s_blocksize_bits;
|
| 4549 |
|
| 4550 | |
| 4551 | * given current request is allocated
|
| 4552 | size = extent_logical_end(sbi, &ac->ac_o_ex);
|
| 4553 | size = size << bsbits;
|
| 4554 | if (size < i_size_read(ac->ac_inode))
|
| 4555 | size = i_size_read(ac->ac_inode);
|
| 4556 | orig_size = size;
|
| 4557 |
|
| 4558 |
|
| 4559 | max = 2 << bsbits;
|
| 4560 |
|
| 4561 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
|
| 4562 | (req <= (size) || max <= (chunk_size))
|
| 4563 |
|
| 4564 |
|
| 4565 | start_off = 0;
|
| 4566 | if (size <= SZ_1M) {
|
| 4567 | |
| 4568 | * For files up to 1MB, round up the preallocation size to
|
| 4569 | * the next power of two, with a minimum of 16KB.
|
| 4570 |
|
| 4571 | if (size <= (unsigned long)SZ_16K)
|
| 4572 | size = SZ_16K;
|
| 4573 | else
|
| 4574 | size = roundup_pow_of_two(size);
|
| 4575 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
|
| 4576 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
|
| 4577 | (21 - bsbits)) << 21;
|
| 4578 | size = 2 * 1024 * 1024;
|
| 4579 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
|
| 4580 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
|
| 4581 | (22 - bsbits)) << 22;
|
| 4582 | size = 4 * 1024 * 1024;
|
| 4583 | } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len),
|
| 4584 | (8<<20)>>bsbits, max, 8 * 1024)) {
|
| 4585 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
|
| 4586 | (23 - bsbits)) << 23;
|
| 4587 | size = 8 * 1024 * 1024;
|
| 4588 | } else {
|
| 4589 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
|
| 4590 | size = (loff_t) EXT4_C2B(sbi,
|
| 4591 | ac->ac_o_ex.fe_len) << bsbits;
|
| 4592 | }
|
| 4593 | size = size >> bsbits;
|
| 4594 | start = start_off >> bsbits;
|
| 4595 |
|
| 4596 | |
| 4597 | * For tiny groups (smaller than 8MB) the chosen allocation
|
| 4598 | * alignment may be larger than group size. Make sure the
|
| 4599 | * alignment does not move allocation to a different group which
|
| 4600 | * makes mballoc fail assertions later.
|
| 4601 |
|
| 4602 | start = max(start, rounddown(ac->ac_o_ex.fe_logical,
|
| 4603 | (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
|
| 4604 |
|
| 4605 |
|
| 4606 | if (start + size > EXT_MAX_BLOCKS)
|
| 4607 | size = EXT_MAX_BLOCKS - start;
|
| 4608 |
|
| 4609 |
|
| 4610 | if (ar->pleft && start <= ar->lleft) {
|
| 4611 | size -= ar->lleft + 1 - start;
|
| 4612 | start = ar->lleft + 1;
|
| 4613 | }
|
| 4614 | if (ar->pright && start + size - 1 >= ar->lright)
|
| 4615 | size -= start + size - ar->lright;
|
| 4616 |
|
| 4617 | |
| 4618 | * Trim allocation request for filesystems with artificially small
|
| 4619 | * groups.
|
| 4620 |
|
| 4621 | if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
|
| 4622 | size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
|
| 4623 |
|
| 4624 | end = start + size;
|
| 4625 |
|
| 4626 | ext4_mb_pa_adjust_overlap(ac, &start, &end);
|
| 4627 |
|
| 4628 | size = end - start;
|
| 4629 |
|
| 4630 | |
| 4631 | * In this function "start" and "size" are normalized for better
|
| 4632 | * alignment and length such that we could preallocate more blocks.
|
| 4633 | * This normalization is done such that original request of
|
| 4634 | * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
|
| 4635 | * "size" boundaries.
|
| 4636 | * (Note fe_len can be relaxed since FS block allocation API does not
|
| 4637 | * provide gurantee on number of contiguous blocks allocation since that
|
| 4638 | * depends upon free space left, etc).
|
| 4639 | * In case of inode pa, later we use the allocated blocks
|
| 4640 | * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated
|
| 4641 | * range of goal/best blocks [start, size] to put it at the
|
| 4642 | * ac_o_ex.fe_logical extent of this inode.
|
| 4643 | * (See ext4_mb_use_inode_pa() for more details)
|
| 4644 |
|
| 4645 | if (start + size <= ac->ac_o_ex.fe_logical ||
|
| 4646 | start > ac->ac_o_ex.fe_logical) {
|
| 4647 | ext4_msg(ac->ac_sb, KERN_ERR,
|
| 4648 | "start %lu, size %lu, fe_logical %lu",
|
| 4649 | (unsigned long) start, (unsigned long) size,
|
| 4650 | (unsigned long) ac->ac_o_ex.fe_logical);
|
| 4651 | BUG();
|
| 4652 | }
|
| 4653 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
|
| 4654 |
|
| 4655 |
|
| 4656 |
|
| 4657 | |
| 4658 | * placement or satisfy big request as is
|
| 4659 | ac->ac_g_ex.fe_logical = start;
|
| 4660 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
|
| 4661 | ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
|
| 4662 |
|
| 4663 |
|
| 4664 | if (ar->pright && (ar->lright == (start + size)) &&
|
| 4665 | ar->pright >= size &&
|
| 4666 | ar->pright - size >= le32_to_cpu(es->s_first_data_block)) {
|
| 4667 |
|
| 4668 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
|
| 4669 | &ac->ac_g_ex.fe_group,
|
| 4670 | &ac->ac_g_ex.fe_start);
|
| 4671 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
|
| 4672 | }
|
| 4673 | if (ar->pleft && (ar->lleft + 1 == start) &&
|
| 4674 | ar->pleft + 1 < ext4_blocks_count(es)) {
|
| 4675 |
|
| 4676 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
|
| 4677 | &ac->ac_g_ex.fe_group,
|
| 4678 | &ac->ac_g_ex.fe_start);
|
| 4679 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
|
| 4680 | }
|
| 4681 |
|
| 4682 | mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
|
| 4683 | orig_size, start);
|
| 4684 | }
|
| 4685 |
|
| 4686 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
|
| 4687 | {
|
| 4688 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4689 |
|
| 4690 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
|
| 4691 | atomic_inc(&sbi->s_bal_reqs);
|
| 4692 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
|
| 4693 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
|
| 4694 | atomic_inc(&sbi->s_bal_success);
|
| 4695 |
|
| 4696 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
|
| 4697 | for (int i=0; i<EXT4_MB_NUM_CRS; i++) {
|
| 4698 | atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]);
|
| 4699 | }
|
| 4700 |
|
| 4701 | atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
|
| 4702 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
|
| 4703 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
|
| 4704 | atomic_inc(&sbi->s_bal_goals);
|
| 4705 |
|
| 4706 | if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len)
|
| 4707 | atomic_inc(&sbi->s_bal_len_goals);
|
| 4708 |
|
| 4709 | if (ac->ac_found > sbi->s_mb_max_to_scan)
|
| 4710 | atomic_inc(&sbi->s_bal_breaks);
|
| 4711 | }
|
| 4712 |
|
| 4713 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
|
| 4714 | trace_ext4_mballoc_alloc(ac);
|
| 4715 | else
|
| 4716 | trace_ext4_mballoc_prealloc(ac);
|
| 4717 | }
|
| 4718 |
|
| 4719 | |
| 4720 | * Called on failure; free up any blocks from the inode PA for this
|
| 4721 | * context. We don't need this for MB_GROUP_PA because we only change
|
| 4722 | * pa_free in ext4_mb_release_context(), but on failure, we've already
|
| 4723 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
|
| 4724 |
|
| 4725 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
|
| 4726 | {
|
| 4727 | struct ext4_prealloc_space *pa = ac->ac_pa;
|
| 4728 | struct ext4_buddy e4b;
|
| 4729 | int err;
|
| 4730 |
|
| 4731 | if (pa == NULL) {
|
| 4732 | if (ac->ac_f_ex.fe_len == 0)
|
| 4733 | return;
|
| 4734 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
|
| 4735 | if (WARN_RATELIMIT(err,
|
| 4736 | "ext4: mb_load_buddy failed (%d)", err))
|
| 4737 | |
| 4738 | * This should never happen since we pin the
|
| 4739 | * folios in the ext4_allocation_context so
|
| 4740 | * ext4_mb_load_buddy() should never fail.
|
| 4741 |
|
| 4742 | return;
|
| 4743 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
|
| 4744 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
|
| 4745 | ac->ac_f_ex.fe_len);
|
| 4746 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
|
| 4747 | ext4_mb_unload_buddy(&e4b);
|
| 4748 | return;
|
| 4749 | }
|
| 4750 | if (pa->pa_type == MB_INODE_PA) {
|
| 4751 | spin_lock(&pa->pa_lock);
|
| 4752 | pa->pa_free += ac->ac_b_ex.fe_len;
|
| 4753 | spin_unlock(&pa->pa_lock);
|
| 4754 | }
|
| 4755 | }
|
| 4756 |
|
| 4757 | |
| 4758 | * use blocks preallocated to inode
|
| 4759 |
|
| 4760 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
|
| 4761 | struct ext4_prealloc_space *pa)
|
| 4762 | {
|
| 4763 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4764 | ext4_fsblk_t start;
|
| 4765 | ext4_fsblk_t end;
|
| 4766 | int len;
|
| 4767 |
|
| 4768 |
|
| 4769 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
|
| 4770 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
|
| 4771 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
|
| 4772 | len = EXT4_NUM_B2C(sbi, end - start);
|
| 4773 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
|
| 4774 | &ac->ac_b_ex.fe_start);
|
| 4775 | ac->ac_b_ex.fe_len = len;
|
| 4776 | ac->ac_status = AC_STATUS_FOUND;
|
| 4777 | ac->ac_pa = pa;
|
| 4778 |
|
| 4779 | BUG_ON(start < pa->pa_pstart);
|
| 4780 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
|
| 4781 | BUG_ON(pa->pa_free < len);
|
| 4782 | BUG_ON(ac->ac_b_ex.fe_len <= 0);
|
| 4783 | pa->pa_free -= len;
|
| 4784 |
|
| 4785 | mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
|
| 4786 | }
|
| 4787 |
|
| 4788 | |
| 4789 | * use blocks preallocated to locality group
|
| 4790 |
|
| 4791 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
|
| 4792 | struct ext4_prealloc_space *pa)
|
| 4793 | {
|
| 4794 | unsigned int len = ac->ac_o_ex.fe_len;
|
| 4795 |
|
| 4796 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
|
| 4797 | &ac->ac_b_ex.fe_group,
|
| 4798 | &ac->ac_b_ex.fe_start);
|
| 4799 | ac->ac_b_ex.fe_len = len;
|
| 4800 | ac->ac_status = AC_STATUS_FOUND;
|
| 4801 | ac->ac_pa = pa;
|
| 4802 |
|
| 4803 | |
| 4804 | * possible race when the group is being loaded concurrently
|
| 4805 | * instead we correct pa later, after blocks are marked
|
| 4806 | * in on-disk bitmap -- see ext4_mb_release_context()
|
| 4807 | * Other CPUs are prevented from allocating from this pa by lg_mutex
|
| 4808 |
|
| 4809 | mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
|
| 4810 | pa->pa_lstart, len, pa);
|
| 4811 | }
|
| 4812 |
|
| 4813 | |
| 4814 | * Return the prealloc space that have minimal distance
|
| 4815 | * from the goal block. @cpa is the prealloc
|
| 4816 | * space that is having currently known minimal distance
|
| 4817 | * from the goal block.
|
| 4818 |
|
| 4819 | static struct ext4_prealloc_space *
|
| 4820 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
|
| 4821 | struct ext4_prealloc_space *pa,
|
| 4822 | struct ext4_prealloc_space *cpa)
|
| 4823 | {
|
| 4824 | ext4_fsblk_t cur_distance, new_distance;
|
| 4825 |
|
| 4826 | if (cpa == NULL) {
|
| 4827 | atomic_inc(&pa->pa_count);
|
| 4828 | return pa;
|
| 4829 | }
|
| 4830 | cur_distance = abs(goal_block - cpa->pa_pstart);
|
| 4831 | new_distance = abs(goal_block - pa->pa_pstart);
|
| 4832 |
|
| 4833 | if (cur_distance <= new_distance)
|
| 4834 | return cpa;
|
| 4835 |
|
| 4836 |
|
| 4837 | atomic_dec(&cpa->pa_count);
|
| 4838 | atomic_inc(&pa->pa_count);
|
| 4839 | return pa;
|
| 4840 | }
|
| 4841 |
|
| 4842 | |
| 4843 | * check if found pa meets EXT4_MB_HINT_GOAL_ONLY
|
| 4844 |
|
| 4845 | static bool
|
| 4846 | ext4_mb_pa_goal_check(struct ext4_allocation_context *ac,
|
| 4847 | struct ext4_prealloc_space *pa)
|
| 4848 | {
|
| 4849 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4850 | ext4_fsblk_t start;
|
| 4851 |
|
| 4852 | if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)))
|
| 4853 | return true;
|
| 4854 |
|
| 4855 | |
| 4856 | * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted
|
| 4857 | * in ext4_mb_normalize_request and will keep same with ac_o_ex
|
| 4858 | * from ext4_mb_initialize_context. Choose ac_g_ex here to keep
|
| 4859 | * consistent with ext4_mb_find_by_goal.
|
| 4860 |
|
| 4861 | start = pa->pa_pstart +
|
| 4862 | (ac->ac_g_ex.fe_logical - pa->pa_lstart);
|
| 4863 | if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start)
|
| 4864 | return false;
|
| 4865 |
|
| 4866 | if (ac->ac_g_ex.fe_len > pa->pa_len -
|
| 4867 | EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart))
|
| 4868 | return false;
|
| 4869 |
|
| 4870 | return true;
|
| 4871 | }
|
| 4872 |
|
| 4873 | |
| 4874 | * search goal blocks in preallocated space
|
| 4875 |
|
| 4876 | static noinline_for_stack bool
|
| 4877 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
|
| 4878 | {
|
| 4879 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 4880 | int order, i;
|
| 4881 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
|
| 4882 | struct ext4_locality_group *lg;
|
| 4883 | struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL;
|
| 4884 | struct rb_node *iter;
|
| 4885 | ext4_fsblk_t goal_block;
|
| 4886 |
|
| 4887 |
|
| 4888 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
|
| 4889 | return false;
|
| 4890 |
|
| 4891 | |
| 4892 | * first, try per-file preallocation by searching the inode pa rbtree.
|
| 4893 | *
|
| 4894 | * Here, we can't do a direct traversal of the tree because
|
| 4895 | * ext4_mb_discard_group_preallocation() can paralelly mark the pa
|
| 4896 | * deleted and that can cause direct traversal to skip some entries.
|
| 4897 |
|
| 4898 | read_lock(&ei->i_prealloc_lock);
|
| 4899 |
|
| 4900 | if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) {
|
| 4901 | goto try_group_pa;
|
| 4902 | }
|
| 4903 |
|
| 4904 | |
| 4905 | * Step 1: Find a pa with logical start immediately adjacent to the
|
| 4906 | * original logical start. This could be on the left or right.
|
| 4907 | *
|
| 4908 | * (tmp_pa->pa_lstart never changes so we can skip locking for it).
