1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 4 * Written by Alex Tomas <alex@clusterfs.com> 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 <trace/events/ext4.h> 20 21 /* 22 * MUSTDO: 23 * - test ext4_ext_search_left() and ext4_ext_search_right() 24 * - search for metadata in few groups 25 * 26 * TODO v4: 27 * - normalization should take into account whether file is still open 28 * - discard preallocations if no free space left (policy?) 29 * - don't normalize tails 30 * - quota 31 * - reservation for superuser 32 * 33 * TODO v3: 34 * - bitmap read-ahead (proposed by Oleg Drokin aka green) 35 * - track min/max extents in each group for better group selection 36 * - mb_mark_used() may allocate chunk right after splitting buddy 37 * - tree of groups sorted by number of free blocks 38 * - error handling 39 */ 40 41 /* 42 * The allocation request involve request for multiple number of blocks 43 * near to the goal(block) value specified. 44 * 45 * During initialization phase of the allocator we decide to use the 46 * group preallocation or inode preallocation depending on the size of 47 * the file. The size of the file could be the resulting file size we 48 * would have after allocation, or the current file size, which ever 49 * is larger. If the size is less than sbi->s_mb_stream_request we 50 * select to use the group preallocation. The default value of 51 * s_mb_stream_request is 16 blocks. This can also be tuned via 52 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 53 * terms of number of blocks. 54 * 55 * The main motivation for having small file use group preallocation is to 56 * ensure that we have small files closer together on the disk. 57 * 58 * First stage the allocator looks at the inode prealloc list, 59 * ext4_inode_info->i_prealloc_list, which contains list of prealloc 60 * spaces for this particular inode. The inode prealloc space is 61 * represented as: 62 * 63 * pa_lstart -> the logical start block for this prealloc space 64 * pa_pstart -> the physical start block for this prealloc space 65 * pa_len -> length for this prealloc space (in clusters) 66 * pa_free -> free space available in this prealloc space (in clusters) 67 * 68 * The inode preallocation space is used looking at the _logical_ start 69 * block. If only the logical file block falls within the range of prealloc 70 * space we will consume the particular prealloc space. This makes sure that 71 * we have contiguous physical blocks representing the file blocks 72 * 73 * The important thing to be noted in case of inode prealloc space is that 74 * we don't modify the values associated to inode prealloc space except 75 * pa_free. 76 * 77 * If we are not able to find blocks in the inode prealloc space and if we 78 * have the group allocation flag set then we look at the locality group 79 * prealloc space. These are per CPU prealloc list represented as 80 * 81 * ext4_sb_info.s_locality_groups[smp_processor_id()] 82 * 83 * The reason for having a per cpu locality group is to reduce the contention 84 * between CPUs. It is possible to get scheduled at this point. 85 * 86 * The locality group prealloc space is used looking at whether we have 87 * enough free space (pa_free) within the prealloc space. 88 * 89 * If we can't allocate blocks via inode prealloc or/and locality group 90 * prealloc then we look at the buddy cache. The buddy cache is represented 91 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 92 * mapped to the buddy and bitmap information regarding different 93 * groups. The buddy information is attached to buddy cache inode so that 94 * we can access them through the page cache. The information regarding 95 * each group is loaded via ext4_mb_load_buddy. The information involve 96 * block bitmap and buddy information. The information are stored in the 97 * inode as: 98 * 99 * { page } 100 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 101 * 102 * 103 * one block each for bitmap and buddy information. So for each group we 104 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / 105 * blocksize) blocks. So it can have information regarding groups_per_page 106 * which is blocks_per_page/2 107 * 108 * The buddy cache inode is not stored on disk. The inode is thrown 109 * away when the filesystem is unmounted. 110 * 111 * We look for count number of blocks in the buddy cache. If we were able 112 * to locate that many free blocks we return with additional information 113 * regarding rest of the contiguous physical block available 114 * 115 * Before allocating blocks via buddy cache we normalize the request 116 * blocks. This ensure we ask for more blocks that we needed. The extra 117 * blocks that we get after allocation is added to the respective prealloc 118 * list. In case of inode preallocation we follow a list of heuristics 119 * based on file size. This can be found in ext4_mb_normalize_request. If 120 * we are doing a group prealloc we try to normalize the request to 121 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 122 * dependent on the cluster size; for non-bigalloc file systems, it is 123 * 512 blocks. This can be tuned via 124 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 125 * terms of number of blocks. If we have mounted the file system with -O 126 * stripe=<value> option the group prealloc request is normalized to the 127 * smallest multiple of the stripe value (sbi->s_stripe) which is 128 * greater than the default mb_group_prealloc. 129 * 130 * If "mb_optimize_scan" mount option is set, we maintain in memory group info 131 * structures in two data structures: 132 * 133 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders) 134 * 135 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks) 136 * 137 * This is an array of lists where the index in the array represents the 138 * largest free order in the buddy bitmap of the participating group infos of 139 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total 140 * number of buddy bitmap orders possible) number of lists. Group-infos are 141 * placed in appropriate lists. 142 * 143 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size) 144 * 145 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks) 146 * 147 * This is an array of lists where in the i-th list there are groups with 148 * average fragment size >= 2^i and < 2^(i+1). The average fragment size 149 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments. 150 * Note that we don't bother with a special list for completely empty groups 151 * so we only have MB_NUM_ORDERS(sb) lists. 152 * 153 * When "mb_optimize_scan" mount option is set, mballoc consults the above data 154 * structures to decide the order in which groups are to be traversed for 155 * fulfilling an allocation request. 156 * 157 * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order 158 * >= the order of the request. We directly look at the largest free order list 159 * in the data structure (1) above where largest_free_order = order of the 160 * request. If that list is empty, we look at remaining list in the increasing 161 * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED 162 * lookup in O(1) time. 163 * 164 * At CR_GOAL_LEN_FAST, we only consider groups where 165 * average fragment size > request size. So, we lookup a group which has average 166 * fragment size just above or equal to request size using our average fragment 167 * size group lists (data structure 2) in O(1) time. 168 * 169 * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied 170 * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in 171 * CR_GOAL_LEN_FAST suggests that there is no BG that has avg 172 * fragment size > goal length. So before falling to the slower 173 * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and 174 * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big 175 * enough average fragment size. This increases the chances of finding a 176 * suitable block group in O(1) time and results in faster allocation at the 177 * cost of reduced size of allocation. 178 * 179 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in 180 * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and 181 * CR_GOAL_LEN_FAST phase. 182 * 183 * The regular allocator (using the buddy cache) supports a few tunables. 