|
| 4909 |
|
| 4910 | for (iter = ei->i_prealloc_node.rb_node; iter;
|
| 4911 | iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
|
| 4912 | tmp_pa->pa_lstart, iter)) {
|
| 4913 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 4914 | pa_node.inode_node);
|
| 4915 | }
|
| 4916 |
|
| 4917 | |
| 4918 | * Step 2: The adjacent pa might be to the right of logical start, find
|
| 4919 | * the left adjacent pa. After this step we'd have a valid tmp_pa whose
|
| 4920 | * logical start is towards the left of original request's logical start
|
| 4921 |
|
| 4922 | if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) {
|
| 4923 | struct rb_node *tmp;
|
| 4924 | tmp = rb_prev(&tmp_pa->pa_node.inode_node);
|
| 4925 |
|
| 4926 | if (tmp) {
|
| 4927 | tmp_pa = rb_entry(tmp, struct ext4_prealloc_space,
|
| 4928 | pa_node.inode_node);
|
| 4929 | } else {
|
| 4930 | |
| 4931 | * If there is no adjacent pa to the left then finding
|
| 4932 | * an overlapping pa is not possible hence stop searching
|
| 4933 | * inode pa tree
|
| 4934 |
|
| 4935 | goto try_group_pa;
|
| 4936 | }
|
| 4937 | }
|
| 4938 |
|
| 4939 | BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
|
| 4940 |
|
| 4941 | |
| 4942 | * Step 3: If the left adjacent pa is deleted, keep moving left to find
|
| 4943 | * the first non deleted adjacent pa. After this step we should have a
|
| 4944 | * valid tmp_pa which is guaranteed to be non deleted.
|
| 4945 |
|
| 4946 | for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) {
|
| 4947 | if (!iter) {
|
| 4948 | |
| 4949 | * no non deleted left adjacent pa, so stop searching
|
| 4950 | * inode pa tree
|
| 4951 |
|
| 4952 | goto try_group_pa;
|
| 4953 | }
|
| 4954 | tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 4955 | pa_node.inode_node);
|
| 4956 | spin_lock(&tmp_pa->pa_lock);
|
| 4957 | if (tmp_pa->pa_deleted == 0) {
|
| 4958 | |
| 4959 | * We will keep holding the pa_lock from
|
| 4960 | * this point on because we don't want group discard
|
| 4961 | * to delete this pa underneath us. Since group
|
| 4962 | * discard is anyways an ENOSPC operation it
|
| 4963 | * should be okay for it to wait a few more cycles.
|
| 4964 |
|
| 4965 | break;
|
| 4966 | } else {
|
| 4967 | spin_unlock(&tmp_pa->pa_lock);
|
| 4968 | }
|
| 4969 | }
|
| 4970 |
|
| 4971 | BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
|
| 4972 | BUG_ON(tmp_pa->pa_deleted == 1);
|
| 4973 |
|
| 4974 | |
| 4975 | * Step 4: We now have the non deleted left adjacent pa. Only this
|
| 4976 | * pa can possibly satisfy the request hence check if it overlaps
|
| 4977 | * original logical start and stop searching if it doesn't.
|
| 4978 |
|
| 4979 | if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) {
|
| 4980 | spin_unlock(&tmp_pa->pa_lock);
|
| 4981 | goto try_group_pa;
|
| 4982 | }
|
| 4983 |
|
| 4984 |
|
| 4985 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
|
| 4986 | (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) >
|
| 4987 | EXT4_MAX_BLOCK_FILE_PHYS)) {
|
| 4988 | |
| 4989 | * Since PAs don't overlap, we won't find any other PA to
|
| 4990 | * satisfy this.
|
| 4991 |
|
| 4992 | spin_unlock(&tmp_pa->pa_lock);
|
| 4993 | goto try_group_pa;
|
| 4994 | }
|
| 4995 |
|
| 4996 | if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) {
|
| 4997 | atomic_inc(&tmp_pa->pa_count);
|
| 4998 | ext4_mb_use_inode_pa(ac, tmp_pa);
|
| 4999 | spin_unlock(&tmp_pa->pa_lock);
|
| 5000 | read_unlock(&ei->i_prealloc_lock);
|
| 5001 | return true;
|
| 5002 | } else {
|
| 5003 | |
| 5004 | * We found a valid overlapping pa but couldn't use it because
|
| 5005 | * it had no free blocks. This should ideally never happen
|
| 5006 | * because:
|
| 5007 | *
|
| 5008 | * 1. When a new inode pa is added to rbtree it must have
|
| 5009 | * pa_free > 0 since otherwise we won't actually need
|
| 5010 | * preallocation.
|
| 5011 | *
|
| 5012 | * 2. An inode pa that is in the rbtree can only have it's
|
| 5013 | * pa_free become zero when another thread calls:
|
| 5014 | * ext4_mb_new_blocks
|
| 5015 | * ext4_mb_use_preallocated
|
| 5016 | * ext4_mb_use_inode_pa
|
| 5017 | *
|
| 5018 | * 3. Further, after the above calls make pa_free == 0, we will
|
| 5019 | * immediately remove it from the rbtree in:
|
| 5020 | * ext4_mb_new_blocks
|
| 5021 | * ext4_mb_release_context
|
| 5022 | * ext4_mb_put_pa
|
| 5023 | *
|
| 5024 | * 4. Since the pa_free becoming 0 and pa_free getting removed
|
| 5025 | * from tree both happen in ext4_mb_new_blocks, which is always
|
| 5026 | * called with i_data_sem held for data allocations, we can be
|
| 5027 | * sure that another process will never see a pa in rbtree with
|
| 5028 | * pa_free == 0.
|
| 5029 |
|
| 5030 | WARN_ON_ONCE(tmp_pa->pa_free == 0);
|
| 5031 | }
|
| 5032 | spin_unlock(&tmp_pa->pa_lock);
|
| 5033 | try_group_pa:
|
| 5034 | read_unlock(&ei->i_prealloc_lock);
|
| 5035 |
|
| 5036 |
|
| 5037 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
|
| 5038 | return false;
|
| 5039 |
|
| 5040 |
|
| 5041 | lg = ac->ac_lg;
|
| 5042 | if (lg == NULL)
|
| 5043 | return false;
|
| 5044 | order = fls(ac->ac_o_ex.fe_len) - 1;
|
| 5045 | if (order > PREALLOC_TB_SIZE - 1)
|
| 5046 |
|
| 5047 | order = PREALLOC_TB_SIZE - 1;
|
| 5048 |
|
| 5049 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
|
| 5050 | |
| 5051 | * search for the prealloc space that is having
|
| 5052 | * minimal distance from the goal block.
|
| 5053 |
|
| 5054 | for (i = order; i < PREALLOC_TB_SIZE; i++) {
|
| 5055 | rcu_read_lock();
|
| 5056 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
|
| 5057 | pa_node.lg_list) {
|
| 5058 | spin_lock(&tmp_pa->pa_lock);
|
| 5059 | if (tmp_pa->pa_deleted == 0 &&
|
| 5060 | tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {
|
| 5061 |
|
| 5062 | cpa = ext4_mb_check_group_pa(goal_block,
|
| 5063 | tmp_pa, cpa);
|
| 5064 | }
|
| 5065 | spin_unlock(&tmp_pa->pa_lock);
|
| 5066 | }
|
| 5067 | rcu_read_unlock();
|
| 5068 | }
|
| 5069 | if (cpa) {
|
| 5070 | ext4_mb_use_group_pa(ac, cpa);
|
| 5071 | return true;
|
| 5072 | }
|
| 5073 | return false;
|
| 5074 | }
|
| 5075 |
|
| 5076 | |
| 5077 | * the function goes through all preallocation in this group and marks them
|
| 5078 | * used in in-core bitmap. buddy must be generated from this bitmap
|
| 5079 | * Need to be called with ext4 group lock held
|
| 5080 |
|
| 5081 | static noinline_for_stack
|
| 5082 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
|
| 5083 | ext4_group_t group)
|
| 5084 | {
|
| 5085 | struct ext4_group_info *grp = ext4_get_group_info(sb, group);
|
| 5086 | struct ext4_prealloc_space *pa;
|
| 5087 | struct list_head *cur;
|
| 5088 | ext4_group_t groupnr;
|
| 5089 | ext4_grpblk_t start;
|
| 5090 | int preallocated = 0;
|
| 5091 | int len;
|
| 5092 |
|
| 5093 | if (!grp)
|
| 5094 | return;
|
| 5095 |
|
| 5096 | |
| 5097 | * so the only competing code is preallocation use.
|
| 5098 | * we don't need any locking here
|
| 5099 | * notice we do NOT ignore preallocations with pa_deleted
|
| 5100 | * otherwise we could leave used blocks available for
|
| 5101 | * allocation in buddy when concurrent ext4_mb_put_pa()
|
| 5102 | * is dropping preallocation
|
| 5103 |
|
| 5104 | list_for_each(cur, &grp->bb_prealloc_list) {
|
| 5105 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
|
| 5106 | spin_lock(&pa->pa_lock);
|
| 5107 | ext4_get_group_no_and_offset(sb, pa->pa_pstart,
|
| 5108 | &groupnr, &start);
|
| 5109 | len = pa->pa_len;
|
| 5110 | spin_unlock(&pa->pa_lock);
|
| 5111 | if (unlikely(len == 0))
|
| 5112 | continue;
|
| 5113 | BUG_ON(groupnr != group);
|
| 5114 | mb_set_bits(bitmap, start, len);
|
| 5115 | preallocated += len;
|
| 5116 | }
|
| 5117 | mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
|
| 5118 | }
|
| 5119 |
|
| 5120 | static void ext4_mb_mark_pa_deleted(struct super_block *sb,
|
| 5121 | struct ext4_prealloc_space *pa)
|
| 5122 | {
|
| 5123 | struct ext4_inode_info *ei;
|
| 5124 |
|
| 5125 | if (pa->pa_deleted) {
|
| 5126 | ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
|
| 5127 | pa->pa_type, pa->pa_pstart, pa->pa_lstart,
|
| 5128 | pa->pa_len);
|
| 5129 | return;
|
| 5130 | }
|
| 5131 |
|
| 5132 | pa->pa_deleted = 1;
|
| 5133 |
|
| 5134 | if (pa->pa_type == MB_INODE_PA) {
|
| 5135 | ei = EXT4_I(pa->pa_inode);
|
| 5136 | atomic_dec(&ei->i_prealloc_active);
|
| 5137 | }
|
| 5138 | }
|
| 5139 |
|
| 5140 | static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa)
|
| 5141 | {
|
| 5142 | BUG_ON(!pa);
|
| 5143 | BUG_ON(atomic_read(&pa->pa_count));
|
| 5144 | BUG_ON(pa->pa_deleted == 0);
|
| 5145 | kmem_cache_free(ext4_pspace_cachep, pa);
|
| 5146 | }
|
| 5147 |
|
| 5148 | static void ext4_mb_pa_callback(struct rcu_head *head)
|
| 5149 | {
|
| 5150 | struct ext4_prealloc_space *pa;
|
| 5151 |
|
| 5152 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
|
| 5153 | ext4_mb_pa_free(pa);
|
| 5154 | }
|
| 5155 |
|
| 5156 | |
| 5157 | * drops a reference to preallocated space descriptor
|
| 5158 | * if this was the last reference and the space is consumed
|
| 5159 |
|
| 5160 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
|
| 5161 | struct super_block *sb, struct ext4_prealloc_space *pa)
|
| 5162 | {
|
| 5163 | ext4_group_t grp;
|
| 5164 | ext4_fsblk_t grp_blk;
|
| 5165 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
|
| 5166 |
|
| 5167 |
|
| 5168 | spin_lock(&pa->pa_lock);
|
| 5169 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
|
| 5170 | spin_unlock(&pa->pa_lock);
|
| 5171 | return;
|
| 5172 | }
|
| 5173 |
|
| 5174 | if (pa->pa_deleted == 1) {
|
| 5175 | spin_unlock(&pa->pa_lock);
|
| 5176 | return;
|
| 5177 | }
|
| 5178 |
|
| 5179 | ext4_mb_mark_pa_deleted(sb, pa);
|
| 5180 | spin_unlock(&pa->pa_lock);
|
| 5181 |
|
| 5182 | grp_blk = pa->pa_pstart;
|
| 5183 | |
| 5184 | * If doing group-based preallocation, pa_pstart may be in the
|
| 5185 | * next group when pa is used up
|
| 5186 |
|
| 5187 | if (pa->pa_type == MB_GROUP_PA)
|
| 5188 | grp_blk--;
|
| 5189 |
|
| 5190 | grp = ext4_get_group_number(sb, grp_blk);
|
| 5191 |
|
| 5192 | |
| 5193 | * possible race:
|
| 5194 | *
|
| 5195 | * P1 (buddy init) P2 (regular allocation)
|
| 5196 | * find block B in PA
|
| 5197 | * copy on-disk bitmap to buddy
|
| 5198 | * mark B in on-disk bitmap
|
| 5199 | * drop PA from group
|
| 5200 | * mark all PAs in buddy
|
| 5201 | *
|
| 5202 | * thus, P1 initializes buddy with B available. to prevent this
|
| 5203 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
|
| 5204 | * against that pair
|
| 5205 |
|
| 5206 | ext4_lock_group(sb, grp);
|
| 5207 | list_del(&pa->pa_group_list);
|
| 5208 | ext4_unlock_group(sb, grp);
|
| 5209 |
|
| 5210 | if (pa->pa_type == MB_INODE_PA) {
|
| 5211 | write_lock(pa->pa_node_lock.inode_lock);
|
| 5212 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
|
| 5213 | write_unlock(pa->pa_node_lock.inode_lock);
|
| 5214 | ext4_mb_pa_free(pa);
|
| 5215 | } else {
|
| 5216 | spin_lock(pa->pa_node_lock.lg_lock);
|
| 5217 | list_del_rcu(&pa->pa_node.lg_list);
|
| 5218 | spin_unlock(pa->pa_node_lock.lg_lock);
|
| 5219 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
|
| 5220 | }
|
| 5221 | }
|
| 5222 |
|
| 5223 | static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new)
|
| 5224 | {
|
| 5225 | struct rb_node **iter = &root->rb_node, *parent = NULL;
|
| 5226 | struct ext4_prealloc_space *iter_pa, *new_pa;
|
| 5227 | ext4_lblk_t iter_start, new_start;
|
| 5228 |
|
| 5229 | while (*iter) {
|
| 5230 | iter_pa = rb_entry(*iter, struct ext4_prealloc_space,
|
| 5231 | pa_node.inode_node);
|
| 5232 | new_pa = rb_entry(new, struct ext4_prealloc_space,
|
| 5233 | pa_node.inode_node);
|
| 5234 | iter_start = iter_pa->pa_lstart;
|
| 5235 | new_start = new_pa->pa_lstart;
|
| 5236 |
|
| 5237 | parent = *iter;
|
| 5238 | if (new_start < iter_start)
|
| 5239 | iter = &((*iter)->rb_left);
|
| 5240 | else
|
| 5241 | iter = &((*iter)->rb_right);
|
| 5242 | }
|
| 5243 |
|
| 5244 | rb_link_node(new, parent, iter);
|
| 5245 | rb_insert_color(new, root);
|
| 5246 | }
|
| 5247 |
|
| 5248 | |
| 5249 | * creates new preallocated space for given inode
|
| 5250 |
|
| 5251 | static noinline_for_stack void
|
| 5252 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
|
| 5253 | {
|
| 5254 | struct super_block *sb = ac->ac_sb;
|
| 5255 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 5256 | struct ext4_prealloc_space *pa;
|
| 5257 | struct ext4_group_info *grp;
|
| 5258 | struct ext4_inode_info *ei;
|
| 5259 |
|
| 5260 |
|
| 5261 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
|
| 5262 | BUG_ON(ac->ac_status != AC_STATUS_FOUND);
|
| 5263 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
|
| 5264 | BUG_ON(ac->ac_pa == NULL);
|
| 5265 |
|
| 5266 | pa = ac->ac_pa;
|
| 5267 |
|
| 5268 | if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) {
|
| 5269 | struct ext4_free_extent ex = {
|
| 5270 | .fe_logical = ac->ac_g_ex.fe_logical,
|
| 5271 | .fe_len = ac->ac_orig_goal_len,
|
| 5272 | };
|
| 5273 | loff_t orig_goal_end = extent_logical_end(sbi, &ex);
|
| 5274 | loff_t o_ex_end = extent_logical_end(sbi, &ac->ac_o_ex);
|
| 5275 |
|
| 5276 | |
| 5277 | * We can't allocate as much as normalizer wants, so we try
|
| 5278 | * to get proper lstart to cover the original request, except
|
| 5279 | * when the goal doesn't cover the original request as below:
|
| 5280 | *
|
| 5281 | * orig_ex:2045/2055(10), isize:8417280 -> normalized:0/2048
|
| 5282 | * best_ex:0/200(200) -> adjusted: 1848/2048(200)
|
| 5283 |
|
| 5284 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
|
| 5285 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
|
| 5286 |
|
| 5287 | |
| 5288 | * Use the below logic for adjusting best extent as it keeps
|
| 5289 | * fragmentation in check while ensuring logical range of best
|
| 5290 | * extent doesn't overflow out of goal extent:
|
| 5291 | *
|
| 5292 | * 1. Check if best ex can be kept at end of goal (before
|
| 5293 | * cr_best_avail trimmed it) and still cover original start
|
| 5294 | * 2. Else, check if best ex can be kept at start of goal and
|
| 5295 | * still cover original end
|
| 5296 | * 3. Else, keep the best ex at start of original request.
|
| 5297 |
|
| 5298 | ex.fe_len = ac->ac_b_ex.fe_len;
|
| 5299 |
|
| 5300 | ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len);
|
| 5301 | if (ac->ac_o_ex.fe_logical >= ex.fe_logical)
|
| 5302 | goto adjust_bex;
|
| 5303 |
|
| 5304 | ex.fe_logical = ac->ac_g_ex.fe_logical;
|
| 5305 | if (o_ex_end <= extent_logical_end(sbi, &ex))
|
| 5306 | goto adjust_bex;
|
| 5307 |
|
| 5308 | ex.fe_logical = ac->ac_o_ex.fe_logical;
|
| 5309 | adjust_bex:
|
| 5310 | ac->ac_b_ex.fe_logical = ex.fe_logical;
|
| 5311 |
|
| 5312 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
|
| 5313 | BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end);
|
| 5314 | }
|
| 5315 |
|
| 5316 | pa->pa_lstart = ac->ac_b_ex.fe_logical;
|
| 5317 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
|
| 5318 | pa->pa_len = ac->ac_b_ex.fe_len;
|
| 5319 | pa->pa_free = pa->pa_len;
|
| 5320 | spin_lock_init(&pa->pa_lock);
|
| 5321 | INIT_LIST_HEAD(&pa->pa_group_list);
|
| 5322 | pa->pa_deleted = 0;
|
| 5323 | pa->pa_type = MB_INODE_PA;
|
| 5324 |
|
| 5325 | mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
|
| 5326 | pa->pa_len, pa->pa_lstart);
|
| 5327 | trace_ext4_mb_new_inode_pa(ac, pa);
|
| 5328 |
|
| 5329 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
|
| 5330 | ext4_mb_use_inode_pa(ac, pa);
|
| 5331 |
|
| 5332 | ei = EXT4_I(ac->ac_inode);
|
| 5333 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
|
| 5334 | if (!grp)
|
| 5335 | return;
|
| 5336 |
|
| 5337 | pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock;
|
| 5338 | pa->pa_inode = ac->ac_inode;
|
| 5339 |
|
| 5340 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
|
| 5341 |
|
| 5342 | write_lock(pa->pa_node_lock.inode_lock);
|
| 5343 | ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node);
|
| 5344 | write_unlock(pa->pa_node_lock.inode_lock);
|
| 5345 | atomic_inc(&ei->i_prealloc_active);
|
| 5346 | }
|
| 5347 |
|
| 5348 | |
| 5349 | * creates new preallocated space for locality group inodes belongs to
|
| 5350 |
|
| 5351 | static noinline_for_stack void
|
| 5352 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
|
| 5353 | {
|
| 5354 | struct super_block *sb = ac->ac_sb;
|
| 5355 | struct ext4_locality_group *lg;
|
| 5356 | struct ext4_prealloc_space *pa;
|
| 5357 | struct ext4_group_info *grp;
|
| 5358 |
|
| 5359 |
|
| 5360 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
|
| 5361 | BUG_ON(ac->ac_status != AC_STATUS_FOUND);
|
| 5362 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
|
| 5363 | BUG_ON(ac->ac_pa == NULL);
|
| 5364 |
|
| 5365 | pa = ac->ac_pa;
|
| 5366 |
|
| 5367 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
|
| 5368 | pa->pa_lstart = pa->pa_pstart;
|
| 5369 | pa->pa_len = ac->ac_b_ex.fe_len;
|
| 5370 | pa->pa_free = pa->pa_len;
|
| 5371 | spin_lock_init(&pa->pa_lock);
|
| 5372 | INIT_LIST_HEAD(&pa->pa_node.lg_list);
|
| 5373 | INIT_LIST_HEAD(&pa->pa_group_list);
|
| 5374 | pa->pa_deleted = 0;
|
| 5375 | pa->pa_type = MB_GROUP_PA;
|
| 5376 |
|
| 5377 | mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
|
| 5378 | pa->pa_len, pa->pa_lstart);
|
| 5379 | trace_ext4_mb_new_group_pa(ac, pa);
|
| 5380 |
|
| 5381 | ext4_mb_use_group_pa(ac, pa);
|
| 5382 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
|
| 5383 |
|
| 5384 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
|
| 5385 | if (!grp)
|
| 5386 | return;
|
| 5387 | lg = ac->ac_lg;
|
| 5388 | BUG_ON(lg == NULL);
|
| 5389 |
|
| 5390 | pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock;
|
| 5391 | pa->pa_inode = NULL;
|
| 5392 |
|
| 5393 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
|
| 5394 |
|
| 5395 | |
| 5396 | * We will later add the new pa to the right bucket
|
| 5397 | * after updating the pa_free in ext4_mb_release_context
|
| 5398 |
|
| 5399 | }
|
| 5400 |
|
| 5401 | static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
|
| 5402 | {
|
| 5403 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
|
| 5404 | ext4_mb_new_group_pa(ac);
|
| 5405 | else
|
| 5406 | ext4_mb_new_inode_pa(ac);
|
| 5407 | }
|
| 5408 |
|
| 5409 | |
| 5410 | * finds all unused blocks in on-disk bitmap, frees them in
|
| 5411 | * in-core bitmap and buddy.
|
| 5412 | * @pa must be unlinked from inode and group lists, so that
|
| 5413 | * nobody else can find/use it.
|
| 5414 | * the caller MUST hold group/inode locks.
|
| 5415 | * TODO: optimize the case when there are no in-core structures yet
|
| 5416 |
|
| 5417 | static noinline_for_stack void
|
| 5418 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
|
| 5419 | struct ext4_prealloc_space *pa)
|
| 5420 | {
|
| 5421 | struct super_block *sb = e4b->bd_sb;
|
| 5422 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 5423 | unsigned int end;
|
| 5424 | unsigned int next;
|
| 5425 | ext4_group_t group;
|
| 5426 | ext4_grpblk_t bit;
|
| 5427 | unsigned long long grp_blk_start;
|
| 5428 | int free = 0;
|
| 5429 |
|
| 5430 | BUG_ON(pa->pa_deleted == 0);
|
| 5431 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
|
| 5432 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
|
| 5433 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
|
| 5434 | end = bit + pa->pa_len;
|
| 5435 |
|
| 5436 | while (bit < end) {
|
| 5437 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
|
| 5438 | if (bit >= end)
|
| 5439 | break;
|
| 5440 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
|
| 5441 | mb_debug(sb, "free preallocated %u/%u in group %u\n",
|
| 5442 | (unsigned) ext4_group_first_block_no(sb, group) + bit,
|
| 5443 | (unsigned) next - bit, (unsigned) group);
|
| 5444 | free += next - bit;
|
| 5445 |
|
| 5446 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
|
| 5447 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
|
| 5448 | EXT4_C2B(sbi, bit)),
|
| 5449 | next - bit);
|
| 5450 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
|
| 5451 | bit = next + 1;
|
| 5452 | }
|
| 5453 | if (free != pa->pa_free) {
|
| 5454 | ext4_msg(e4b->bd_sb, KERN_CRIT,
|
| 5455 | "pa %p: logic %lu, phys. %lu, len %d",
|
| 5456 | pa, (unsigned long) pa->pa_lstart,
|
| 5457 | (unsigned long) pa->pa_pstart,
|
| 5458 | pa->pa_len);
|
| 5459 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
|
| 5460 | free, pa->pa_free);
|
| 5461 | |
| 5462 | * pa is already deleted so we use the value obtained
|
| 5463 | * from the bitmap and continue.
|
| 5464 |
|
| 5465 | }
|
| 5466 | atomic_add(free, &sbi->s_mb_discarded);
|
| 5467 | }
|
| 5468 |
|
| 5469 | static noinline_for_stack void
|
| 5470 | ext4_mb_release_group_pa(struct ext4_buddy *e4b,
|
| 5471 | struct ext4_prealloc_space *pa)
|
| 5472 | {
|
| 5473 | struct super_block *sb = e4b->bd_sb;
|
| 5474 | ext4_group_t group;
|
| 5475 | ext4_grpblk_t bit;
|
| 5476 |
|
| 5477 | trace_ext4_mb_release_group_pa(sb, pa);
|
| 5478 | BUG_ON(pa->pa_deleted == 0);
|
| 5479 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
|
| 5480 | if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) {
|
| 5481 | ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu",
|
| 5482 | e4b->bd_group, group, pa->pa_pstart);
|
| 5483 | return;
|
| 5484 | }
|
| 5485 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
|
| 5486 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
|
| 5487 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
|
| 5488 | }
|
| 5489 |
|
| 5490 | |
| 5491 | * releases all preallocations in given group
|
| 5492 | *
|
| 5493 | * first, we need to decide discard policy:
|
| 5494 | * - when do we discard
|
| 5495 | * 1) ENOSPC
|
| 5496 | * - how many do we discard
|
| 5497 | * 1) how many requested
|
| 5498 |
|
| 5499 | static noinline_for_stack int
|
| 5500 | ext4_mb_discard_group_preallocations(struct super_block *sb,
|
| 5501 | ext4_group_t group, int *busy)
|
| 5502 | {
|
| 5503 | struct ext4_group_info *grp = ext4_get_group_info(sb, group);
|
| 5504 | struct buffer_head *bitmap_bh = NULL;
|
| 5505 | struct ext4_prealloc_space *pa, *tmp;
|
| 5506 | LIST_HEAD(list);
|
| 5507 | struct ext4_buddy e4b;
|
| 5508 | struct ext4_inode_info *ei;
|
| 5509 | int err;
|
| 5510 | int free = 0;
|
| 5511 |
|
| 5512 | if (!grp)
|
| 5513 | return 0;
|
| 5514 | mb_debug(sb, "discard preallocation for group %u\n", group);
|
| 5515 | if (list_empty(&grp->bb_prealloc_list))
|
| 5516 | goto out_dbg;
|
| 5517 |
|
| 5518 | bitmap_bh = ext4_read_block_bitmap(sb, group);
|
| 5519 | if (IS_ERR(bitmap_bh)) {
|
| 5520 | err = PTR_ERR(bitmap_bh);
|
| 5521 | ext4_error_err(sb, -err,
|
| 5522 | "Error %d reading block bitmap for %u",
|
| 5523 | err, group);
|
| 5524 | goto out_dbg;
|
| 5525 | }
|
| 5526 |
|
| 5527 | err = ext4_mb_load_buddy(sb, group, &e4b);
|
| 5528 | if (err) {
|
| 5529 | ext4_warning(sb, "Error %d loading buddy information for %u",
|
| 5530 | err, group);
|
| 5531 | put_bh(bitmap_bh);
|
| 5532 | goto out_dbg;
|
| 5533 | }
|
| 5534 |
|
| 5535 | ext4_lock_group(sb, group);
|
| 5536 | list_for_each_entry_safe(pa, tmp,
|
| 5537 | &grp->bb_prealloc_list, pa_group_list) {
|
| 5538 | spin_lock(&pa->pa_lock);
|
| 5539 | if (atomic_read(&pa->pa_count)) {
|
| 5540 | spin_unlock(&pa->pa_lock);
|
| 5541 | *busy = 1;
|
| 5542 | continue;
|
| 5543 | }
|
| 5544 | if (pa->pa_deleted) {
|
| 5545 | spin_unlock(&pa->pa_lock);
|
| 5546 | continue;
|
| 5547 | }
|
| 5548 |
|
| 5549 |
|
| 5550 | ext4_mb_mark_pa_deleted(sb, pa);
|
| 5551 |
|
| 5552 | if (!free)
|
| 5553 | this_cpu_inc(discard_pa_seq);
|
| 5554 |
|
| 5555 |
|
| 5556 | free += pa->pa_free;
|
| 5557 |
|
| 5558 | spin_unlock(&pa->pa_lock);
|
| 5559 |
|
| 5560 | list_del(&pa->pa_group_list);
|
| 5561 | list_add(&pa->u.pa_tmp_list, &list);
|
| 5562 | }
|
| 5563 |
|
| 5564 |
|
| 5565 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
|
| 5566 |
|
| 5567 |
|
| 5568 | if (pa->pa_type == MB_GROUP_PA) {
|
| 5569 | spin_lock(pa->pa_node_lock.lg_lock);
|
| 5570 | list_del_rcu(&pa->pa_node.lg_list);
|
| 5571 | spin_unlock(pa->pa_node_lock.lg_lock);
|
| 5572 | } else {
|
| 5573 | write_lock(pa->pa_node_lock.inode_lock);
|
| 5574 | ei = EXT4_I(pa->pa_inode);
|
| 5575 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
|
| 5576 | write_unlock(pa->pa_node_lock.inode_lock);
|
| 5577 | }
|
| 5578 |
|
| 5579 | list_del(&pa->u.pa_tmp_list);
|
| 5580 |
|
| 5581 | if (pa->pa_type == MB_GROUP_PA) {
|
| 5582 | ext4_mb_release_group_pa(&e4b, pa);
|
| 5583 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
|
| 5584 | } else {
|
| 5585 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
|
| 5586 | ext4_mb_pa_free(pa);
|
| 5587 | }
|
| 5588 | }
|
| 5589 |
|
| 5590 | ext4_unlock_group(sb, group);
|
| 5591 | ext4_mb_unload_buddy(&e4b);
|
| 5592 | put_bh(bitmap_bh);
|
| 5593 | out_dbg:
|
| 5594 | mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
|
| 5595 | free, group, grp->bb_free);
|
| 5596 | return free;
|
| 5597 | }
|
| 5598 |
|
| 5599 | |
| 5600 | * releases all non-used preallocated blocks for given inode
|
| 5601 | *
|
| 5602 | * It's important to discard preallocations under i_data_sem
|
| 5603 | * We don't want another block to be served from the prealloc
|
| 5604 | * space when we are discarding the inode prealloc space.