184 * 185 * /sys/fs/ext4/<partition>/mb_min_to_scan 186 * /sys/fs/ext4/<partition>/mb_max_to_scan 187 * /sys/fs/ext4/<partition>/mb_order2_req 188 * /sys/fs/ext4/<partition>/mb_linear_limit 189 * 190 * The regular allocator uses buddy scan only if the request len is power of 191 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 192 * value of s_mb_order2_reqs can be tuned via 193 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 194 * stripe size (sbi->s_stripe), we try to search for contiguous block in 195 * stripe size. This should result in better allocation on RAID setups. If 196 * not, we search in the specific group using bitmap for best extents. The 197 * tunable min_to_scan and max_to_scan control the behaviour here. 198 * min_to_scan indicate how long the mballoc __must__ look for a best 199 * extent and max_to_scan indicates how long the mballoc __can__ look for a 200 * best extent in the found extents. Searching for the blocks starts with 201 * the group specified as the goal value in allocation context via 202 * ac_g_ex. Each group is first checked based on the criteria whether it 203 * can be used for allocation. ext4_mb_good_group explains how the groups are 204 * checked. 205 * 206 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not 207 * get traversed linearly. That may result in subsequent allocations being not 208 * close to each other. And so, the underlying device may get filled up in a 209 * non-linear fashion. While that may not matter on non-rotational devices, for 210 * rotational devices that may result in higher seek times. "mb_linear_limit" 211 * tells mballoc how many groups mballoc should search linearly before 212 * performing consulting above data structures for more efficient lookups. For 213 * non rotational devices, this value defaults to 0 and for rotational devices 214 * this is set to MB_DEFAULT_LINEAR_LIMIT. 215 * 216 * Both the prealloc space are getting populated as above. So for the first 217 * request we will hit the buddy cache which will result in this prealloc 218 * space getting filled. The prealloc space is then later used for the 219 * subsequent request. 220 */ 221 222 /* 223 * mballoc operates on the following data: 224 * - on-disk bitmap 225 * - in-core buddy (actually includes buddy and bitmap) 226 * - preallocation descriptors (PAs) 227 * 228 * there are two types of preallocations: 229 * - inode 230 * assiged to specific inode and can be used for this inode only. 231 * it describes part of inode's space preallocated to specific 232 * physical blocks. any block from that preallocated can be used 233 * independent. the descriptor just tracks number of blocks left 234 * unused. so, before taking some block from descriptor, one must 235 * make sure corresponded logical block isn't allocated yet. this 236 * also means that freeing any block within descriptor's range 237 * must discard all preallocated blocks. 238 * - locality group 239 * assigned to specific locality group which does not translate to 240 * permanent set of inodes: inode can join and leave group. space 241 * from this type of preallocation can be used for any inode. thus 242 * it's consumed from the beginning to the end. 243 * 244 * relation between them can be expressed as: 245 * in-core buddy = on-disk bitmap + preallocation descriptors 246 * 247 * this mean blocks mballoc considers used are: 248 * - allocated blocks (persistent) 249 * - preallocated blocks (non-persistent) 250 * 251 * consistency in mballoc world means that at any time a block is either 252 * free or used in ALL structures. notice: "any time" should not be read 253 * literally -- time is discrete and delimited by locks. 254 * 255 * to keep it simple, we don't use block numbers, instead we count number of 256 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 257 * 258 * all operations can be expressed as: 259 * - init buddy: buddy = on-disk + PAs 260 * - new PA: buddy += N; PA = N 261 * - use inode PA: on-disk += N; PA -= N 262 * - discard inode PA buddy -= on-disk - PA; PA = 0 263 * - use locality group PA on-disk += N; PA -= N 264 * - discard locality group PA buddy -= PA; PA = 0 265 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 266 * is used in real operation because we can't know actual used 267 * bits from PA, only from on-disk bitmap 268 * 269 * if we follow this strict logic, then all operations above should be atomic. 270 * given some of them can block, we'd have to use something like semaphores 271 * killing performance on high-end SMP hardware. let's try to relax it using 272 * the following knowledge: 273 * 1) if buddy is referenced, it's already initialized 274 * 2) while block is used in buddy and the buddy is referenced, 275 * nobody can re-allocate that block 276 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 277 * bit set and PA claims same block, it's OK. IOW, one can set bit in 278 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 279 * block 280 * 281 * so, now we're building a concurrency table: 282 * - init buddy vs. 283 * - new PA 284 * blocks for PA are allocated in the buddy, buddy must be referenced 285 * until PA is linked to allocation group to avoid concurrent buddy init 286 * - use inode PA 287 * we need to make sure that either on-disk bitmap or PA has uptodate data 288 * given (3) we care that PA-=N operation doesn't interfere with init 289 * - discard inode PA 290 * the simplest way would be to have buddy initialized by the discard 291 * - use locality group PA 292 * again PA-=N must be serialized with init 293 * - discard locality group PA 294 * the simplest way would be to have buddy initialized by the discard 295 * - new PA vs. 296 * - use inode PA 297 * i_data_sem serializes them 298 * - discard inode PA 299 * discard process must wait until PA isn't used by another process 300 * - use locality group PA 301 * some mutex should serialize them 302 * - discard locality group PA 303 * discard process must wait until PA isn't used by another process 304 * - use inode PA 305 * - use inode PA 306 * i_data_sem or another mutex should serializes them 307 * - discard inode PA 308 * discard process must wait until PA isn't used by another process 309 * - use locality group PA 310 * nothing wrong here -- they're different PAs covering different blocks 311 * - discard locality group PA 312 * discard process must wait until PA isn't used by another process 313 * 314 * now we're ready to make few consequences: 315 * - PA is referenced and while it is no discard is possible 316 * - PA is referenced until block isn't marked in on-disk bitmap 317 * - PA changes only after on-disk bitmap 318 * - discard must not compete with init. either init is done before 319 * any discard or they're serialized somehow 320 * - buddy init as sum of on-disk bitmap and PAs is done atomically 321 * 322 * a special case when we've used PA to emptiness. no need to modify buddy 323 * in this case, but we should care about concurrent init 324 * 325 */ 326 327 /* 328 * Logic in few words: 329 * 330 * - allocation: 331 * load group 332 * find blocks 333 * mark bits in on-disk bitmap 334 * release group 335 * 336 * - use preallocation: 337 * find proper PA (per-inode or group) 338 * load group 339 * mark bits in on-disk bitmap 340 * release group 341 * release PA 342 * 343 * - free: 344 * load group 345 * mark bits in on-disk bitmap 346 * release group 347 * 348 * - discard preallocations in group: 349 * mark PAs deleted 350 * move them onto local list 351 * load on-disk bitmap 352 * load group 353 * remove PA from object (inode or locality group) 354 * mark free blocks in-core 355 * 356 * - discard inode's preallocations: 357 */ 358 359 /* 360 * Locking rules 361 * 362 * Locks: 363 * - bitlock on a group (group) 364 * - object (inode/locality) (object) 365 * - per-pa lock (pa) 366 * - cr_power2_aligned lists lock (cr_power2_aligned) 367 * - cr_goal_len_fast lists lock (cr_goal_len_fast) 368 * 369 * Paths: 370 * - new pa 371 * object 372 * group 373 * 374 * - find and use pa: 375 * pa 376 * 377 * - release consumed pa: 378 * pa 379 * group 380 * object 381 * 382 * - generate in-core bitmap: 383 * group 384 * pa 385 * 386 * - discard all for given object (inode, locality group): 387 * object 388 * pa 389 * group 390 * 391 * - discard all for given group: 392 * group 393 * pa 394 * group 395 * object 396 * 397 * - allocation path (ext4_mb_regular_allocator) 398 * group 399 * cr_power2_aligned/cr_goal_len_fast 400 */ 401 static struct kmem_cache *ext4_pspace_cachep; 402 static struct kmem_cache *ext4_ac_cachep; 403 static struct kmem_cache *ext4_free_data_cachep; 404 405 /* We create slab caches for groupinfo data structures based on the 406 * superblock block size. There will be one per mounted filesystem for 407 * each unique s_blocksize_bits */ 408 #define NR_GRPINFO_CACHES 8 409 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 410 411 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 412 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 413 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 414 "ext4_groupinfo_64k", "ext4_groupinfo_128k" 415 }; 416 417 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 418 ext4_group_t group); 419 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 420 ext4_group_t group); 421 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); 422 423 static bool ext4_mb_good_group(struct ext4_allocation_context *ac, 424 ext4_group_t group, enum criteria cr); 425 426 static int ext4_try_to_trim_range(struct super_block *sb, 427 struct ext4_buddy *e4b, ext4_grpblk_t start, 428 ext4_grpblk_t max, ext4_grpblk_t minblocks); 429 430 /* 431 * The algorithm using this percpu seq counter goes below: 432 * 1. We sample the percpu discard_pa_seq counter before trying for block 433 * allocation in ext4_mb_new_blocks(). 