|
| 5605 | *
|
| 5606 | * FIXME!! Make sure it is valid at all the call sites
|
| 5607 |
|
| 5608 | void ext4_discard_preallocations(struct inode *inode)
|
| 5609 | {
|
| 5610 | struct ext4_inode_info *ei = EXT4_I(inode);
|
| 5611 | struct super_block *sb = inode->i_sb;
|
| 5612 | struct buffer_head *bitmap_bh = NULL;
|
| 5613 | struct ext4_prealloc_space *pa, *tmp;
|
| 5614 | ext4_group_t group = 0;
|
| 5615 | LIST_HEAD(list);
|
| 5616 | struct ext4_buddy e4b;
|
| 5617 | struct rb_node *iter;
|
| 5618 | int err;
|
| 5619 |
|
| 5620 | if (!S_ISREG(inode->i_mode))
|
| 5621 | return;
|
| 5622 |
|
| 5623 | if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
|
| 5624 | return;
|
| 5625 |
|
| 5626 | mb_debug(sb, "discard preallocation for inode %llu\n",
|
| 5627 | inode->i_ino);
|
| 5628 | trace_ext4_discard_preallocations(inode,
|
| 5629 | atomic_read(&ei->i_prealloc_active));
|
| 5630 |
|
| 5631 | repeat:
|
| 5632 |
|
| 5633 | write_lock(&ei->i_prealloc_lock);
|
| 5634 | for (iter = rb_first(&ei->i_prealloc_node); iter;
|
| 5635 | iter = rb_next(iter)) {
|
| 5636 | pa = rb_entry(iter, struct ext4_prealloc_space,
|
| 5637 | pa_node.inode_node);
|
| 5638 | BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock);
|
| 5639 |
|
| 5640 | spin_lock(&pa->pa_lock);
|
| 5641 | if (atomic_read(&pa->pa_count)) {
|
| 5642 | |
| 5643 | * use preallocation while we're discarding it
|
| 5644 | spin_unlock(&pa->pa_lock);
|
| 5645 | write_unlock(&ei->i_prealloc_lock);
|
| 5646 | ext4_msg(sb, KERN_ERR,
|
| 5647 | "uh-oh! used pa while discarding");
|
| 5648 | WARN_ON(1);
|
| 5649 | schedule_timeout_uninterruptible(HZ);
|
| 5650 | goto repeat;
|
| 5651 |
|
| 5652 | }
|
| 5653 | if (pa->pa_deleted == 0) {
|
| 5654 | ext4_mb_mark_pa_deleted(sb, pa);
|
| 5655 | spin_unlock(&pa->pa_lock);
|
| 5656 | rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
|
| 5657 | list_add(&pa->u.pa_tmp_list, &list);
|
| 5658 | continue;
|
| 5659 | }
|
| 5660 |
|
| 5661 |
|
| 5662 | spin_unlock(&pa->pa_lock);
|
| 5663 | write_unlock(&ei->i_prealloc_lock);
|
| 5664 |
|
| 5665 | |
| 5666 | * doesn't mean pa is already unlinked from
|
| 5667 | * the list. as we might be called from
|
| 5668 | * ->clear_inode() the inode will get freed
|
| 5669 | * and concurrent thread which is unlinking
|
| 5670 | * pa from inode's list may access already
|
| 5671 | * freed memory, bad-bad-bad
|
| 5672 |
|
| 5673 | |
| 5674 | * add a flag to force wait only in case
|
| 5675 | * of ->clear_inode(), but not in case of
|
| 5676 | * regular truncate
|
| 5677 | schedule_timeout_uninterruptible(HZ);
|
| 5678 | goto repeat;
|
| 5679 | }
|
| 5680 | write_unlock(&ei->i_prealloc_lock);
|
| 5681 |
|
| 5682 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
|
| 5683 | BUG_ON(pa->pa_type != MB_INODE_PA);
|
| 5684 | group = ext4_get_group_number(sb, pa->pa_pstart);
|
| 5685 |
|
| 5686 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
|
| 5687 | GFP_NOFS|__GFP_NOFAIL);
|
| 5688 | if (err) {
|
| 5689 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
|
| 5690 | err, group);
|
| 5691 | continue;
|
| 5692 | }
|
| 5693 |
|
| 5694 | bitmap_bh = ext4_read_block_bitmap(sb, group);
|
| 5695 | if (IS_ERR(bitmap_bh)) {
|
| 5696 | err = PTR_ERR(bitmap_bh);
|
| 5697 | ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
|
| 5698 | err, group);
|
| 5699 | ext4_mb_unload_buddy(&e4b);
|
| 5700 | continue;
|
| 5701 | }
|
| 5702 |
|
| 5703 | ext4_lock_group(sb, group);
|
| 5704 | list_del(&pa->pa_group_list);
|
| 5705 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
|
| 5706 | ext4_unlock_group(sb, group);
|
| 5707 |
|
| 5708 | ext4_mb_unload_buddy(&e4b);
|
| 5709 | put_bh(bitmap_bh);
|
| 5710 |
|
| 5711 | list_del(&pa->u.pa_tmp_list);
|
| 5712 | ext4_mb_pa_free(pa);
|
| 5713 | }
|
| 5714 | }
|
| 5715 |
|
| 5716 | static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
|
| 5717 | {
|
| 5718 | struct ext4_prealloc_space *pa;
|
| 5719 |
|
| 5720 | BUG_ON(ext4_pspace_cachep == NULL);
|
| 5721 | pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
|
| 5722 | if (!pa)
|
| 5723 | return -ENOMEM;
|
| 5724 | atomic_set(&pa->pa_count, 1);
|
| 5725 | ac->ac_pa = pa;
|
| 5726 | return 0;
|
| 5727 | }
|
| 5728 |
|
| 5729 | static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac)
|
| 5730 | {
|
| 5731 | struct ext4_prealloc_space *pa = ac->ac_pa;
|
| 5732 |
|
| 5733 | BUG_ON(!pa);
|
| 5734 | ac->ac_pa = NULL;
|
| 5735 | WARN_ON(!atomic_dec_and_test(&pa->pa_count));
|
| 5736 | |
| 5737 | * current function is only called due to an error or due to
|
| 5738 | * len of found blocks < len of requested blocks hence the PA has not
|
| 5739 | * been added to grp->bb_prealloc_list. So we don't need to lock it
|
| 5740 |
|
| 5741 | pa->pa_deleted = 1;
|
| 5742 | ext4_mb_pa_free(pa);
|
| 5743 | }
|
| 5744 |
|
| 5745 | #ifdef CONFIG_EXT4_DEBUG
|
| 5746 | static inline void ext4_mb_show_pa(struct super_block *sb)
|
| 5747 | {
|
| 5748 | ext4_group_t i, ngroups;
|
| 5749 |
|
| 5750 | if (ext4_emergency_state(sb))
|
| 5751 | return;
|
| 5752 |
|
| 5753 | ngroups = ext4_get_groups_count(sb);
|
| 5754 | mb_debug(sb, "groups: ");
|
| 5755 | for (i = 0; i < ngroups; i++) {
|
| 5756 | struct ext4_group_info *grp = ext4_get_group_info(sb, i);
|
| 5757 | struct ext4_prealloc_space *pa;
|
| 5758 | ext4_grpblk_t start;
|
| 5759 | struct list_head *cur;
|
| 5760 |
|
| 5761 | if (!grp)
|
| 5762 | continue;
|
| 5763 | ext4_lock_group(sb, i);
|
| 5764 | list_for_each(cur, &grp->bb_prealloc_list) {
|
| 5765 | pa = list_entry(cur, struct ext4_prealloc_space,
|
| 5766 | pa_group_list);
|
| 5767 | spin_lock(&pa->pa_lock);
|
| 5768 | ext4_get_group_no_and_offset(sb, pa->pa_pstart,
|
| 5769 | NULL, &start);
|
| 5770 | spin_unlock(&pa->pa_lock);
|
| 5771 | mb_debug(sb, "PA:%u:%d:%d\n", i, start,
|
| 5772 | pa->pa_len);
|
| 5773 | }
|
| 5774 | ext4_unlock_group(sb, i);
|
| 5775 | mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
|
| 5776 | grp->bb_fragments);
|
| 5777 | }
|
| 5778 | }
|
| 5779 |
|
| 5780 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
|
| 5781 | {
|
| 5782 | struct super_block *sb = ac->ac_sb;
|
| 5783 |
|
| 5784 | if (ext4_emergency_state(sb))
|
| 5785 | return;
|
| 5786 |
|
| 5787 | mb_debug(sb, "Can't allocate:"
|
| 5788 | " Allocation context details:");
|
| 5789 | mb_debug(sb, "status %u flags 0x%x",
|
| 5790 | ac->ac_status, ac->ac_flags);
|
| 5791 | mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
|
| 5792 | "goal %lu/%lu/%lu@%lu, "
|
| 5793 | "best %lu/%lu/%lu@%lu cr %d",
|
| 5794 | (unsigned long)ac->ac_o_ex.fe_group,
|
| 5795 | (unsigned long)ac->ac_o_ex.fe_start,
|
| 5796 | (unsigned long)ac->ac_o_ex.fe_len,
|
| 5797 | (unsigned long)ac->ac_o_ex.fe_logical,
|
| 5798 | (unsigned long)ac->ac_g_ex.fe_group,
|
| 5799 | (unsigned long)ac->ac_g_ex.fe_start,
|
| 5800 | (unsigned long)ac->ac_g_ex.fe_len,
|
| 5801 | (unsigned long)ac->ac_g_ex.fe_logical,
|
| 5802 | (unsigned long)ac->ac_b_ex.fe_group,
|
| 5803 | (unsigned long)ac->ac_b_ex.fe_start,
|
| 5804 | (unsigned long)ac->ac_b_ex.fe_len,
|
| 5805 | (unsigned long)ac->ac_b_ex.fe_logical,
|
| 5806 | (int)ac->ac_criteria);
|
| 5807 | mb_debug(sb, "%u found", ac->ac_found);
|
| 5808 | mb_debug(sb, "used pa: %s, ", str_yes_no(ac->ac_pa));
|
| 5809 | if (ac->ac_pa)
|
| 5810 | mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ?
|
| 5811 | "group pa" : "inode pa");
|
| 5812 | ext4_mb_show_pa(sb);
|
| 5813 | }
|
| 5814 | #else
|
| 5815 | static inline void ext4_mb_show_pa(struct super_block *sb)
|
| 5816 | {
|
| 5817 | }
|
| 5818 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
|
| 5819 | {
|
| 5820 | ext4_mb_show_pa(ac->ac_sb);
|
| 5821 | }
|
| 5822 | #endif
|
| 5823 |
|
| 5824 | |
| 5825 | * We use locality group preallocation for small size file. The size of the
|
| 5826 | * file is determined by the current size or the resulting size after
|
| 5827 | * allocation which ever is larger
|
| 5828 | *
|
| 5829 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
|
| 5830 |
|
| 5831 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
|
| 5832 | {
|
| 5833 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 5834 | int bsbits = ac->ac_sb->s_blocksize_bits;
|
| 5835 | loff_t size, isize;
|
| 5836 | bool inode_pa_eligible, group_pa_eligible;
|
| 5837 |
|
| 5838 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
|
| 5839 | return;
|
| 5840 |
|
| 5841 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
|
| 5842 | return;
|
| 5843 |
|
| 5844 | group_pa_eligible = sbi->s_mb_group_prealloc > 0;
|
| 5845 | inode_pa_eligible = true;
|
| 5846 | size = extent_logical_end(sbi, &ac->ac_o_ex);
|
| 5847 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
|
| 5848 | >> bsbits;
|
| 5849 |
|
| 5850 |
|
| 5851 | if ((size == isize) && !ext4_fs_is_busy(sbi) &&
|
| 5852 | !inode_is_open_for_write(ac->ac_inode))
|
| 5853 | inode_pa_eligible = false;
|
| 5854 |
|
| 5855 | size = max(size, isize);
|
| 5856 |
|
| 5857 | if (size > sbi->s_mb_stream_request)
|
| 5858 | group_pa_eligible = false;
|
| 5859 |
|
| 5860 | if (!group_pa_eligible) {
|
| 5861 | if (inode_pa_eligible)
|
| 5862 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
|
| 5863 | else
|
| 5864 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
|
| 5865 | return;
|
| 5866 | }
|
| 5867 |
|
| 5868 | BUG_ON(ac->ac_lg != NULL);
|
| 5869 | |
| 5870 | * locality group prealloc space are per cpu. The reason for having
|
| 5871 | * per cpu locality group is to reduce the contention between block
|
| 5872 | * request from multiple CPUs.