434 * 2. We increment this percpu discard_pa_seq counter when we either allocate 435 * or free these blocks i.e. while marking those blocks as used/free in 436 * mb_mark_used()/mb_free_blocks(). 437 * 3. We also increment this percpu seq counter when we successfully identify 438 * that the bb_prealloc_list is not empty and hence proceed for discarding 439 * of those PAs inside ext4_mb_discard_group_preallocations(). 440 * 441 * Now to make sure that the regular fast path of block allocation is not 442 * affected, as a small optimization we only sample the percpu seq counter 443 * on that cpu. Only when the block allocation fails and when freed blocks 444 * found were 0, that is when we sample percpu seq counter for all cpus using 445 * below function ext4_get_discard_pa_seq_sum(). This happens after making 446 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. 447 */ 448 static DEFINE_PER_CPU(u64, discard_pa_seq); 449 static inline u64 ext4_get_discard_pa_seq_sum(void) 450 { 451 int __cpu; 452 u64 __seq = 0; 453 454 for_each_possible_cpu(__cpu) 455 __seq += per_cpu(discard_pa_seq, __cpu); 456 return __seq; 457 } 458 459 static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 460 { 461 #if BITS_PER_LONG == 64 462 *bit += ((unsigned long) addr & 7UL) << 3; 463 addr = (void *) ((unsigned long) addr & ~7UL); 464 #elif BITS_PER_LONG == 32 465 *bit += ((unsigned long) addr & 3UL) << 3; 466 addr = (void *) ((unsigned long) addr & ~3UL); 467 #else 468 #error "how many bits you are?!" 469 #endif 470 return addr; 471 } 472 473 static inline int mb_test_bit(int bit, void *addr) 474 { 475 /* 476 * ext4_test_bit on architecture like powerpc 477 * needs unsigned long aligned address 478 */ 479 addr = mb_correct_addr_and_bit(&bit, addr); 480 return ext4_test_bit(bit, addr); 481 } 482 483 static inline void mb_set_bit(int bit, void *addr) 484 { 485 addr = mb_correct_addr_and_bit(&bit, addr); 486 ext4_set_bit(bit, addr); 487 } 488 489 static inline void mb_clear_bit(int bit, void *addr) 490 { 491 addr = mb_correct_addr_and_bit(&bit, addr); 492 ext4_clear_bit(bit, addr); 493 } 494 495 static inline int mb_test_and_clear_bit(int bit, void *addr) 496 { 497 addr = mb_correct_addr_and_bit(&bit, addr); 498 return ext4_test_and_clear_bit(bit, addr); 499 } 500 501 static inline int mb_find_next_zero_bit(void *addr, int max, int start) 502 { 503 int fix = 0, ret, tmpmax; 504 addr = mb_correct_addr_and_bit(&fix, addr); 505 tmpmax = max + fix; 506 start += fix; 507 508 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 509 if (ret > max) 510 return max; 511 return ret; 512 } 513 514 static inline int mb_find_next_bit(void *addr, int max, int start) 515 { 516 int fix = 0, ret, tmpmax; 517 addr = mb_correct_addr_and_bit(&fix, addr); 518 tmpmax = max + fix; 519 start += fix; 520 521 ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 522 if (ret > max) 523 return max; 524 return ret; 525 } 526 527 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 528 { 529 char *bb; 530 531 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 532 BUG_ON(max == NULL); 533 534 if (order > e4b->bd_blkbits + 1) { 535 *max = 0; 536 return NULL; 537 } 538 539 /* at order 0 we see each particular block */ 540 if (order == 0) { 541 *max = 1 << (e4b->bd_blkbits + 3); 542 return e4b->bd_bitmap; 543 } 544 545 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 546 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 547 548 return bb; 549 } 550 551 #ifdef DOUBLE_CHECK 552 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 553 int first, int count) 554 { 555 int i; 556 struct super_block *sb = e4b->bd_sb; 557 558 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 559 return; 560 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 561 for (i = 0; i < count; i++) { 562 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 563 ext4_fsblk_t blocknr; 564 565 blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 566 blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 567 ext4_grp_locked_error(sb, e4b->bd_group, 568 inode ? inode->i_ino : 0, 569 blocknr, 570 "freeing block already freed " 571 "(bit %u)", 572 first + i); 573 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 574 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 575 } 576 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 577 } 578 } 579 580 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 581 { 582 int i; 583 584 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 585 return; 586 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 587 for (i = 0; i < count; i++) { 588 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 589 mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 590 } 591 } 592 593 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 594 { 595 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 596 return; 597 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 598 unsigned char *b1, *b2; 599 int i; 600 b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 601 b2 = (unsigned char *) bitmap; 602 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 603 if (b1[i] != b2[i]) { 604 ext4_msg(e4b->bd_sb, KERN_ERR, 605 "corruption in group %u " 606 "at byte %u(%u): %x in copy != %x " 607 "on disk/prealloc", 608 e4b->bd_group, i, i * 8, b1[i], b2[i]); 609 BUG(); 610 } 611 } 612 } 613 } 614 615 static void mb_group_bb_bitmap_alloc(struct super_block *sb, 616 struct ext4_group_info *grp, ext4_group_t group) 617 { 618 struct buffer_head *bh; 619 620 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); 621 if (!grp->bb_bitmap) 622 return; 623 624 bh = ext4_read_block_bitmap(sb, group); 625 if (IS_ERR_OR_NULL(bh)) { 626 kfree(grp->bb_bitmap); 627 grp->bb_bitmap = NULL; 628 return; 629 } 630 631 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); 632 put_bh(bh); 633 } 634 635 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) 636 { 637 kfree(grp->bb_bitmap); 638 } 639 640 #else 641 static inline void mb_free_blocks_double(struct inode *inode, 642 struct ext4_buddy *e4b, int first, int count) 643 { 644 return; 645 } 646 static inline void mb_mark_used_double(struct ext4_buddy *e4b, 647 int first, int count) 648 { 649 return; 650 } 651 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 652 { 653 return; 654 } 655 656 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, 657 struct ext4_group_info *grp, ext4_group_t group) 658 { 659 return; 660 } 661 662 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) 663 { 664 return; 665 } 666 #endif 667 668 #ifdef AGGRESSIVE_CHECK 669 670 #define MB_CHECK_ASSERT(assert) \ 671 do { \ 672 if (!(assert)) { \ 673 printk(KERN_EMERG \ 674 "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 675 function, file, line, # assert); \ 676 BUG(); \ 677 } \ 678 } while (0) 679 680 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 681 const char *function, int line) 682 { 683 struct super_block *sb = e4b->bd_sb; 684 int order = e4b->bd_blkbits + 1; 685 int max; 686 int max2; 687 int i; 688 int j; 689 int k; 690 int count; 691 struct ext4_group_info *grp; 692 int fragments = 0; 693 int fstart; 694 struct list_head *cur; 695 void *buddy; 696 void *buddy2; 697 698 if (e4b->bd_info->bb_check_counter++ % 10) 699 return 0; 700 701 while (order > 1) { 702 buddy = mb_find_buddy(e4b, order, &max); 703 MB_CHECK_ASSERT(buddy); 704 buddy2 = mb_find_buddy(e4b, order - 1, &max2); 705 MB_CHECK_ASSERT(buddy2); 706 MB_CHECK_ASSERT(buddy != buddy2); 707 MB_CHECK_ASSERT(max * 2 == max2); 708 709 count = 0; 710 for (i = 0; i < max; i++) { 711 712 if (mb_test_bit(i, buddy)) { 713 /* only single bit in buddy2 may be 0 */ 714 if (!mb_test_bit(i << 1, buddy2)) { 715 MB_CHECK_ASSERT( 716 mb_test_bit((i<<1)+1, buddy2)); 717 } 718 continue; 719 } 720 721 /* both bits in buddy2 must be 1 */ 722 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 723 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 724 725 for (j = 0; j < (1 << order); j++) { 726 k = (i * (1 << order)) + j; 727 MB_CHECK_ASSERT( 728 !mb_test_bit(k, e4b->bd_bitmap)); 729 } 730 count++; 731 } 732 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 733 order--; 734 } 735 736 fstart = -1; 737 buddy = mb_find_buddy(e4b, 0, &max); 738 for (i = 0; i < max; i++) { 739 if (!mb_test_bit(i, buddy)) { 740 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 741 if (fstart == -1) { 742 fragments++; 743 fstart = i; 744 } 745 continue; 746 } 747 fstart = -1; 748 /* check used bits only */ 749 for (j = 0; j < e4b->bd_blkbits + 1; j++) { 750 buddy2 = mb_find_buddy(e4b, j, &max2); 751 k = i >> j; 752 MB_CHECK_ASSERT(k < max2); 753 MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 754 } 755 } 756 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 757 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 758 759 grp = ext4_get_group_info(sb, e4b->bd_group); 760 if (!grp) 761 return NULL; 762 list_for_each(cur, &grp->bb_prealloc_list) { 763 ext4_group_t groupnr; 764 struct ext4_prealloc_space *pa; 765 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 766 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 767 MB_CHECK_ASSERT(groupnr == e4b->bd_group); 768 for (i = 0; i < pa->pa_len; i++) 769 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 770 } 771 return 0; 772 } 773 #undef MB_CHECK_ASSERT 774 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 775 __FILE__, __func__, __LINE__) 776 #else 777 #define mb_check_buddy(e4b) 778 #endif 779 780 /* 781 * Divide blocks started from @first with length @len into 782 * smaller chunks with power of 2 blocks. 783 * Clear the bits in bitmap which the blocks of the chunk(s) covered, 784 * then increase bb_counters[] for corresponded chunk size. 785 */ 786 static void ext4_mb_mark_free_simple(struct super_block *sb, 787 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 788 struct ext4_group_info *grp) 789 { 790 struct ext4_sb_info *sbi = EXT4_SB(sb); 791 ext4_grpblk_t min; 792 ext4_grpblk_t max; 793 ext4_grpblk_t chunk; 794 unsigned int border; 795 796 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 797 798 border = 2 << sb->s_blocksize_bits; 799 800 while (len > 0) { 801 /* find how many blocks can be covered since this position */ 802 max = ffs(first | border) - 1; 803 804 /* find how many blocks of power 2 we need to mark */ 805 min = fls(len) - 1; 806 807 if (max < min) 808 min = max; 809 chunk = 1 << min; 810 811 /* mark multiblock chunks only */ 812 grp->bb_counters[min]++; 813 if (min > 0) 814 mb_clear_bit(first >> min, 815 buddy + sbi->s_mb_offsets[min]); 816 817 len -= chunk; 818 first += chunk; 819 } 820 } 821 822 static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) 823 { 824 int order; 825 826 /* 827 * We don't bother with a special lists groups with only 1 block free 828 * extents and for completely empty groups. 829 */ 830 order = fls(len) - 2; 831 if (order < 0) 832 return 0; 833 if (order == MB_NUM_ORDERS(sb)) 834 order--; 835 return order; 836 } 837 838 /* Move group to appropriate avg_fragment_size list */ 839 static void 840 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) 841 { 842 struct ext4_sb_info *sbi = EXT4_SB(sb); 843 int new_order; 844 845 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) 846 return; 847 848 new_order = mb_avg_fragment_size_order(sb, 849 grp->bb_free / grp->bb_fragments); 850 if (new_order == grp->bb_avg_fragment_size_order) 851 return; 852 853 if (grp->bb_avg_fragment_size_order != -1) { 854 write_lock(&sbi->s_mb_avg_fragment_size_locks[ 855 grp->bb_avg_fragment_size_order]); 856 list_del(&grp->bb_avg_fragment_size_node); 857 write_unlock(&sbi->s_mb_avg_fragment_size_locks[ 858 grp->bb_avg_fragment_size_order]); 859 } 860 grp->bb_avg_fragment_size_order = new_order; 861 write_lock(&sbi->s_mb_avg_fragment_size_locks[ 862 grp->bb_avg_fragment_size_order]); 863 list_add_tail(&grp->bb_avg_fragment_size_node, 864 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]); 865 write_unlock(&sbi->s_mb_avg_fragment_size_locks[ 866 grp->bb_avg_fragment_size_order]); 867 } 868 869 /* 870 * Choose next group by traversing largest_free_order lists. Updates *new_cr if 871 * cr level needs an update. 872 */ 873 static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac, 874 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 875 { 876 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 877 struct ext4_group_info *iter; 878 int i; 879 880 if (ac->ac_status == AC_STATUS_FOUND) 881 return; 882 883 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED)) 884 atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions); 885 886 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { 887 if (list_empty(&sbi->s_mb_largest_free_orders[i])) 888 continue; 889 read_lock(&sbi->s_mb_largest_free_orders_locks[i]); 890 if (list_empty(&sbi->s_mb_largest_free_orders[i])) { 891 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); 892 continue; 893 } 894 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], 895 bb_largest_free_order_node) { 896 if (sbi->s_mb_stats) 897 atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]); 898 if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) { 899 *group = iter->bb_group; 900 ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED; 901 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); 902 return; 903 } 904 } 905 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); 906 } 907 908 /* Increment cr and search again if no group is found */ 909 *new_cr = CR_GOAL_LEN_FAST; 910 } 911 912 /* 913 * Find a suitable group of given order from the average fragments list. 914 */ 915 static struct ext4_group_info * 916 ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order) 917 { 918 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 919 struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order]; 920 rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order]; 921 struct ext4_group_info *grp = NULL, *iter; 922 enum criteria cr = ac->ac_criteria; 923 924 if (list_empty(frag_list)) 925 return NULL; 926 read_lock(frag_list_lock); 927 if (list_empty(frag_list)) { 928 read_unlock(frag_list_lock); 929 return NULL; 930 } 931 list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) { 932 if (sbi->s_mb_stats) 933 atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]); 934 if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) { 935 grp = iter; 936 break; 937 } 938 } 939 read_unlock(frag_list_lock); 940 return grp; 941 } 942 943 /* 944 * Choose next group by traversing average fragment size list of suitable 945 * order. Updates *new_cr if cr level needs an update. 946 */ 947 static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac, 948 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 949 { 950 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 951 struct ext4_group_info *grp = NULL; 952 int i; 953 954 if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) { 955 if (sbi->s_mb_stats) 956 atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions); 957 } 958 959 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); 960 i < MB_NUM_ORDERS(ac->ac_sb); i++) { 961 grp = ext4_mb_find_good_group_avg_frag_lists(ac, i); 962 if (grp) { 963 *group = grp->bb_group; 964 ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED; 965 return; 966 } 967 } 968 969 /* 970 * CR_BEST_AVAIL_LEN works based on the concept that we have 971 * a larger normalized goal len request which can be trimmed to 972 * a smaller goal len such that it can still satisfy original 973 * request len. However, allocation request for non-regular 974 * files never gets normalized. 975 * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA). 976 */ 977 if (ac->ac_flags & EXT4_MB_HINT_DATA) 978 *new_cr = CR_BEST_AVAIL_LEN; 979 else 980 *new_cr = CR_GOAL_LEN_SLOW; 981 } 982 983 /* 984 * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment 985 * order we have and proactively trim the goal request length to that order to 986 * find a suitable group faster. 987 * 988 * This optimizes allocation speed at the cost of slightly reduced 989 * preallocations. However, we make sure that we don't trim the request too 990 * much and fall to CR_GOAL_LEN_SLOW in that case. 991 */ 992 static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac, 993 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 994 { 995 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 996 struct ext4_group_info *grp = NULL; 997 int i, order, min_order; 998 unsigned long num_stripe_clusters = 0; 999 1000 if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) { 1001 if (sbi->s_mb_stats) 1002 atomic_inc(&sbi->s_bal_best_avail_bad_suggestions); 1003 } 1004 1005 /* 1006 * mb_avg_fragment_size_order() returns order in a way that makes 1007 * retrieving back the length using (1 << order) inaccurate. Hence, use 1008 * fls() instead since we need to know the actual length while modifying 1009 * goal length. 