|
| 5873 |
|
| 5874 | ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
|
| 5875 |
|
| 5876 |
|
| 5877 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
|
| 5878 |
|
| 5879 |
|
| 5880 | mutex_lock(&ac->ac_lg->lg_mutex);
|
| 5881 | }
|
| 5882 |
|
| 5883 | static noinline_for_stack void
|
| 5884 | ext4_mb_initialize_context(struct ext4_allocation_context *ac,
|
| 5885 | struct ext4_allocation_request *ar)
|
| 5886 | {
|
| 5887 | struct super_block *sb = ar->inode->i_sb;
|
| 5888 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 5889 | struct ext4_super_block *es = sbi->s_es;
|
| 5890 | ext4_group_t group;
|
| 5891 | unsigned int len;
|
| 5892 | ext4_fsblk_t goal;
|
| 5893 | ext4_grpblk_t block;
|
| 5894 |
|
| 5895 |
|
| 5896 | len = ar->len;
|
| 5897 |
|
| 5898 |
|
| 5899 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
|
| 5900 | len = EXT4_CLUSTERS_PER_GROUP(sb);
|
| 5901 |
|
| 5902 |
|
| 5903 | goal = ar->goal;
|
| 5904 | if (goal < le32_to_cpu(es->s_first_data_block) ||
|
| 5905 | goal >= ext4_blocks_count(es))
|
| 5906 | goal = le32_to_cpu(es->s_first_data_block);
|
| 5907 | ext4_get_group_no_and_offset(sb, goal, &group, &block);
|
| 5908 |
|
| 5909 |
|
| 5910 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
|
| 5911 | ac->ac_status = AC_STATUS_CONTINUE;
|
| 5912 | ac->ac_sb = sb;
|
| 5913 | ac->ac_inode = ar->inode;
|
| 5914 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
|
| 5915 | ac->ac_o_ex.fe_group = group;
|
| 5916 | ac->ac_o_ex.fe_start = block;
|
| 5917 | ac->ac_o_ex.fe_len = len;
|
| 5918 | ac->ac_g_ex = ac->ac_o_ex;
|
| 5919 | ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
|
| 5920 | ac->ac_flags = ar->flags;
|
| 5921 |
|
| 5922 | |
| 5923 | * locality group. this is a policy, actually
|
| 5924 | ext4_mb_group_or_file(ac);
|
| 5925 |
|
| 5926 | mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
|
| 5927 | "left: %u/%u, right %u/%u to %swritable\n",
|
| 5928 | (unsigned) ar->len, (unsigned) ar->logical,
|
| 5929 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
|
| 5930 | (unsigned) ar->lleft, (unsigned) ar->pleft,
|
| 5931 | (unsigned) ar->lright, (unsigned) ar->pright,
|
| 5932 | inode_is_open_for_write(ar->inode) ? "" : "non-");
|
| 5933 | }
|
| 5934 |
|
| 5935 | static noinline_for_stack void
|
| 5936 | ext4_mb_discard_lg_preallocations(struct super_block *sb,
|
| 5937 | struct ext4_locality_group *lg,
|
| 5938 | int order, int total_entries)
|
| 5939 | {
|
| 5940 | ext4_group_t group = 0;
|
| 5941 | struct ext4_buddy e4b;
|
| 5942 | LIST_HEAD(discard_list);
|
| 5943 | struct ext4_prealloc_space *pa, *tmp;
|
| 5944 |
|
| 5945 | mb_debug(sb, "discard locality group preallocation\n");
|
| 5946 |
|
| 5947 | spin_lock(&lg->lg_prealloc_lock);
|
| 5948 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
|
| 5949 | pa_node.lg_list,
|
| 5950 | lockdep_is_held(&lg->lg_prealloc_lock)) {
|
| 5951 | spin_lock(&pa->pa_lock);
|
| 5952 | if (atomic_read(&pa->pa_count)) {
|
| 5953 | |
| 5954 | * This is the pa that we just used
|
| 5955 | * for block allocation. So don't
|
| 5956 | * free that
|
| 5957 |
|
| 5958 | spin_unlock(&pa->pa_lock);
|
| 5959 | continue;
|
| 5960 | }
|
| 5961 | if (pa->pa_deleted) {
|
| 5962 | spin_unlock(&pa->pa_lock);
|
| 5963 | continue;
|
| 5964 | }
|
| 5965 |
|
| 5966 | BUG_ON(pa->pa_type != MB_GROUP_PA);
|
| 5967 |
|
| 5968 |
|
| 5969 | ext4_mb_mark_pa_deleted(sb, pa);
|
| 5970 | spin_unlock(&pa->pa_lock);
|
| 5971 |
|
| 5972 | list_del_rcu(&pa->pa_node.lg_list);
|
| 5973 | list_add(&pa->u.pa_tmp_list, &discard_list);
|
| 5974 |
|
| 5975 | total_entries--;
|
| 5976 | if (total_entries <= 5) {
|
| 5977 | |
| 5978 | * we want to keep only 5 entries
|
| 5979 | * allowing it to grow to 8. This
|
| 5980 | * mak sure we don't call discard
|
| 5981 | * soon for this list.
|
| 5982 |
|
| 5983 | break;
|
| 5984 | }
|
| 5985 | }
|
| 5986 | spin_unlock(&lg->lg_prealloc_lock);
|
| 5987 |
|
| 5988 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
|
| 5989 | int err;
|
| 5990 |
|
| 5991 | group = ext4_get_group_number(sb, pa->pa_pstart);
|
| 5992 | err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
|
| 5993 | GFP_NOFS|__GFP_NOFAIL);
|
| 5994 | if (err) {
|
| 5995 | ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
|
| 5996 | err, group);
|
| 5997 | continue;
|
| 5998 | }
|
| 5999 | ext4_lock_group(sb, group);
|
| 6000 | list_del(&pa->pa_group_list);
|
| 6001 | ext4_mb_release_group_pa(&e4b, pa);
|
| 6002 | ext4_unlock_group(sb, group);
|
| 6003 |
|
| 6004 | ext4_mb_unload_buddy(&e4b);
|
| 6005 | list_del(&pa->u.pa_tmp_list);
|
| 6006 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
|
| 6007 | }
|
| 6008 | }
|
| 6009 |
|
| 6010 | |
| 6011 | * We have incremented pa_count. So it cannot be freed at this
|
| 6012 | * point. Also we hold lg_mutex. So no parallel allocation is
|
| 6013 | * possible from this lg. That means pa_free cannot be updated.
|
| 6014 | *
|
| 6015 | * A parallel ext4_mb_discard_group_preallocations is possible.
|
| 6016 | * which can cause the lg_prealloc_list to be updated.
|
| 6017 |
|
| 6018 |
|
| 6019 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
|
| 6020 | {
|
| 6021 | int order, added = 0, lg_prealloc_count = 1;
|
| 6022 | struct super_block *sb = ac->ac_sb;
|
| 6023 | struct ext4_locality_group *lg = ac->ac_lg;
|
| 6024 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
|
| 6025 |
|
| 6026 | order = fls(pa->pa_free) - 1;
|
| 6027 | if (order > PREALLOC_TB_SIZE - 1)
|
| 6028 |
|
| 6029 | order = PREALLOC_TB_SIZE - 1;
|
| 6030 |
|
| 6031 | spin_lock(&lg->lg_prealloc_lock);
|
| 6032 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
|
| 6033 | pa_node.lg_list,
|
| 6034 | lockdep_is_held(&lg->lg_prealloc_lock)) {
|
| 6035 | spin_lock(&tmp_pa->pa_lock);
|
| 6036 | if (tmp_pa->pa_deleted) {
|
| 6037 | spin_unlock(&tmp_pa->pa_lock);
|
| 6038 | continue;
|
| 6039 | }
|
| 6040 | if (!added && pa->pa_free < tmp_pa->pa_free) {
|
| 6041 |
|
| 6042 | list_add_tail_rcu(&pa->pa_node.lg_list,
|
| 6043 | &tmp_pa->pa_node.lg_list);
|
| 6044 | added = 1;
|
| 6045 | |
| 6046 | * we want to count the total
|
| 6047 | * number of entries in the list
|
| 6048 |
|
| 6049 | }
|
| 6050 | spin_unlock(&tmp_pa->pa_lock);
|
| 6051 | lg_prealloc_count++;
|
| 6052 | }
|
| 6053 | if (!added)
|
| 6054 | list_add_tail_rcu(&pa->pa_node.lg_list,
|
| 6055 | &lg->lg_prealloc_list[order]);
|
| 6056 | spin_unlock(&lg->lg_prealloc_lock);
|
| 6057 |
|
| 6058 |
|
| 6059 | if (lg_prealloc_count > 8)
|
| 6060 | ext4_mb_discard_lg_preallocations(sb, lg,
|
| 6061 | order, lg_prealloc_count);
|
| 6062 | }
|
| 6063 |
|
| 6064 | |
| 6065 | * release all resource we used in allocation
|
| 6066 |
|
| 6067 | static void ext4_mb_release_context(struct ext4_allocation_context *ac)
|
| 6068 | {
|
| 6069 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
|
| 6070 | struct ext4_prealloc_space *pa = ac->ac_pa;
|
| 6071 | if (pa) {
|
| 6072 | if (pa->pa_type == MB_GROUP_PA) {
|
| 6073 |
|
| 6074 | spin_lock(&pa->pa_lock);
|
| 6075 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
|
| 6076 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
|
| 6077 | pa->pa_free -= ac->ac_b_ex.fe_len;
|
| 6078 | pa->pa_len -= ac->ac_b_ex.fe_len;
|
| 6079 | spin_unlock(&pa->pa_lock);
|
| 6080 |
|
| 6081 | |
| 6082 | * We want to add the pa to the right bucket.
|
| 6083 | * Remove it from the list and while adding
|
| 6084 | * make sure the list to which we are adding
|
| 6085 | * doesn't grow big.
|
| 6086 |
|
| 6087 | if (likely(pa->pa_free)) {
|
| 6088 | spin_lock(pa->pa_node_lock.lg_lock);
|
| 6089 | list_del_rcu(&pa->pa_node.lg_list);
|
| 6090 | spin_unlock(pa->pa_node_lock.lg_lock);
|
| 6091 | ext4_mb_add_n_trim(ac);
|
| 6092 | }
|
| 6093 | }
|
| 6094 |
|
| 6095 | ext4_mb_put_pa(ac, ac->ac_sb, pa);
|
| 6096 | }
|
| 6097 | if (ac->ac_bitmap_folio)
|
| 6098 | folio_put(ac->ac_bitmap_folio);
|
| 6099 | if (ac->ac_buddy_folio)
|
| 6100 | folio_put(ac->ac_buddy_folio);
|
| 6101 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
|
| 6102 | mutex_unlock(&ac->ac_lg->lg_mutex);
|
| 6103 | ext4_mb_collect_stats(ac);
|
| 6104 | }
|
| 6105 |
|
| 6106 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
|
| 6107 | {
|
| 6108 | ext4_group_t i, ngroups = ext4_get_groups_count(sb);
|
| 6109 | int ret;
|
| 6110 | int freed = 0, busy = 0;
|
| 6111 | int retry = 0;
|
| 6112 |
|
| 6113 | trace_ext4_mb_discard_preallocations(sb, needed);
|
| 6114 |
|
| 6115 | if (needed == 0)
|
| 6116 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
|
| 6117 | repeat:
|
| 6118 | for (i = 0; i < ngroups && needed > 0; i++) {
|
| 6119 | ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
|
| 6120 | freed += ret;
|
| 6121 | needed -= ret;
|
| 6122 | cond_resched();
|
| 6123 | }
|
| 6124 |
|
| 6125 | if (needed > 0 && busy && ++retry < 3) {
|
| 6126 | busy = 0;
|
| 6127 | goto repeat;
|
| 6128 | }
|
| 6129 |
|
| 6130 | return freed;
|
| 6131 | }
|
| 6132 |
|
| 6133 | static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
|
| 6134 | struct ext4_allocation_context *ac, u64 *seq)
|
| 6135 | {
|
| 6136 | int freed;
|
| 6137 | u64 seq_retry = 0;
|
| 6138 | bool ret = false;
|
| 6139 |
|
| 6140 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
|
| 6141 | if (freed) {
|
| 6142 | ret = true;
|
| 6143 | goto out_dbg;
|
| 6144 | }
|
| 6145 | seq_retry = ext4_get_discard_pa_seq_sum();
|
| 6146 | if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
|
| 6147 | ac->ac_flags |= EXT4_MB_STRICT_CHECK;
|
| 6148 | *seq = seq_retry;
|
| 6149 | ret = true;
|
| 6150 | }
|
| 6151 |
|
| 6152 | out_dbg:
|
| 6153 | mb_debug(sb, "freed %d, retry ? %s\n", freed, str_yes_no(ret));
|
| 6154 | return ret;
|
| 6155 | }
|
| 6156 |
|
| 6157 | |
| 6158 | * Simple allocator for Ext4 fast commit replay path. It searches for blocks
|
| 6159 | * linearly starting at the goal block and also excludes the blocks which
|
| 6160 | * are going to be in use after fast commit replay.
|
| 6161 |
|
| 6162 | static ext4_fsblk_t
|
| 6163 | ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp)
|
| 6164 | {
|
| 6165 | struct buffer_head *bitmap_bh;
|
| 6166 | struct super_block *sb = ar->inode->i_sb;
|
| 6167 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 6168 | ext4_group_t group, nr;
|
| 6169 | ext4_grpblk_t blkoff;
|
| 6170 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
|
| 6171 | ext4_grpblk_t i = 0;
|
| 6172 | ext4_fsblk_t goal, block;
|
| 6173 | struct ext4_super_block *es = sbi->s_es;
|
| 6174 |
|
| 6175 | goal = ar->goal;
|
| 6176 | if (goal < le32_to_cpu(es->s_first_data_block) ||
|
| 6177 | goal >= ext4_blocks_count(es))
|
| 6178 | goal = le32_to_cpu(es->s_first_data_block);
|
| 6179 |
|
| 6180 | ar->len = 0;
|
| 6181 | ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
|
| 6182 | for (nr = ext4_get_groups_count(sb); nr > 0; nr--) {
|
| 6183 | bitmap_bh = ext4_read_block_bitmap(sb, group);
|
| 6184 | if (IS_ERR(bitmap_bh)) {
|
| 6185 | *errp = PTR_ERR(bitmap_bh);
|
| 6186 | pr_warn("Failed to read block bitmap\n");
|
| 6187 | return 0;
|
| 6188 | }
|
| 6189 |
|
| 6190 | while (1) {
|
| 6191 | i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
|
| 6192 | blkoff);
|
| 6193 | if (i >= max)
|
| 6194 | break;
|
| 6195 | if (ext4_fc_replay_check_excluded(sb,
|
| 6196 | ext4_group_first_block_no(sb, group) +
|
| 6197 | EXT4_C2B(sbi, i))) {
|
| 6198 | blkoff = i + 1;
|
| 6199 | } else
|
| 6200 | break;
|
| 6201 | }
|
| 6202 | brelse(bitmap_bh);
|
| 6203 | if (i < max)
|
| 6204 | break;
|
| 6205 |
|
| 6206 | if (++group >= ext4_get_groups_count(sb))
|
| 6207 | group = 0;
|
| 6208 |
|
| 6209 | blkoff = 0;
|
| 6210 | }
|
| 6211 |
|
| 6212 | if (i >= max) {
|
| 6213 | *errp = -ENOSPC;
|
| 6214 | return 0;
|
| 6215 | }
|
| 6216 |
|
| 6217 | block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i);
|
| 6218 | ext4_mb_mark_bb(sb, block, 1, true);
|
| 6219 | ar->len = 1;
|
| 6220 |
|
| 6221 | *errp = 0;
|
| 6222 | return block;
|
| 6223 | }
|
| 6224 |
|
| 6225 | |
| 6226 | * Main entry point into mballoc to allocate blocks
|
| 6227 | * it tries to use preallocation first, then falls back
|
| 6228 | * to usual allocation
|
| 6229 |
|
| 6230 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
|
| 6231 | struct ext4_allocation_request *ar, int *errp)
|
| 6232 | {
|
| 6233 | struct ext4_allocation_context *ac = NULL;
|
| 6234 | struct ext4_sb_info *sbi;
|
| 6235 | struct super_block *sb;
|
| 6236 | ext4_fsblk_t block = 0;
|
| 6237 | unsigned int inquota = 0;
|
| 6238 | unsigned int reserv_clstrs = 0;
|
| 6239 | int retries = 0;
|
| 6240 | u64 seq;
|
| 6241 |
|
| 6242 | might_sleep();
|
| 6243 | sb = ar->inode->i_sb;
|
| 6244 | sbi = EXT4_SB(sb);
|
| 6245 |
|
| 6246 | trace_ext4_request_blocks(ar);
|
| 6247 | if (sbi->s_mount_state & EXT4_FC_REPLAY)
|
| 6248 | return ext4_mb_new_blocks_simple(ar, errp);
|
| 6249 |
|
| 6250 |
|
| 6251 | if (ext4_is_quota_file(ar->inode))
|
| 6252 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
|
| 6253 |
|
| 6254 | if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
|
| 6255 | |
| 6256 | * there is enough free blocks to do block allocation
|
| 6257 | * and verify allocation doesn't exceed the quota limits.