1010 */ 1011 order = fls(ac->ac_g_ex.fe_len) - 1; 1012 min_order = order - sbi->s_mb_best_avail_max_trim_order; 1013 if (min_order < 0) 1014 min_order = 0; 1015 1016 if (sbi->s_stripe > 0) { 1017 /* 1018 * We are assuming that stripe size is always a multiple of 1019 * cluster ratio otherwise __ext4_fill_super exists early. 1020 */ 1021 num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe); 1022 if (1 << min_order < num_stripe_clusters) 1023 /* 1024 * We consider 1 order less because later we round 1025 * up the goal len to num_stripe_clusters 1026 */ 1027 min_order = fls(num_stripe_clusters) - 1; 1028 } 1029 1030 if (1 << min_order < ac->ac_o_ex.fe_len) 1031 min_order = fls(ac->ac_o_ex.fe_len); 1032 1033 for (i = order; i >= min_order; i--) { 1034 int frag_order; 1035 /* 1036 * Scale down goal len to make sure we find something 1037 * in the free fragments list. Basically, reduce 1038 * preallocations. 1039 */ 1040 ac->ac_g_ex.fe_len = 1 << i; 1041 1042 if (num_stripe_clusters > 0) { 1043 /* 1044 * Try to round up the adjusted goal length to 1045 * stripe size (in cluster units) multiple for 1046 * efficiency. 1047 */ 1048 ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len, 1049 num_stripe_clusters); 1050 } 1051 1052 frag_order = mb_avg_fragment_size_order(ac->ac_sb, 1053 ac->ac_g_ex.fe_len); 1054 1055 grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order); 1056 if (grp) { 1057 *group = grp->bb_group; 1058 ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED; 1059 return; 1060 } 1061 } 1062 1063 /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */ 1064 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; 1065 *new_cr = CR_GOAL_LEN_SLOW; 1066 } 1067 1068 static inline int should_optimize_scan(struct ext4_allocation_context *ac) 1069 { 1070 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) 1071 return 0; 1072 if (ac->ac_criteria >= CR_GOAL_LEN_SLOW) 1073 return 0; 1074 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) 1075 return 0; 1076 return 1; 1077 } 1078 1079 /* 1080 * Return next linear group for allocation. If linear traversal should not be 1081 * performed, this function just returns the same group 1082 */ 1083 static ext4_group_t 1084 next_linear_group(struct ext4_allocation_context *ac, ext4_group_t group, 1085 ext4_group_t ngroups) 1086 { 1087 if (!should_optimize_scan(ac)) 1088 goto inc_and_return; 1089 1090 if (ac->ac_groups_linear_remaining) { 1091 ac->ac_groups_linear_remaining--; 1092 goto inc_and_return; 1093 } 1094 1095 return group; 1096 inc_and_return: 1097 /* 1098 * Artificially restricted ngroups for non-extent 1099 * files makes group > ngroups possible on first loop. 1100 */ 1101 return group + 1 >= ngroups ? 0 : group + 1; 1102 } 1103 1104 /* 1105 * ext4_mb_choose_next_group: choose next group for allocation. 1106 * 1107 * @ac Allocation Context 1108 * @new_cr This is an output parameter. If the there is no good group 1109 * available at current CR level, this field is updated to indicate 1110 * the new cr level that should be used. 1111 * @group This is an input / output parameter. As an input it indicates the 1112 * next group that the allocator intends to use for allocation. As 1113 * output, this field indicates the next group that should be used as 1114 * determined by the optimization functions. 1115 * @ngroups Total number of groups 1116 */ 1117 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, 1118 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 1119 { 1120 *new_cr = ac->ac_criteria; 1121 1122 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) { 1123 *group = next_linear_group(ac, *group, ngroups); 1124 return; 1125 } 1126 1127 if (*new_cr == CR_POWER2_ALIGNED) { 1128 ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group, ngroups); 1129 } else if (*new_cr == CR_GOAL_LEN_FAST) { 1130 ext4_mb_choose_next_group_goal_fast(ac, new_cr, group, ngroups); 1131 } else if (*new_cr == CR_BEST_AVAIL_LEN) { 1132 ext4_mb_choose_next_group_best_avail(ac, new_cr, group, ngroups); 1133 } else { 1134 /* 1135 * TODO: For CR=2, we can arrange groups in an rb tree sorted by 1136 * bb_free. But until that happens, we should never come here. 1137 */ 1138 WARN_ON(1); 1139 } 1140 } 1141 1142 /* 1143 * Cache the order of the largest free extent we have available in this block 1144 * group. 1145 */ 1146 static void 1147 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 1148 { 1149 struct ext4_sb_info *sbi = EXT4_SB(sb); 1150 int i; 1151 1152 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) 1153 if (grp->bb_counters[i] > 0) 1154 break; 1155 /* No need to move between order lists? */ 1156 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || 1157 i == grp->bb_largest_free_order) { 1158 grp->bb_largest_free_order = i; 1159 return; 1160 } 1161 1162 if (grp->bb_largest_free_order >= 0) { 1163 write_lock(&sbi->s_mb_largest_free_orders_locks[ 1164 grp->bb_largest_free_order]); 1165 list_del_init(&grp->bb_largest_free_order_node); 1166 write_unlock(&sbi->s_mb_largest_free_orders_locks[ 1167 grp->bb_largest_free_order]); 1168 } 1169 grp->bb_largest_free_order = i; 1170 if (grp->bb_largest_free_order >= 0 && grp->bb_free) { 1171 write_lock(&sbi->s_mb_largest_free_orders_locks[ 1172 grp->bb_largest_free_order]); 1173 list_add_tail(&grp->bb_largest_free_order_node, 1174 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); 1175 write_unlock(&sbi->s_mb_largest_free_orders_locks[ 1176 grp->bb_largest_free_order]); 1177 } 1178 } 1179 1180 static noinline_for_stack 1181 void ext4_mb_generate_buddy(struct super_block *sb, 1182 void *buddy, void *bitmap, ext4_group_t group, 1183 struct ext4_group_info *grp) 1184 { 1185 struct ext4_sb_info *sbi = EXT4_SB(sb); 1186 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 1187 ext4_grpblk_t i = 0; 1188 ext4_grpblk_t first; 1189 ext4_grpblk_t len; 1190 unsigned free = 0; 1191 unsigned fragments = 0; 1192 unsigned long long period = get_cycles(); 1193 1194 /* initialize buddy from bitmap which is aggregation 1195 * of on-disk bitmap and preallocations */ 1196 i = mb_find_next_zero_bit(bitmap, max, 0); 1197 grp->bb_first_free = i; 1198 while (i < max) { 1199 fragments++; 1200 first = i; 1201 i = mb_find_next_bit(bitmap, max, i); 1202 len = i - first; 1203 free += len; 1204 if (len > 1) 1205 ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 1206 else 1207 grp->bb_counters[0]++; 1208 if (i < max) 1209 i = mb_find_next_zero_bit(bitmap, max, i); 1210 } 1211 grp->bb_fragments = fragments; 1212 1213 if (free != grp->bb_free) { 1214 ext4_grp_locked_error(sb, group, 0, 0, 1215 "block bitmap and bg descriptor " 1216 "inconsistent: %u vs %u free clusters", 1217 free, grp->bb_free); 1218 /* 1219 * If we intend to continue, we consider group descriptor 1220 * corrupt and update bb_free using bitmap value 1221 */ 1222 grp->bb_free = free; 1223 ext4_mark_group_bitmap_corrupted(sb, group, 1224 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1225 } 1226 mb_set_largest_free_order(sb, grp); 1227 mb_update_avg_fragment_size(sb, grp); 1228 1229 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 1230 1231 period = get_cycles() - period; 1232 atomic_inc(&sbi->s_mb_buddies_generated); 1233 atomic64_add(period, &sbi->s_mb_generation_time); 1234 } 1235 1236 /* The buddy information is attached the buddy cache inode 1237 * for convenience. The information regarding each group 1238 * is loaded via ext4_mb_load_buddy. The information involve 1239 * block bitmap and buddy information. The information are 1240 * stored in the inode as 1241 * 1242 * { page } 1243 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 1244 * 1245 * 1246 * one block each for bitmap and buddy information. 1247 * So for each group we take up 2 blocks. A page can 1248 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. 1249 * So it can have information regarding groups_per_page which 1250 * is blocks_per_page/2 1251 * 1252 * Locking note: This routine takes the block group lock of all groups 1253 * for this page; do not hold this lock when calling this routine! 