|
| 6258 |
|
| 6259 | while (ar->len &&
|
| 6260 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
|
| 6261 |
|
| 6262 |
|
| 6263 | cond_resched();
|
| 6264 | ar->len = ar->len >> 1;
|
| 6265 | }
|
| 6266 | if (!ar->len) {
|
| 6267 | ext4_mb_show_pa(sb);
|
| 6268 | *errp = -ENOSPC;
|
| 6269 | return 0;
|
| 6270 | }
|
| 6271 | reserv_clstrs = ar->len;
|
| 6272 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
|
| 6273 | dquot_alloc_block_nofail(ar->inode,
|
| 6274 | EXT4_C2B(sbi, ar->len));
|
| 6275 | } else {
|
| 6276 | while (ar->len &&
|
| 6277 | dquot_alloc_block(ar->inode,
|
| 6278 | EXT4_C2B(sbi, ar->len))) {
|
| 6279 |
|
| 6280 | ar->flags |= EXT4_MB_HINT_NOPREALLOC;
|
| 6281 | ar->len--;
|
| 6282 | }
|
| 6283 | }
|
| 6284 | inquota = ar->len;
|
| 6285 | if (ar->len == 0) {
|
| 6286 | *errp = -EDQUOT;
|
| 6287 | goto out;
|
| 6288 | }
|
| 6289 | }
|
| 6290 |
|
| 6291 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
|
| 6292 | if (!ac) {
|
| 6293 | ar->len = 0;
|
| 6294 | *errp = -ENOMEM;
|
| 6295 | goto out;
|
| 6296 | }
|
| 6297 |
|
| 6298 | ext4_mb_initialize_context(ac, ar);
|
| 6299 |
|
| 6300 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
|
| 6301 | seq = this_cpu_read(discard_pa_seq);
|
| 6302 | if (!ext4_mb_use_preallocated(ac)) {
|
| 6303 | ac->ac_op = EXT4_MB_HISTORY_ALLOC;
|
| 6304 | ext4_mb_normalize_request(ac, ar);
|
| 6305 |
|
| 6306 | *errp = ext4_mb_pa_alloc(ac);
|
| 6307 | if (*errp)
|
| 6308 | goto errout;
|
| 6309 | repeat:
|
| 6310 |
|
| 6311 | *errp = ext4_mb_regular_allocator(ac);
|
| 6312 | |
| 6313 | * pa allocated above is added to grp->bb_prealloc_list only
|
| 6314 | * when we were able to allocate some block i.e. when
|
| 6315 | * ac->ac_status == AC_STATUS_FOUND.
|
| 6316 | * And error from above mean ac->ac_status != AC_STATUS_FOUND
|
| 6317 | * So we have to free this pa here itself.
|
| 6318 |
|
| 6319 | if (*errp) {
|
| 6320 | ext4_mb_pa_put_free(ac);
|
| 6321 | ext4_discard_allocated_blocks(ac);
|
| 6322 | goto errout;
|
| 6323 | }
|
| 6324 | if (ac->ac_status == AC_STATUS_FOUND &&
|
| 6325 | ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
|
| 6326 | ext4_mb_pa_put_free(ac);
|
| 6327 | }
|
| 6328 | if (likely(ac->ac_status == AC_STATUS_FOUND)) {
|
| 6329 | *errp = ext4_mb_mark_diskspace_used(ac, handle);
|
| 6330 | if (*errp) {
|
| 6331 | ext4_discard_allocated_blocks(ac);
|
| 6332 | goto errout;
|
| 6333 | } else {
|
| 6334 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
|
| 6335 | ar->len = ac->ac_b_ex.fe_len;
|
| 6336 | }
|
| 6337 | } else {
|
| 6338 | if (++retries < 3 &&
|
| 6339 | ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
|
| 6340 | goto repeat;
|
| 6341 | |
| 6342 | * If block allocation fails then the pa allocated above
|
| 6343 | * needs to be freed here itself.
|
| 6344 |
|
| 6345 | ext4_mb_pa_put_free(ac);
|
| 6346 | *errp = -ENOSPC;
|
| 6347 | }
|
| 6348 |
|
| 6349 | if (*errp) {
|
| 6350 | errout:
|
| 6351 | ac->ac_b_ex.fe_len = 0;
|
| 6352 | ar->len = 0;
|
| 6353 | ext4_mb_show_ac(ac);
|
| 6354 | }
|
| 6355 | ext4_mb_release_context(ac);
|
| 6356 | kmem_cache_free(ext4_ac_cachep, ac);
|
| 6357 | out:
|
| 6358 | if (inquota && ar->len < inquota)
|
| 6359 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
|
| 6360 |
|
| 6361 | if (reserv_clstrs)
|
| 6362 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, reserv_clstrs);
|
| 6363 |
|
| 6364 | trace_ext4_allocate_blocks(ar, (unsigned long long)block);
|
| 6365 |
|
| 6366 | return block;
|
| 6367 | }
|
| 6368 |
|
| 6369 | |
| 6370 | * We can merge two free data extents only if the physical blocks
|
| 6371 | * are contiguous, AND the extents were freed by the same transaction,
|
| 6372 | * AND the blocks are associated with the same group.
|
| 6373 |
|
| 6374 | static inline bool
|
| 6375 | ext4_freed_extents_can_be_merged(struct ext4_free_data *entry1,
|
| 6376 | struct ext4_free_data *entry2)
|
| 6377 | {
|
| 6378 | if (entry1->efd_tid != entry2->efd_tid)
|
| 6379 | return false;
|
| 6380 | if (entry1->efd_start_cluster + entry1->efd_count !=
|
| 6381 | entry2->efd_start_cluster)
|
| 6382 | return false;
|
| 6383 | if (WARN_ON_ONCE(entry1->efd_group != entry2->efd_group))
|
| 6384 | return false;
|
| 6385 | return true;
|
| 6386 | }
|
| 6387 |
|
| 6388 | static inline void
|
| 6389 | ext4_merge_freed_extents(struct ext4_sb_info *sbi, struct rb_root *root,
|
| 6390 | struct ext4_free_data *entry1,
|
| 6391 | struct ext4_free_data *entry2)
|
| 6392 | {
|
| 6393 | entry1->efd_count += entry2->efd_count;
|
| 6394 | spin_lock(&sbi->s_md_lock);
|
| 6395 | list_del(&entry2->efd_list);
|
| 6396 | spin_unlock(&sbi->s_md_lock);
|
| 6397 | rb_erase(&entry2->efd_node, root);
|
| 6398 | kmem_cache_free(ext4_free_data_cachep, entry2);
|
| 6399 | }
|
| 6400 |
|
| 6401 | static inline void
|
| 6402 | ext4_try_merge_freed_extent_prev(struct ext4_sb_info *sbi, struct rb_root *root,
|
| 6403 | struct ext4_free_data *entry)
|
| 6404 | {
|
| 6405 | struct ext4_free_data *prev;
|
| 6406 | struct rb_node *node;
|
| 6407 |
|
| 6408 | node = rb_prev(&entry->efd_node);
|
| 6409 | if (!node)
|
| 6410 | return;
|
| 6411 |
|
| 6412 | prev = rb_entry(node, struct ext4_free_data, efd_node);
|
| 6413 | if (ext4_freed_extents_can_be_merged(prev, entry))
|
| 6414 | ext4_merge_freed_extents(sbi, root, prev, entry);
|
| 6415 | }
|
| 6416 |
|
| 6417 | static inline void
|
| 6418 | ext4_try_merge_freed_extent_next(struct ext4_sb_info *sbi, struct rb_root *root,
|
| 6419 | struct ext4_free_data *entry)
|
| 6420 | {
|
| 6421 | struct ext4_free_data *next;
|
| 6422 | struct rb_node *node;
|
| 6423 |
|
| 6424 | node = rb_next(&entry->efd_node);
|
| 6425 | if (!node)
|
| 6426 | return;
|
| 6427 |
|
| 6428 | next = rb_entry(node, struct ext4_free_data, efd_node);
|
| 6429 | if (ext4_freed_extents_can_be_merged(entry, next))
|
| 6430 | ext4_merge_freed_extents(sbi, root, entry, next);
|
| 6431 | }
|
| 6432 |
|
| 6433 | static noinline_for_stack void
|
| 6434 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
|
| 6435 | struct ext4_free_data *new_entry)
|
| 6436 | {
|
| 6437 | ext4_group_t group = e4b->bd_group;
|
| 6438 | ext4_grpblk_t cluster;
|
| 6439 | ext4_grpblk_t clusters = new_entry->efd_count;
|
| 6440 | struct ext4_free_data *entry = NULL;
|
| 6441 | struct ext4_group_info *db = e4b->bd_info;
|
| 6442 | struct super_block *sb = e4b->bd_sb;
|
| 6443 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 6444 | struct rb_root *root = &db->bb_free_root;
|
| 6445 | struct rb_node **n = &root->rb_node;
|
| 6446 | struct rb_node *parent = NULL, *new_node;
|
| 6447 |
|
| 6448 | BUG_ON(!ext4_handle_valid(handle));
|
| 6449 | BUG_ON(e4b->bd_bitmap_folio == NULL);
|
| 6450 | BUG_ON(e4b->bd_buddy_folio == NULL);
|
| 6451 |
|
| 6452 | new_node = &new_entry->efd_node;
|
| 6453 | cluster = new_entry->efd_start_cluster;
|
| 6454 |
|
| 6455 | if (!*n) {
|
| 6456 | |
| 6457 | protect buddy cache from being freed,
|
| 6458 | * otherwise we'll refresh it from
|
| 6459 | * on-disk bitmap and lose not-yet-available
|
| 6460 | * blocks
|
| 6461 | folio_get(e4b->bd_buddy_folio);
|
| 6462 | folio_get(e4b->bd_bitmap_folio);
|
| 6463 | }
|
| 6464 | while (*n) {
|
| 6465 | parent = *n;
|
| 6466 | entry = rb_entry(parent, struct ext4_free_data, efd_node);
|
| 6467 | if (cluster < entry->efd_start_cluster)
|
| 6468 | n = &(*n)->rb_left;
|
| 6469 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
|
| 6470 | n = &(*n)->rb_right;
|
| 6471 | else {
|
| 6472 | ext4_grp_locked_error(sb, group, 0,
|
| 6473 | ext4_group_first_block_no(sb, group) +
|
| 6474 | EXT4_C2B(sbi, cluster),
|
| 6475 | "Block already on to-be-freed list");
|
| 6476 | kmem_cache_free(ext4_free_data_cachep, new_entry);
|
| 6477 | return;
|
| 6478 | }
|
| 6479 | }
|
| 6480 |
|
| 6481 | atomic_add(clusters, &sbi->s_mb_free_pending);
|
| 6482 | if (!entry)
|
| 6483 | goto insert;
|
| 6484 |
|
| 6485 |
|
| 6486 | if (ext4_freed_extents_can_be_merged(new_entry, entry)) {
|
| 6487 | entry->efd_start_cluster = cluster;
|
| 6488 | entry->efd_count += new_entry->efd_count;
|
| 6489 | kmem_cache_free(ext4_free_data_cachep, new_entry);
|
| 6490 | ext4_try_merge_freed_extent_prev(sbi, root, entry);
|
| 6491 | return;
|
| 6492 | }
|
| 6493 | if (ext4_freed_extents_can_be_merged(entry, new_entry)) {
|
| 6494 | entry->efd_count += new_entry->efd_count;
|
| 6495 | kmem_cache_free(ext4_free_data_cachep, new_entry);
|
| 6496 | ext4_try_merge_freed_extent_next(sbi, root, entry);
|
| 6497 | return;
|
| 6498 | }
|
| 6499 | insert:
|
| 6500 | rb_link_node(new_node, parent, n);
|
| 6501 | rb_insert_color(new_node, root);
|
| 6502 |
|
| 6503 | spin_lock(&sbi->s_md_lock);
|
| 6504 | list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]);
|
| 6505 | spin_unlock(&sbi->s_md_lock);
|
| 6506 | }
|
| 6507 |
|
| 6508 | static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
|
| 6509 | unsigned long count)
|
| 6510 | {
|
| 6511 | struct super_block *sb = inode->i_sb;
|
| 6512 | ext4_group_t group;
|
| 6513 | ext4_grpblk_t blkoff;
|
| 6514 |
|
| 6515 | ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
|
| 6516 | ext4_mb_mark_context(NULL, sb, false, group, blkoff, count,
|
| 6517 | EXT4_MB_BITMAP_MARKED_CHECK |
|
| 6518 | EXT4_MB_SYNC_UPDATE,
|
| 6519 | NULL);
|
| 6520 | }
|
| 6521 |
|
| 6522 | |
| 6523 | * ext4_mb_clear_bb() -- helper function for freeing blocks.
|
| 6524 | * Used by ext4_free_blocks()
|
| 6525 | * @handle: handle for this transaction
|
| 6526 | * @inode: inode
|
| 6527 | * @block: starting physical block to be freed
|
| 6528 | * @count: number of blocks to be freed
|
| 6529 | * @flags: flags used by ext4_free_blocks
|
| 6530 |
|
| 6531 | static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
|
| 6532 | ext4_fsblk_t block, unsigned long count,
|
| 6533 | int flags)
|
| 6534 | {
|
| 6535 | struct super_block *sb = inode->i_sb;
|
| 6536 | struct ext4_group_info *grp;
|
| 6537 | unsigned int overflow;
|
| 6538 | ext4_grpblk_t bit;
|
| 6539 | ext4_group_t block_group;
|
| 6540 | struct ext4_sb_info *sbi;
|
| 6541 | struct ext4_buddy e4b;
|
| 6542 | unsigned int count_clusters;
|
| 6543 | int err = 0;
|
| 6544 | int mark_flags = 0;
|
| 6545 | ext4_grpblk_t changed;
|
| 6546 |
|
| 6547 | sbi = EXT4_SB(sb);
|
| 6548 |
|
| 6549 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
|
| 6550 | !ext4_inode_block_valid(inode, block, count)) {
|
| 6551 | ext4_error(sb, "Freeing blocks in system zone - "
|
| 6552 | "Block = %llu, count = %lu", block, count);
|
| 6553 |
|
| 6554 | goto error_out;
|
| 6555 | }
|
| 6556 | flags |= EXT4_FREE_BLOCKS_VALIDATED;
|
| 6557 |
|
| 6558 | do_more:
|
| 6559 | overflow = 0;
|
| 6560 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
|
| 6561 |
|
| 6562 | grp = ext4_get_group_info(sb, block_group);
|
| 6563 | if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
|
| 6564 | return;
|
| 6565 |
|
| 6566 | |
| 6567 | * Check to see if we are freeing blocks across a group
|
| 6568 | * boundary.