1254 */ 1255 1256 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) 1257 { 1258 ext4_group_t ngroups; 1259 unsigned int blocksize; 1260 int blocks_per_page; 1261 int groups_per_page; 1262 int err = 0; 1263 int i; 1264 ext4_group_t first_group, group; 1265 int first_block; 1266 struct super_block *sb; 1267 struct buffer_head *bhs; 1268 struct buffer_head **bh = NULL; 1269 struct inode *inode; 1270 char *data; 1271 char *bitmap; 1272 struct ext4_group_info *grinfo; 1273 1274 inode = page->mapping->host; 1275 sb = inode->i_sb; 1276 ngroups = ext4_get_groups_count(sb); 1277 blocksize = i_blocksize(inode); 1278 blocks_per_page = PAGE_SIZE / blocksize; 1279 1280 mb_debug(sb, "init page %lu\n", page->index); 1281 1282 groups_per_page = blocks_per_page >> 1; 1283 if (groups_per_page == 0) 1284 groups_per_page = 1; 1285 1286 /* allocate buffer_heads to read bitmaps */ 1287 if (groups_per_page > 1) { 1288 i = sizeof(struct buffer_head *) * groups_per_page; 1289 bh = kzalloc(i, gfp); 1290 if (bh == NULL) 1291 return -ENOMEM; 1292 } else 1293 bh = &bhs; 1294 1295 first_group = page->index * blocks_per_page / 2; 1296 1297 /* read all groups the page covers into the cache */ 1298 for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 1299 if (group >= ngroups) 1300 break; 1301 1302 grinfo = ext4_get_group_info(sb, group); 1303 if (!grinfo) 1304 continue; 1305 /* 1306 * If page is uptodate then we came here after online resize 1307 * which added some new uninitialized group info structs, so 1308 * we must skip all initialized uptodate buddies on the page, 1309 * which may be currently in use by an allocating task. 1310 */ 1311 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { 1312 bh[i] = NULL; 1313 continue; 1314 } 1315 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); 1316 if (IS_ERR(bh[i])) { 1317 err = PTR_ERR(bh[i]); 1318 bh[i] = NULL; 1319 goto out; 1320 } 1321 mb_debug(sb, "read bitmap for group %u\n", group); 1322 } 1323 1324 /* wait for I/O completion */ 1325 for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 1326 int err2; 1327 1328 if (!bh[i]) 1329 continue; 1330 err2 = ext4_wait_block_bitmap(sb, group, bh[i]); 1331 if (!err) 1332 err = err2; 1333 } 1334 1335 first_block = page->index * blocks_per_page; 1336 for (i = 0; i < blocks_per_page; i++) { 1337 group = (first_block + i) >> 1; 1338 if (group >= ngroups) 1339 break; 1340 1341 if (!bh[group - first_group]) 1342 /* skip initialized uptodate buddy */ 1343 continue; 1344 1345 if (!buffer_verified(bh[group - first_group])) 1346 /* Skip faulty bitmaps */ 1347 continue; 1348 err = 0; 1349 1350 /* 1351 * data carry information regarding this 1352 * particular group in the format specified 1353 * above 1354 * 1355 */ 1356 data = page_address(page) + (i * blocksize); 1357 bitmap = bh[group - first_group]->b_data; 1358 1359 /* 1360 * We place the buddy block and bitmap block 1361 * close together 1362 */ 1363 if ((first_block + i) & 1) { 1364 /* this is block of buddy */ 1365 BUG_ON(incore == NULL); 1366 mb_debug(sb, "put buddy for group %u in page %lu/%x\n", 1367 group, page->index, i * blocksize); 1368 trace_ext4_mb_buddy_bitmap_load(sb, group); 1369 grinfo = ext4_get_group_info(sb, group); 1370 if (!grinfo) { 1371 err = -EFSCORRUPTED; 1372 goto out; 1373 } 1374 grinfo->bb_fragments = 0; 1375 memset(grinfo->bb_counters, 0, 1376 sizeof(*grinfo->bb_counters) * 1377 (MB_NUM_ORDERS(sb))); 1378 /* 1379 * incore got set to the group block bitmap below 1380 */ 1381 ext4_lock_group(sb, group); 1382 /* init the buddy */ 1383 memset(data, 0xff, blocksize); 1384 ext4_mb_generate_buddy(sb, data, incore, group, grinfo); 1385 ext4_unlock_group(sb, group); 1386 incore = NULL; 1387 } else { 1388 /* this is block of bitmap */ 1389 BUG_ON(incore != NULL); 1390 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", 1391 group, page->index, i * blocksize); 1392 trace_ext4_mb_bitmap_load(sb, group); 1393 1394 /* see comments in ext4_mb_put_pa() */ 1395 ext4_lock_group(sb, group); 1396 memcpy(data, bitmap, blocksize); 1397 1398 /* mark all preallocated blks used in in-core bitmap */ 1399 ext4_mb_generate_from_pa(sb, data, group); 1400 ext4_mb_generate_from_freelist(sb, data, group); 1401 ext4_unlock_group(sb, group); 1402 1403 /* set incore so that the buddy information can be 1404 * generated using this 1405 */ 1406 incore = data; 1407 } 1408 } 1409 SetPageUptodate(page); 1410 1411 out: 1412 if (bh) { 1413 for (i = 0; i < groups_per_page; i++) 1414 brelse(bh[i]); 1415 if (bh != &bhs) 1416 kfree(bh); 1417 } 1418 return err; 1419 } 1420 1421 /* 1422 * Lock the buddy and bitmap pages. This make sure other parallel init_group 1423 * on the same buddy page doesn't happen whild holding the buddy page lock. 1424 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 1425 * are on the same page e4b->bd_buddy_page is NULL and return value is 0. 1426 */ 1427 static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 1428 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) 1429 { 1430 struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 1431 int block, pnum, poff; 1432 int blocks_per_page; 1433 struct page *page; 1434 1435 e4b->bd_buddy_page = NULL; 1436 e4b->bd_bitmap_page = NULL; 1437 1438 blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1439 /* 1440 * the buddy cache inode stores the block bitmap 1441 * and buddy information in consecutive blocks. 1442 * So for each group we need two blocks. 1443 */ 1444 block = group * 2; 1445 pnum = block / blocks_per_page; 1446 poff = block % blocks_per_page; 1447 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1448 if (!page) 1449 return -ENOMEM; 1450 BUG_ON(page->mapping != inode->i_mapping); 1451 e4b->bd_bitmap_page = page; 1452 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1453 1454 if (blocks_per_page >= 2) { 1455 /* buddy and bitmap are on the same page */ 1456 return 0; 1457 } 1458 1459 block++; 1460 pnum = block / blocks_per_page; 1461 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1462 if (!page) 1463 return -ENOMEM; 1464 BUG_ON(page->mapping != inode->i_mapping); 1465 e4b->bd_buddy_page = page; 1466 return 0; 1467 } 1468 1469 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 1470 { 1471 if (e4b->bd_bitmap_page) { 1472 unlock_page(e4b->bd_bitmap_page); 1473 put_page(e4b->bd_bitmap_page); 1474 } 1475 if (e4b->bd_buddy_page) { 1476 unlock_page(e4b->bd_buddy_page); 1477 put_page(e4b->bd_buddy_page); 1478 } 1479 } 1480 1481 /* 1482 * Locking note: This routine calls ext4_mb_init_cache(), which takes the 1483 * block group lock of all groups for this page; do not hold the BG lock when 1484 * calling this routine! 1485 */ 1486 static noinline_for_stack 1487 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) 1488 { 1489 1490 struct ext4_group_info *this_grp; 1491 struct ext4_buddy e4b; 1492 struct page *page; 1493 int ret = 0; 1494 1495 might_sleep(); 1496 mb_debug(sb, "init group %u\n", group); 1497 this_grp = ext4_get_group_info(sb, group); 1498 if (!this_grp) 1499 return -EFSCORRUPTED; 1500 1501 /* 1502 * This ensures that we don't reinit the buddy cache 1503 * page which map to the group from which we are already 1504 * allocating. If we are looking at the buddy cache we would 1505 * have taken a reference using ext4_mb_load_buddy and that 1506 * would have pinned buddy page to page cache. 1507 * The call to ext4_mb_get_buddy_page_lock will mark the 1508 * page accessed. 1509 */ 1510 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); 1511 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1512 /* 1513 * somebody initialized the group 1514 * return without doing anything 1515 */ 1516 goto err; 1517 } 1518 1519 page = e4b.bd_bitmap_page; 1520 ret = ext4_mb_init_cache(page, NULL, gfp); 1521 if (ret) 1522 goto err; 1523 if (!PageUptodate(page)) { 1524 ret = -EIO; 1525 goto err; 1526 } 1527 1528 if (e4b.bd_buddy_page == NULL) { 1529 /* 1530 * If both the bitmap and buddy are in 1531 * the same page we don't need to force 1532 * init the buddy 1533 */ 1534 ret = 0; 1535 goto err; 1536 } 1537 /* init buddy cache */ 1538 page = e4b.bd_buddy_page; 1539 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); 1540 if (ret) 1541 goto err; 1542 if (!PageUptodate(page)) { 1543 ret = -EIO; 1544 goto err; 1545 } 1546 err: 1547 ext4_mb_put_buddy_page_lock(&e4b); 1548 return ret; 1549 } 1550 1551 /* 1552 * Locking note: This routine calls ext4_mb_init_cache(), which takes the 1553 * block group lock of all groups for this page; do not hold the BG lock when 1554 * calling this routine! 1555 */ 1556 static noinline_for_stack int 1557 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, 1558 struct ext4_buddy *e4b, gfp_t gfp) 1559 { 1560 int blocks_per_page; 1561 int block; 1562 int pnum; 1563 int poff; 1564 struct page *page; 1565 int ret; 1566 struct ext4_group_info *grp; 1567 struct ext4_sb_info *sbi = EXT4_SB(sb); 1568 struct inode *inode = sbi->s_buddy_cache; 1569 1570 might_sleep(); 1571 mb_debug(sb, "load group %u\n", group); 1572 1573 blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1574 grp = ext4_get_group_info(sb, group); 1575 if (!grp) 1576 return -EFSCORRUPTED; 1577 1578 e4b->bd_blkbits = sb->s_blocksize_bits; 1579 e4b->bd_info = grp; 1580 e4b->bd_sb = sb; 1581 e4b->bd_group = group; 1582 e4b->bd_buddy_page = NULL; 1583 e4b->bd_bitmap_page = NULL; 1584 1585 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1586 /* 1587 * we need full data about the group 1588 * to make a good selection 1589 */ 1590 ret = ext4_mb_init_group(sb, group, gfp); 1591 if (ret) 1592 return ret; 1593 } 1594 1595 /* 1596 * the buddy cache inode stores the block bitmap 1597 * and buddy information in consecutive blocks. 1598 * So for each group we need two blocks. 1599 */ 1600 block = group * 2; 1601 pnum = block / blocks_per_page; 1602 poff = block % blocks_per_page; 1603 1604 /* we could use find_or_create_page(), but it locks page 1605 * what we'd like to avoid in fast path ... */ 1606 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1607 if (page == NULL || !PageUptodate(page)) { 1608 if (page) 1609 /* 1610 * drop the page reference and try 1611 * to get the page with lock. If we 1612 * are not uptodate that implies 1613 * somebody just created the page but 1614 * is yet to initialize the same. So 1615 * wait for it to initialize. 1616 */ 1617 put_page(page); 1618 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1619 if (page) { 1620 if (WARN_RATELIMIT(page->mapping != inode->i_mapping, 1621 "ext4: bitmap's paging->mapping != inode->i_mapping\n")) { 1622 /* should never happen */ 1623 unlock_page(page); 1624 ret = -EINVAL; 1625 goto err; 1626 } 1627 if (!PageUptodate(page)) { 1628 ret = ext4_mb_init_cache(page, NULL, gfp); 1629 if (ret) { 1630 unlock_page(page); 1631 goto err; 1632 } 1633 mb_cmp_bitmaps(e4b, page_address(page) + 1634 (poff * sb->s_blocksize)); 1635 } 1636 unlock_page(page); 1637 } 1638 } 1639 if (page == NULL) { 1640 ret = -ENOMEM; 1641 goto err; 1642 } 1643 if (!PageUptodate(page)) { 1644 ret = -EIO; 1645 goto err; 1646 } 1647 1648 /* Pages marked accessed already */ 1649 e4b->bd_bitmap_page = page; 1650 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1651 1652 block++; 1653 pnum = block / blocks_per_page; 1654 poff = block % blocks_per_page; 1655 1656 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1657 if (page == NULL || !PageUptodate(page)) { 1658 if (page) 1659 put_page(page); 1660 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1661 if (page) { 1662 if (WARN_RATELIMIT(page->mapping != inode->i_mapping, 1663 "ext4: buddy bitmap's page->mapping != inode->i_mapping\n")) { 1664 /* should never happen */ 1665 unlock_page(page); 1666 ret = -EINVAL; 1667 goto err; 1668 } 1669 if (!PageUptodate(page)) { 1670 ret = ext4_mb_init_cache(page, e4b->bd_bitmap, 1671 gfp); 1672 if (ret) { 1673 unlock_page(page); 1674 goto err; 1675 } 1676 } 1677 unlock_page(page); 1678 } 1679 } 1680 if (page == NULL) { 1681 ret = -ENOMEM; 1682 goto err; 1683 } 1684 if (!PageUptodate(page)) { 1685 ret = -EIO; 1686 goto err; 1687 } 1688 1689 /* Pages marked accessed already */ 1690 e4b->bd_buddy_page = page; 1691 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1692 1693 return 0; 1694 1695 err: 1696 if (page) 1697 put_page(page); 1698 if (e4b->bd_bitmap_page) 1699 put_page(e4b->bd_bitmap_page); 1700 1701 e4b->bd_buddy = NULL; 1702 e4b->bd_bitmap = NULL; 1703 return ret; 1704 } 1705 1706 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1707 struct ext4_buddy *e4b) 1708 { 1709 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); 1710 } 1711 1712 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1713 { 1714 if (e4b->bd_bitmap_page) 1715 put_page(e4b->bd_bitmap_page); 1716 if (e4b->bd_buddy_page) 1717 put_page(e4b->bd_buddy_page); 1718 } 1719 1720 1721 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1722 { 1723 int order = 1, max; 1724 void *bb; 1725 1726 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1727 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1728 1729 while (order <= e4b->bd_blkbits + 1) { 1730 bb = mb_find_buddy(e4b, order, &max); 1731 if (!mb_test_bit(block >> order, bb)) { 1732 /* this block is part of buddy of order 'order' */ 1733 return order; 1734 } 1735 order++; 1736 } 1737 return 0; 1738 } 1739 1740 static void mb_clear_bits(void *bm, int cur, int len) 1741 { 1742 __u32 *addr; 1743 1744 len = cur + len; 1745 while (cur < len) { 1746 if ((cur & 31) == 0 && (len - cur) >= 32) { 1747 /* fast path: clear whole word at once */ 1748 addr = bm + (cur >> 3); 1749 *addr = 0; 1750 cur += 32; 1751 continue; 1752 } 1753 mb_clear_bit(cur, bm); 1754 cur++; 1755 } 1756 } 1757 1758 /* clear bits in given range 1759 * will return first found zero bit if any, -1 otherwise 1760 */ 1761 static int mb_test_and_clear_bits(void *bm, int cur, int len) 1762 { 1763 __u32 *addr; 1764 int zero_bit = -1; 1765 1766 len = cur + len; 1767 while (cur < len) { 1768 if ((cur & 31) == 0 && (len - cur) >= 32) { 1769 /* fast path: clear whole word at once */ 1770 addr = bm + (cur >> 3); 1771 if (*addr != (__u32)(-1) && zero_bit == -1) 1772 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); 1773 *addr = 0; 1774 cur += 32; 1775 continue; 1776 } 1777 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) 1778 zero_bit = cur; 1779 cur++; 1780 } 1781 1782 return zero_bit; 1783 } 1784 1785 void mb_set_bits(void *bm, int cur, int len) 1786 { 1787 __u32 *addr; 1788 1789 len = cur + len; 1790 while (cur < len) { 1791 if ((cur & 31) == 0 && (len - cur) >= 32) { 1792 /* fast path: set whole word at once */ 1793 addr = bm + (cur >> 3); 1794 *addr = 0xffffffff; 1795 cur += 32; 1796 continue; 1797 } 1798 mb_set_bit(cur, bm); 1799 cur++; 1800 } 1801 } 1802 1803 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) 1804 { 1805 if (mb_test_bit(*bit + side, bitmap)) { 1806 mb_clear_bit(*bit, bitmap); 1807 (*bit) -= side; 1808 return 1; 1809 } 1810 else { 1811 (*bit) += side; 1812 mb_set_bit(*bit, bitmap); 1813 return -1; 1814 } 1815 } 1816 1817 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) 1818 { 1819 int max; 1820 int order = 1; 1821 void *buddy = mb_find_buddy(e4b, order, &max); 1822 1823 while (buddy) { 1824 void *buddy2; 1825 1826 /* Bits in range [first; last] are known to be set since 1827 * corresponding blocks were allocated. Bits in range 1828 * (first; last) will stay set because they form buddies on 1829 * upper layer. We just deal with borders if they don't 1830 * align with upper layer and then go up. 1831 * Releasing entire group is all about clearing 1832 * single bit of highest order buddy. 1833 */ 1834 1835 /* Example: 1836 * --------------------------------- 1837 * | 1 | 1 | 1 | 1 | 1838 * --------------------------------- 1839 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1840 * --------------------------------- 1841 * 0 1 2 3 4 5 6 7 1842 * \_____________________/ 1843 * 1844 * Neither [1] nor [6] is aligned to above layer. 1845 * Left neighbour [0] is free, so mark it busy, 1846 * decrease bb_counters and extend range to 1847 * [0; 6] 1848 * Right neighbour [7] is busy. It can't be coaleasced with [6], so 1849 * mark [6] free, increase bb_counters and shrink range to 1850 * [0; 5]. 1851 * Then shift range to [0; 2], go up and do the same. 1852 */ 1853 1854 1855 if (first & 1) 1856 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); 1857 if (!(last & 1)) 1858 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); 1859 if (first > last) 1860 break; 1861 order++; 1862 1863 buddy2 = mb_find_buddy(e4b, order, &max); 1864 if (!buddy2) { 1865 mb_clear_bits(buddy, first, last - first + 1); 1866 e4b->bd_info->bb_counters[order - 1] += last - first + 1; 1867 break; 1868 } 1869 first >>= 1; 1870 last >>= 1; 1871 buddy = buddy2; 1872 } 1873 } 1874 1875 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1876 int first, int count) 1877 { 1878 int left_is_free = 0; 1879 int right_is_free = 0; 1880 int block; 1881 int last = first + count - 1; 1882 struct super_block *sb = e4b->bd_sb; 1883 1884 if (WARN_ON(count == 0)) 1885 return; 1886 BUG_ON(last >= (sb->s_blocksize << 3)); 1887 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1888 /* Don't bother if the block group is corrupt. */ 1889 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 1890 return; 1891 1892 mb_check_buddy(e4b); 1893 mb_free_blocks_double(inode, e4b, first, count); 1894 1895 this_cpu_inc(discard_pa_seq); 1896 e4b->bd_info->bb_free += count; 1897 if (first < e4b->bd_info->bb_first_free) 1898 e4b->bd_info->bb_first_free = first; 1899 1900 /* access memory sequentially: check left neighbour, 1901 * clear range and then check right neighbour 1902 */ 1903 if (first != 0) 1904 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); 1905 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); 1906 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) 1907 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); 1908 1909 if (unlikely(block != -1)) { 1910 struct ext4_sb_info *sbi = EXT4_SB(sb); 1911 ext4_fsblk_t blocknr; 1912 1913 blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 1914 blocknr += EXT4_C2B(sbi, block); 1915 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { 1916 ext4_grp_locked_error(sb, e4b->bd_group, 1917 inode ? inode->i_ino : 0, 1918 blocknr, 1919 "freeing already freed block (bit %u); block bitmap corrupt.", 1920 block); 1921 ext4_mark_group_bitmap_corrupted( 1922 sb, e4b->bd_group, 1923 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1924 } 1925 goto done; 1926 } 1927 1928 /* let's maintain fragments counter */ 1929 if (left_is_free && right_is_free) 1930 e4b->bd_info->bb_fragments--; 1931 else if (!left_is_free && !right_is_free) 1932 e4b->bd_info->bb_fragments++; 1933 1934 /* buddy[0] == bd_bitmap is a special case, so handle 1935 * it right away and let mb_buddy_mark_free stay free of 1936 * zero order checks. 