|
| 6569 |
|
| 6570 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
|
| 6571 | overflow = EXT4_C2B(sbi, bit) + count -
|
| 6572 | EXT4_BLOCKS_PER_GROUP(sb);
|
| 6573 | count -= overflow;
|
| 6574 |
|
| 6575 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
|
| 6576 | }
|
| 6577 | count_clusters = EXT4_NUM_B2C(sbi, count);
|
| 6578 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
|
| 6579 |
|
| 6580 |
|
| 6581 | err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
|
| 6582 | GFP_NOFS|__GFP_NOFAIL);
|
| 6583 | if (err)
|
| 6584 | goto error_out;
|
| 6585 |
|
| 6586 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
|
| 6587 | !ext4_inode_block_valid(inode, block, count)) {
|
| 6588 | ext4_error(sb, "Freeing blocks in system zone - "
|
| 6589 | "Block = %llu, count = %lu", block, count);
|
| 6590 |
|
| 6591 | goto error_clean;
|
| 6592 | }
|
| 6593 |
|
| 6594 | #ifdef AGGRESSIVE_CHECK
|
| 6595 | mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK;
|
| 6596 | #endif
|
| 6597 | err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
|
| 6598 | count_clusters, mark_flags, &changed);
|
| 6599 |
|
| 6600 |
|
| 6601 | if (err && changed == 0)
|
| 6602 | goto error_clean;
|
| 6603 |
|
| 6604 | #ifdef AGGRESSIVE_CHECK
|
| 6605 | BUG_ON(changed != count_clusters);
|
| 6606 | #endif
|
| 6607 |
|
| 6608 | |
| 6609 | * We need to make sure we don't reuse the freed block until after the
|
| 6610 | * transaction is committed. We make an exception if the inode is to be
|
| 6611 | * written in writeback mode since writeback mode has weak data
|
| 6612 | * consistency guarantees.
|
| 6613 |
|
| 6614 | if (ext4_handle_valid(handle) &&
|
| 6615 | ((flags & EXT4_FREE_BLOCKS_METADATA) ||
|
| 6616 | !ext4_should_writeback_data(inode))) {
|
| 6617 | struct ext4_free_data *new_entry;
|
| 6618 | |
| 6619 | * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
|
| 6620 | * to fail.
|
| 6621 |
|
| 6622 | new_entry = kmem_cache_alloc(ext4_free_data_cachep,
|
| 6623 | GFP_NOFS|__GFP_NOFAIL);
|
| 6624 | new_entry->efd_start_cluster = bit;
|
| 6625 | new_entry->efd_group = block_group;
|
| 6626 | new_entry->efd_count = count_clusters;
|
| 6627 | new_entry->efd_tid = handle->h_transaction->t_tid;
|
| 6628 |
|
| 6629 | ext4_lock_group(sb, block_group);
|
| 6630 | ext4_mb_free_metadata(handle, &e4b, new_entry);
|
| 6631 | } else {
|
| 6632 | if (test_opt(sb, DISCARD)) {
|
| 6633 | err = ext4_issue_discard(sb, block_group, bit,
|
| 6634 | count_clusters);
|
| 6635 | |
| 6636 | * Ignore EOPNOTSUPP error. This is consistent with
|
| 6637 | * what happens when using journal.
|
| 6638 |
|
| 6639 | if (err == -EOPNOTSUPP)
|
| 6640 | err = 0;
|
| 6641 | if (err)
|
| 6642 | ext4_msg(sb, KERN_WARNING, "discard request in"
|
| 6643 | " group:%u block:%d count:%lu failed"
|
| 6644 | " with %d", block_group, bit, count,
|
| 6645 | err);
|
| 6646 | }
|
| 6647 |
|
| 6648 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
|
| 6649 |
|
| 6650 | ext4_lock_group(sb, block_group);
|
| 6651 | mb_free_blocks(inode, &e4b, bit, count_clusters);
|
| 6652 | }
|
| 6653 |
|
| 6654 | ext4_unlock_group(sb, block_group);
|
| 6655 |
|
| 6656 | |
| 6657 | * on a bigalloc file system, defer the s_freeclusters_counter
|
| 6658 | * update to the caller (ext4_remove_space and friends) so they
|
| 6659 | * can determine if a cluster freed here should be rereserved
|
| 6660 |
|
| 6661 | if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
|
| 6662 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
|
| 6663 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
|
| 6664 | percpu_counter_add(&sbi->s_freeclusters_counter,
|
| 6665 | count_clusters);
|
| 6666 | }
|
| 6667 |
|
| 6668 | if (overflow && !err) {
|
| 6669 | block += count;
|
| 6670 | count = overflow;
|
| 6671 | ext4_mb_unload_buddy(&e4b);
|
| 6672 |
|
| 6673 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
|
| 6674 | goto do_more;
|
| 6675 | }
|
| 6676 |
|
| 6677 | error_clean:
|
| 6678 | ext4_mb_unload_buddy(&e4b);
|
| 6679 | error_out:
|
| 6680 | ext4_std_error(sb, err);
|
| 6681 | }
|
| 6682 |
|
| 6683 | |
| 6684 | * ext4_free_blocks() -- Free given blocks and update quota
|
| 6685 | * @handle: handle for this transaction
|
| 6686 | * @inode: inode
|
| 6687 | * @bh: optional buffer of the block to be freed
|
| 6688 | * @block: starting physical block to be freed
|
| 6689 | * @count: number of blocks to be freed
|
| 6690 | * @flags: flags used by ext4_free_blocks
|
| 6691 |
|
| 6692 | void ext4_free_blocks(handle_t *handle, struct inode *inode,
|
| 6693 | struct buffer_head *bh, ext4_fsblk_t block,
|
| 6694 | unsigned long count, int flags)
|
| 6695 | {
|
| 6696 | struct super_block *sb = inode->i_sb;
|
| 6697 | unsigned int overflow;
|
| 6698 | struct ext4_sb_info *sbi;
|
| 6699 |
|
| 6700 | sbi = EXT4_SB(sb);
|
| 6701 |
|
| 6702 | if (bh) {
|
| 6703 | if (block)
|
| 6704 | BUG_ON(block != bh->b_blocknr);
|
| 6705 | else
|
| 6706 | block = bh->b_blocknr;
|
| 6707 | }
|
| 6708 |
|
| 6709 | if (sbi->s_mount_state & EXT4_FC_REPLAY) {
|
| 6710 | ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count));
|
| 6711 | return;
|
| 6712 | }
|
| 6713 |
|
| 6714 | might_sleep();
|
| 6715 |
|
| 6716 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
|
| 6717 | !ext4_inode_block_valid(inode, block, count)) {
|
| 6718 | ext4_error(sb, "Freeing blocks not in datazone - "
|
| 6719 | "block = %llu, count = %lu", block, count);
|
| 6720 | return;
|
| 6721 | }
|
| 6722 | flags |= EXT4_FREE_BLOCKS_VALIDATED;
|
| 6723 |
|
| 6724 | ext4_debug("freeing block %llu\n", block);
|
| 6725 | trace_ext4_free_blocks(inode, block, count, flags);
|
| 6726 |
|
| 6727 | if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
|
| 6728 | BUG_ON(count > 1);
|
| 6729 |
|
| 6730 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
|
| 6731 | inode, bh, block);
|
| 6732 | }
|
| 6733 |
|
| 6734 | |
| 6735 | * If the extent to be freed does not begin on a cluster
|
| 6736 | * boundary, we need to deal with partial clusters at the
|
| 6737 | * beginning and end of the extent. Normally we will free
|
| 6738 | * blocks at the beginning or the end unless we are explicitly
|
| 6739 | * requested to avoid doing so.
|
| 6740 |
|
| 6741 | overflow = EXT4_PBLK_COFF(sbi, block);
|
| 6742 | if (overflow) {
|
| 6743 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
|
| 6744 | overflow = sbi->s_cluster_ratio - overflow;
|
| 6745 | block += overflow;
|
| 6746 | if (count > overflow)
|
| 6747 | count -= overflow;
|
| 6748 | else
|
| 6749 | return;
|
| 6750 | } else {
|
| 6751 | block -= overflow;
|
| 6752 | count += overflow;
|
| 6753 | }
|
| 6754 |
|
| 6755 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
|
| 6756 | }
|
| 6757 | overflow = EXT4_LBLK_COFF(sbi, count);
|
| 6758 | if (overflow) {
|
| 6759 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
|
| 6760 | if (count > overflow)
|
| 6761 | count -= overflow;
|
| 6762 | else
|
| 6763 | return;
|
| 6764 | } else
|
| 6765 | count += sbi->s_cluster_ratio - overflow;
|
| 6766 |
|
| 6767 | flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
|
| 6768 | }
|
| 6769 |
|
| 6770 | if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
|
| 6771 | int i;
|
| 6772 | int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
|
| 6773 |
|
| 6774 | for (i = 0; i < count; i++) {
|
| 6775 | cond_resched();
|
| 6776 | if (is_metadata)
|
| 6777 | bh = sb_find_get_block_nonatomic(inode->i_sb,
|
| 6778 | block + i);
|
| 6779 | ext4_forget(handle, is_metadata, inode, bh, block + i);
|
| 6780 | }
|
| 6781 | }
|
| 6782 |
|
| 6783 | ext4_mb_clear_bb(handle, inode, block, count, flags);
|
| 6784 | }
|
| 6785 |
|
| 6786 | |
| 6787 | * ext4_group_add_blocks() -- Add given blocks to an existing group
|
| 6788 | * @handle: handle to this transaction
|
| 6789 | * @sb: super block
|
| 6790 | * @block: start physical block to add to the block group
|
| 6791 | * @count: number of blocks to free
|
| 6792 | *
|
| 6793 | * This marks the blocks as free in the bitmap and buddy.
|
| 6794 |
|
| 6795 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
|
| 6796 | ext4_fsblk_t block, unsigned long count)
|
| 6797 | {
|
| 6798 | ext4_group_t block_group;
|
| 6799 | ext4_grpblk_t bit;
|
| 6800 | struct ext4_sb_info *sbi = EXT4_SB(sb);
|
| 6801 | struct ext4_buddy e4b;
|
| 6802 | int err = 0;
|
| 6803 | ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
|
| 6804 | ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
|
| 6805 | unsigned long cluster_count = last_cluster - first_cluster + 1;
|
| 6806 | ext4_grpblk_t changed;
|
| 6807 |
|
| 6808 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
|
| 6809 |
|
| 6810 | if (cluster_count == 0)
|
| 6811 | return 0;
|
| 6812 |
|
| 6813 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
|
| 6814 | |
| 6815 | * Check to see if we are freeing blocks across a group
|
| 6816 | * boundary.
|
| 6817 |
|
| 6818 | if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
|
| 6819 | ext4_warning(sb, "too many blocks added to group %u",
|
| 6820 | block_group);
|
| 6821 | err = -EINVAL;
|
| 6822 | goto error_out;
|
| 6823 | }
|
| 6824 |
|
| 6825 | err = ext4_mb_load_buddy(sb, block_group, &e4b);
|
| 6826 | if (err)
|
| 6827 | goto error_out;
|
| 6828 |
|
| 6829 | if (!ext4_sb_block_valid(sb, NULL, block, count)) {
|
| 6830 | ext4_error(sb, "Adding blocks in system zones - "
|
| 6831 | "Block = %llu, count = %lu",
|
| 6832 | block, count);
|
| 6833 | err = -EINVAL;
|
| 6834 | goto error_clean;
|
| 6835 | }
|
| 6836 |
|
| 6837 | err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
|
| 6838 | cluster_count, EXT4_MB_BITMAP_MARKED_CHECK,
|
| 6839 | &changed);
|
| 6840 | if (err && changed == 0)
|
| 6841 | goto error_clean;
|
| 6842 |
|
| 6843 | if (changed != cluster_count)
|
| 6844 | ext4_error(sb, "bit already cleared in group %u", block_group);
|
| 6845 |
|
| 6846 | ext4_lock_group(sb, block_group);
|
| 6847 | mb_free_blocks(NULL, &e4b, bit, cluster_count);
|
| 6848 | ext4_unlock_group(sb, block_group);
|
| 6849 | percpu_counter_add(&sbi->s_freeclusters_counter,
|
| 6850 | changed);
|
| 6851 |
|
| 6852 | error_clean:
|
| 6853 | ext4_mb_unload_buddy(&e4b);
|
| 6854 | error_out:
|
| 6855 | ext4_std_error(sb, err);
|
| 6856 | return err;
|
| 6857 | }
|
| 6858 |
|
| 6859 | |
| 6860 | * ext4_trim_extent -- function to TRIM one single free extent in the group
|
| 6861 | * @sb: super block for the file system
|
| 6862 | * @start: starting block of the free extent in the alloc. group
|
| 6863 | * @count: number of blocks to TRIM
|
| 6864 | * @e4b: ext4 buddy for the group
|
| 6865 | *
|
| 6866 | * Trim "count" blocks starting at "start" in the "group". To assure that no
|
| 6867 | * one will allocate those blocks, mark it as used in buddy bitmap. This must
|
| 6868 | * be called with under the group lock.
|
| 6869 |
|
| 6870 | static int ext4_trim_extent(struct super_block *sb,
|
| 6871 | int start, int count, struct ext4_buddy *e4b)
|
| 6872 | __releases(bitlock)
|
| 6873 | __acquires(bitlock)
|
| 6874 | {
|
| 6875 | struct ext4_free_extent ex;
|
| 6876 | ext4_group_t group = e4b->bd_group;
|
| 6877 | int ret = 0;
|
| 6878 |
|
| 6879 | trace_ext4_trim_extent(sb, group, start, count);
|
| 6880 |
|
| 6881 | assert_spin_locked(ext4_group_lock_ptr(sb, group));
|
| 6882 |
|
| 6883 | ex.fe_start = start;
|
| 6884 | ex.fe_group = group;
|
| 6885 | ex.fe_len = count;
|
| 6886 |
|
| 6887 | |
| 6888 | * Mark blocks used, so no one can reuse them while
|
| 6889 | * being trimmed.
|
| 6890 |
|
| 6891 | mb_mark_used(e4b, &ex);
|
| 6892 | ext4_unlock_group(sb, group);
|
| 6893 | ret = ext4_issue_discard(sb, group, start, count);
|
| 6894 | ext4_lock_group(sb, group);
|
| 6895 | mb_free_blocks(NULL, e4b, start, ex.fe_len);
|
| 6896 | return ret;
|
| 6897 | }
|
| 6898 |
|
| 6899 | static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb,
|
| 6900 | ext4_group_t grp)
|
| 6901 | {
|
| 6902 | unsigned long nr_clusters_in_group;
|
| 6903 |
|
| 6904 | if (grp < (ext4_get_groups_count(sb) - 1))
|
| 6905 | nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb);
|
| 6906 | else
|
| 6907 | nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) -
|
| 6908 | ext4_group_first_block_no(sb, grp))
|
| 6909 | >> EXT4_CLUSTER_BITS(sb);
|
| 6910 |
|
| 6911 | return nr_clusters_in_group - 1;
|
| 6912 | }
|
| 6913 |
|
| 6914 | static bool ext4_trim_interrupted(void)
|
| 6915 | {
|
| 6916 | return fatal_signal_pending(current) || freezing(current);
|
| 6917 | }
|
| 6918 |
|
| 6919 | static int ext4_try_to_trim_range(struct super_block *sb,
|
| 6920 | struct ext4_buddy *e4b, ext4_grpblk_t start,
|
| 6921 | ext4_grpblk_t max, ext4_grpblk_t minblocks)
|
| 6922 | __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
|
| 6923 | __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
|
| 6924 | {
|
| 6925 | ext4_grpblk_t next, count, free_count, last, origin_start;
|
| 6926 | bool set_trimmed = false;
|
| 6927 | void *bitmap;
|
| 6928 |
|
| 6929 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
|
| 6930 | return 0;
|
| 6931 |
|
| 6932 | last = ext4_last_grp_cluster(sb, e4b->bd_group);
|
| 6933 | bitmap = e4b->bd_bitmap;
|
| 6934 | if (start == 0 && max >= last)
|
| 6935 | set_trimmed = true;
|
| 6936 | origin_start = start;
|
| 6937 | start = max(e4b->bd_info->bb_first_free, start);
|
| 6938 | count = 0;
|
| 6939 | free_count = 0;
|
| 6940 |
|
| 6941 | while (start <= max) {
|
| 6942 | start = mb_find_next_zero_bit(bitmap, max + 1, start);
|
| 6943 | if (start > max)
|
| 6944 | break;
|
| 6945 |
|
| 6946 | next = mb_find_next_bit(bitmap, last + 1, start);
|
| 6947 | if (origin_start == 0 && next >= last)
|
| 6948 | set_trimmed = true;
|
| 6949 |
|
| 6950 | if ((next - start) >= minblocks) {
|
| 6951 | int ret = ext4_trim_extent(sb, start, next - start, e4b);
|
| 6952 |
|
| 6953 | if (ret && ret != -EOPNOTSUPP)
|
| 6954 | return count;
|
| 6955 | count += next - start;
|
| 6956 | }
|
| 6957 | free_count += next - start;
|
| 6958 | start = next + 1;
|
| 6959 |
|
| 6960 | if (ext4_trim_interrupted())
|
| 6961 | return count;
|
| 6962 |
|
| 6963 | if (need_resched()) {
|
| 6964 | ext4_unlock_group(sb, e4b->bd_group);
|
| 6965 | cond_resched();
|
| 6966 | ext4_lock_group(sb, e4b->bd_group);
|
| 6967 | }
|
| 6968 |
|
| 6969 | if ((e4b->bd_info->bb_free - free_count) < minblocks)
|
| 6970 | break;
|
| 6971 | }
|
| 6972 |
|
| 6973 | if (set_trimmed)
|
| 6974 | EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info);
|
| 6975 |
|
| 6976 | return count;
|
| 6977 | }
|
| 6978 |
|
| 6979 | |
| 6980 | * ext4_trim_all_free -- function to trim all free space in alloc. group
|
| 6981 | * @sb: super block for file system
|
| 6982 | * @group: group to be trimmed
|
| 6983 | * @start: first group block to examine
|
| 6984 | * @max: last group block to examine
|
| 6985 | * @minblocks: minimum extent block count
|
| 6986 | *
|
| 6987 | * ext4_trim_all_free walks through group's block bitmap searching for free
|
| 6988 | * extents. When the free extent is found, mark it as used in group buddy
|
| 6989 | * bitmap. Then issue a TRIM command on this extent and free the extent in
|
| 6990 | * the group buddy bitmap.
|
| 6991 |
|
| 6992 | static ext4_grpblk_t
|
| 6993 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
|
| 6994 | ext4_grpblk_t start, ext4_grpblk_t max,
|
| 6995 | ext4_grpblk_t minblocks)
|
| 6996 | {
|
| 6997 | struct ext4_buddy e4b;
|
| 6998 | int ret;
|
| 6999 |
|
| 7000 | trace_ext4_trim_all_free(sb, group, start, max);
|
| 7001 |
|
| 7002 | ret = ext4_mb_load_buddy(sb, group, &e4b);
|
| 7003 | if (ret) {
|
| 7004 | ext4_warning(sb, "Error %d loading buddy information for %u",
|
| 7005 | ret, group);
|
| 7006 | return ret;
|
| 7007 | }
|
| 7008 |
|
| 7009 | ext4_lock_group(sb, group);
|
| 7010 |
|
| 7011 | if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
|
| 7012 | minblocks < EXT4_SB(sb)->s_last_trim_minblks)
|
| 7013 | ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
|
| 7014 | else
|
| 7015 | ret = 0;
|
| 7016 |
|
| 7017 | ext4_unlock_group(sb, group);
|
| 7018 | ext4_mb_unload_buddy(&e4b);
|
| 7019 |
|
| 7020 | ext4_debug("trimmed %d blocks in the group %d\n",
|
| 7021 | ret, group);
|
| 7022 |
|
| 7023 | return ret;
|
| 7024 | }
|
| 7025 |
|
| 7026 | |
| 7027 | * ext4_trim_fs() -- trim ioctl handle function
|
| 7028 | * @sb: superblock for filesystem
|
| 7029 | * @range: fstrim_range structure
|
| 7030 | *
|
| 7031 | * start: First Byte to trim
|
| 7032 | * len: number of Bytes to trim from start
|
| 7033 | * minlen: minimum extent length in Bytes
|
| 7034 | * ext4_trim_fs goes through all allocation groups containing Bytes from
|
| 7035 | * start to start+len. For each such a group ext4_trim_all_free function
|
| 7036 | * is invoked to trim all free space.
|
| 7037 |
|
| 7038 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
|
| 7039 | {
|
| 7040 | unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
|
| 7041 | struct ext4_group_info *grp;
|
| 7042 | ext4_group_t group, first_group, last_group;
|
| 7043 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
|
| 7044 | uint64_t start, end, minlen, trimmed = 0;
|
| 7045 | ext4_fsblk_t first_data_blk =
|
| 7046 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
|
| 7047 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
|
| 7048 | int ret = 0;
|
| 7049 |
|
| 7050 | start = range->start >> sb->s_blocksize_bits;
|
| 7051 | end = start + (range->len >> sb->s_blocksize_bits) - 1;
|
| 7052 | minlen = EXT4_NUM_B2C(EXT4_SB(sb),
|
| 7053 | range->minlen >> sb->s_blocksize_bits);
|
| 7054 |
|
| 7055 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
|
| 7056 | start >= max_blks ||
|
| 7057 | range->len < sb->s_blocksize)
|
| 7058 | return -EINVAL;
|
| 7059 |
|
| 7060 | if (range->minlen < discard_granularity) {
|
| 7061 | minlen = EXT4_NUM_B2C(EXT4_SB(sb),
|
| 7062 | discard_granularity >> sb->s_blocksize_bits);
|
| 7063 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
|
| 7064 | goto out;
|
| 7065 | }
|
| 7066 | if (end >= max_blks - 1)
|
| 7067 | end = max_blks - 1;
|
| 7068 | if (end <= first_data_blk)
|
| 7069 | goto out;
|
| 7070 | if (start < first_data_blk)
|
| 7071 | start = first_data_blk;
|
| 7072 |
|
| 7073 |
|
| 7074 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
|
| 7075 | &first_group, &first_cluster);
|
| 7076 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
|
| 7077 | &last_group, &last_cluster);
|
| 7078 |
|
| 7079 |
|
| 7080 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
|
| 7081 |
|
| 7082 | for (group = first_group; group <= last_group; group++) {
|
| 7083 | if (ext4_trim_interrupted())
|
| 7084 | break;
|
| 7085 | grp = ext4_get_group_info(sb, group);
|
| 7086 | if (!grp)
|
| 7087 | continue;
|
| 7088 |
|
| 7089 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
|
| 7090 | ret = ext4_mb_init_group(sb, group, GFP_NOFS);
|
| 7091 | if (ret)
|
| 7092 | break;
|
| 7093 | }
|
| 7094 |
|
| 7095 | |
| 7096 | * For all the groups except the last one, last cluster will
|
| 7097 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
|
| 7098 | * change it for the last group, note that last_cluster is
|
| 7099 | * already computed earlier by ext4_get_group_no_and_offset()
|
| 7100 |
|
| 7101 | if (group == last_group)
|
| 7102 | end = last_cluster;
|
| 7103 | if (grp->bb_free >= minlen) {
|
| 7104 | cnt = ext4_trim_all_free(sb, group, first_cluster,
|
| 7105 | end, minlen);
|
| 7106 | if (cnt < 0) {
|
| 7107 | ret = cnt;
|
| 7108 | break;
|
| 7109 | }
|
| 7110 | trimmed += cnt;
|
| 7111 | }
|
| 7112 |
|
| 7113 | |
| 7114 | * For every group except the first one, we are sure
|
| 7115 | * that the first cluster to discard will be cluster #0.
|
| 7116 |
|
| 7117 | first_cluster = 0;
|
| 7118 | }
|
| 7119 |
|
| 7120 | if (!ret)
|
| 7121 | EXT4_SB(sb)->s_last_trim_minblks = minlen;
|
| 7122 |
|
| 7123 | out:
|
| 7124 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
|
| 7125 | return ret;
|
| 7126 | }
|
| 7127 |
|
| 7128 |
|
| 7129 | int
|
| 7130 | ext4_mballoc_query_range(
|
| 7131 | struct super_block *sb,
|
| 7132 | ext4_group_t group,
|
| 7133 | ext4_grpblk_t first,
|
| 7134 | ext4_grpblk_t end,
|
| 7135 | ext4_mballoc_query_range_fn meta_formatter,
|
| 7136 | ext4_mballoc_query_range_fn formatter,
|
| 7137 | void *priv)
|
| 7138 | {
|
| 7139 | void *bitmap;
|
| 7140 | ext4_grpblk_t start, next;
|
| 7141 | struct ext4_buddy e4b;
|
| 7142 | int error;
|
| 7143 |
|
| 7144 | error = ext4_mb_load_buddy(sb, group, &e4b);
|
| 7145 | if (error)
|
| 7146 | return error;
|
| 7147 | bitmap = e4b.bd_bitmap;
|
| 7148 |
|
| 7149 | ext4_lock_group(sb, group);
|
| 7150 |
|
| 7151 | start = max(e4b.bd_info->bb_first_free, first);
|
| 7152 | if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
|
| 7153 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
|
| 7154 | if (meta_formatter && start != first) {
|
| 7155 | if (start > end)
|
| 7156 | start = end;
|
| 7157 | ext4_unlock_group(sb, group);
|
| 7158 | error = meta_formatter(sb, group, first, start - first,
|
| 7159 | priv);
|
| 7160 | if (error)
|
| 7161 | goto out_unload;
|
| 7162 | ext4_lock_group(sb, group);
|
| 7163 | }
|
| 7164 | while (start <= end) {
|
| 7165 | start = mb_find_next_zero_bit(bitmap, end + 1, start);
|
| 7166 | if (start > end)
|
| 7167 | break;
|
| 7168 | next = mb_find_next_bit(bitmap, end + 1, start);
|
| 7169 |
|
| 7170 | ext4_unlock_group(sb, group);
|
| 7171 | error = formatter(sb, group, start, next - start, priv);
|
| 7172 | if (error)
|
| 7173 | goto out_unload;
|
| 7174 | ext4_lock_group(sb, group);
|
| 7175 |
|
| 7176 | start = next + 1;
|
| 7177 | }
|
| 7178 |
|
| 7179 | ext4_unlock_group(sb, group);
|
| 7180 | out_unload:
|
| 7181 | ext4_mb_unload_buddy(&e4b);
|
| 7182 |
|
| 7183 | return error;
|
| 7184 | }
|
| 7185 |
|
| 7186 | #if IS_ENABLED(CONFIG_EXT4_KUNIT_TESTS)
|
| 7187 | void mb_clear_bits_test(void *bm, int cur, int len)
|
| 7188 | {
|
| 7189 | mb_clear_bits(bm, cur, len);
|
| 7190 | }
|
| 7191 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_clear_bits_test);
|
| 7192 |
|
| 7193 | ext4_fsblk_t
|
| 7194 | ext4_mb_new_blocks_simple_test(struct ext4_allocation_request *ar,
|
| 7195 | int *errp)
|
| 7196 | {
|
| 7197 | return ext4_mb_new_blocks_simple(ar, errp);
|
| 7198 | }
|
| 7199 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_new_blocks_simple_test);
|
| 7200 |
|
| 7201 | int mb_find_next_zero_bit_test(void *addr, int max, int start)
|
| 7202 | {
|
| 7203 | return mb_find_next_zero_bit(addr, max, start);
|
| 7204 | }
|
| 7205 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_find_next_zero_bit_test);
|
| 7206 |
|
| 7207 | int mb_find_next_bit_test(void *addr, int max, int start)
|
| 7208 | {
|
| 7209 | return mb_find_next_bit(addr, max, start);
|
| 7210 | }
|
| 7211 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_find_next_bit_test);
|
| 7212 |
|
| 7213 | void mb_clear_bit_test(int bit, void *addr)
|
| 7214 | {
|
| 7215 | mb_clear_bit(bit, addr);
|
| 7216 | }
|
| 7217 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_clear_bit_test);
|
| 7218 |
|
| 7219 | int mb_test_bit_test(int bit, void *addr)
|
| 7220 | {
|
| 7221 | return mb_test_bit(bit, addr);
|
| 7222 | }
|
| 7223 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_test_bit_test);
|
| 7224 |
|
| 7225 | int ext4_mb_mark_diskspace_used_test(struct ext4_allocation_context *ac,
|
| 7226 | handle_t *handle)
|
| 7227 | {
|
| 7228 | return ext4_mb_mark_diskspace_used(ac, handle);
|
| 7229 | }
|
| 7230 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_mark_diskspace_used_test);
|
| 7231 |
|
| 7232 | int mb_mark_used_test(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
|
| 7233 | {
|
| 7234 | return mb_mark_used(e4b, ex);
|
| 7235 | }
|
| 7236 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_mark_used_test);
|
| 7237 |
|
| 7238 | void ext4_mb_generate_buddy_test(struct super_block *sb, void *buddy,
|
| 7239 | void *bitmap, ext4_group_t group,
|
| 7240 | struct ext4_group_info *grp)
|
| 7241 | {
|
| 7242 | ext4_mb_generate_buddy(sb, buddy, bitmap, group, grp);
|
| 7243 | }
|
| 7244 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_generate_buddy_test);
|
| 7245 |
|
| 7246 | int ext4_mb_load_buddy_test(struct super_block *sb, ext4_group_t group,
|
| 7247 | struct ext4_buddy *e4b)
|
| 7248 | {
|
| 7249 | return ext4_mb_load_buddy(sb, group, e4b);
|
| 7250 | }
|
| 7251 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_load_buddy_test);
|
| 7252 |
|
| 7253 | void ext4_mb_unload_buddy_test(struct ext4_buddy *e4b)
|
| 7254 | {
|
| 7255 | ext4_mb_unload_buddy(e4b);
|
| 7256 | }
|
| 7257 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_unload_buddy_test);
|
| 7258 |
|
| 7259 | void mb_free_blocks_test(struct inode *inode, struct ext4_buddy *e4b,
|
| 7260 | int first, int count)
|
| 7261 | {
|
| 7262 | mb_free_blocks(inode, e4b, first, count);
|
| 7263 | }
|
| 7264 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_free_blocks_test);
|
| 7265 |
|
| 7266 | void ext4_free_blocks_simple_test(struct inode *inode, ext4_fsblk_t block,
|
| 7267 | unsigned long count)
|
| 7268 | {
|
| 7269 | return ext4_free_blocks_simple(inode, block, count);
|
| 7270 | }
|
| 7271 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_free_blocks_simple_test);
|
| 7272 |
|
| 7273 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_wait_block_bitmap);
|
| 7274 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_init);
|
| 7275 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_get_group_desc);
|
| 7276 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_count_free_clusters);
|
| 7277 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_get_group_info);
|
| 7278 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_free_group_clusters_set);
|
| 7279 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_release);
|
| 7280 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_read_block_bitmap_nowait);
|
| 7281 | EXPORT_SYMBOL_FOR_EXT4_TEST(mb_set_bits);
|
| 7282 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_fc_init_inode);
|
| 7283 | EXPORT_SYMBOL_FOR_EXT4_TEST(ext4_mb_mark_context);
|
| 7284 | #endif
|