1937 * Check if neighbours are to be coaleasced, 1938 * adjust bitmap bb_counters and borders appropriately. 1939 */ 1940 if (first & 1) { 1941 first += !left_is_free; 1942 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; 1943 } 1944 if (!(last & 1)) { 1945 last -= !right_is_free; 1946 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; 1947 } 1948 1949 if (first <= last) 1950 mb_buddy_mark_free(e4b, first >> 1, last >> 1); 1951 1952 done: 1953 mb_set_largest_free_order(sb, e4b->bd_info); 1954 mb_update_avg_fragment_size(sb, e4b->bd_info); 1955 mb_check_buddy(e4b); 1956 } 1957 1958 static int mb_find_extent(struct ext4_buddy *e4b, int block, 1959 int needed, struct ext4_free_extent *ex) 1960 { 1961 int next = block; 1962 int max, order; 1963 void *buddy; 1964 1965 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1966 BUG_ON(ex == NULL); 1967 1968 buddy = mb_find_buddy(e4b, 0, &max); 1969 BUG_ON(buddy == NULL); 1970 BUG_ON(block >= max); 1971 if (mb_test_bit(block, buddy)) { 1972 ex->fe_len = 0; 1973 ex->fe_start = 0; 1974 ex->fe_group = 0; 1975 return 0; 1976 } 1977 1978 /* find actual order */ 1979 order = mb_find_order_for_block(e4b, block); 1980 block = block >> order; 1981 1982 ex->fe_len = 1 << order; 1983 ex->fe_start = block << order; 1984 ex->fe_group = e4b->bd_group; 1985 1986 /* calc difference from given start */ 1987 next = next - ex->fe_start; 1988 ex->fe_len -= next; 1989 ex->fe_start += next; 1990 1991 while (needed > ex->fe_len && 1992 mb_find_buddy(e4b, order, &max)) { 1993 1994 if (block + 1 >= max) 1995 break; 1996 1997 next = (block + 1) * (1 << order); 1998 if (mb_test_bit(next, e4b->bd_bitmap)) 1999 break; 2000 2001 order = mb_find_order_for_block(e4b, next); 2002 2003 block = next >> order; 2004 ex->fe_len += 1 << order; 2005 } 2006 2007 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { 2008 /* Should never happen! (but apparently sometimes does?!?) */ 2009 WARN_ON(1); 2010 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, 2011 "corruption or bug in mb_find_extent " 2012 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", 2013 block, order, needed, ex->fe_group, ex->fe_start, 2014 ex->fe_len, ex->fe_logical); 2015 ex->fe_len = 0; 2016 ex->fe_start = 0; 2017 ex->fe_group = 0; 2018 } 2019 return ex->fe_len; 2020 } 2021 2022 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 2023 { 2024 int ord; 2025 int mlen = 0; 2026 int max = 0; 2027 int cur; 2028 int start = ex->fe_start; 2029 int len = ex->fe_len; 2030 unsigned ret = 0; 2031 int len0 = len; 2032 void *buddy; 2033 bool split = false; 2034 2035 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 2036 BUG_ON(e4b->bd_group != ex->fe_group); 2037 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 2038 mb_check_buddy(e4b); 2039 mb_mark_used_double(e4b, start, len); 2040 2041 this_cpu_inc(discard_pa_seq); 2042 e4b->bd_info->bb_free -= len; 2043 if (e4b->bd_info->bb_first_free == start) 2044 e4b->bd_info->bb_first_free += len; 2045 2046 /* let's maintain fragments counter */ 2047 if (start != 0) 2048 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 2049 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 2050 max = !mb_test_bit(start + len, e4b->bd_bitmap); 2051 if (mlen && max) 2052 e4b->bd_info->bb_fragments++; 2053 else if (!mlen && !max) 2054 e4b->bd_info->bb_fragments--; 2055 2056 /* let's maintain buddy itself */ 2057 while (len) { 2058 if (!split) 2059 ord = mb_find_order_for_block(e4b, start); 2060 2061 if (((start >> ord) << ord) == start && len >= (1 << ord)) { 2062 /* the whole chunk may be allocated at once! */ 2063 mlen = 1 << ord; 2064 if (!split) 2065 buddy = mb_find_buddy(e4b, ord, &max); 2066 else 2067 split = false; 2068 BUG_ON((start >> ord) >= max); 2069 mb_set_bit(start >> ord, buddy); 2070 e4b->bd_info->bb_counters[ord]--; 2071 start += mlen; 2072 len -= mlen; 2073 BUG_ON(len < 0); 2074 continue; 2075 } 2076 2077 /* store for history */ 2078 if (ret == 0) 2079 ret = len | (ord << 16); 2080 2081 /* we have to split large buddy */ 2082 BUG_ON(ord <= 0); 2083 buddy = mb_find_buddy(e4b, ord, &max); 2084 mb_set_bit(start >> ord, buddy); 2085 e4b->bd_info->bb_counters[ord]--; 2086 2087 ord--; 2088 cur = (start >> ord) & ~1U; 2089 buddy = mb_find_buddy(e4b, ord, &max); 2090 mb_clear_bit(cur, buddy); 2091 mb_clear_bit(cur + 1, buddy); 2092 e4b->bd_info->bb_counters[ord]++; 2093 e4b->bd_info->bb_counters[ord]++; 2094 split = true; 2095 } 2096 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 2097 2098 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); 2099 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 2100 mb_check_buddy(e4b); 2101 2102 return ret; 2103 } 2104 2105 /* 2106 * Must be called under group lock! 2107 */ 2108 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 2109 struct ext4_buddy *e4b) 2110 { 2111 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2112 int ret; 2113 2114 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 2115 BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2116 2117 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 2118 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 2119 ret = mb_mark_used(e4b, &ac->ac_b_ex); 2120 2121 /* preallocation can change ac_b_ex, thus we store actually 2122 * allocated blocks for history */ 2123 ac->ac_f_ex = ac->ac_b_ex; 2124 2125 ac->ac_status = AC_STATUS_FOUND; 2126 ac->ac_tail = ret & 0xffff; 2127 ac->ac_buddy = ret >> 16; 2128 2129 /* 2130 * take the page reference. We want the page to be pinned 2131 * so that we don't get a ext4_mb_init_cache_call for this 2132 * group until we update the bitmap. That would mean we 2133 * double allocate blocks. The reference is dropped 2134 * in ext4_mb_release_context 2135 */ 2136 ac->ac_bitmap_page = e4b->bd_bitmap_page; 2137 get_page(ac->ac_bitmap_page); 2138 ac->ac_buddy_page = e4b->bd_buddy_page; 2139 get_page(ac->ac_buddy_page); 2140 /* store last allocated for subsequent stream allocation */ 2141 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2142 spin_lock(&sbi->s_md_lock); 2143 sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 2144 sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 2145 spin_unlock(&sbi->s_md_lock); 2146 } 2147 /* 2148 * As we've just preallocated more space than 2149 * user requested originally, we store allocated 2150 * space in a special descriptor. 2151 */ 2152 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 2153 ext4_mb_new_preallocation(ac); 2154 2155 } 2156 2157 static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 2158 struct ext4_buddy *e4b, 2159 int finish_group) 2160 { 2161 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2162 struct ext4_free_extent *bex = &ac->ac_b_ex; 2163 struct ext4_free_extent *gex = &ac->ac_g_ex; 2164 2165 if (ac->ac_status == AC_STATUS_FOUND) 2166 return; 2167 /* 2168 * We don't want to scan for a whole year 2169 */ 2170 if (ac->ac_found > sbi->s_mb_max_to_scan && 2171 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2172 ac->ac_status = AC_STATUS_BREAK; 2173 return; 2174 } 2175 2176 /* 2177 * Haven't found good chunk so far, let's continue 2178 */ 2179 if (bex->fe_len < gex->fe_len) 2180 return; 2181 2182 if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 2183 ext4_mb_use_best_found(ac, e4b); 2184 } 2185 2186 /* 2187 * The routine checks whether found extent is good enough. If it is, 2188 * then the extent gets marked used and flag is set to the context 2189 * to stop scanning. Otherwise, the extent is compared with the 2190 * previous found extent and if new one is better, then it's stored 2191 * in the context. Later, the best found extent will be used, if 2192 * mballoc can't find good enough extent. 2193 * 2194 * The algorithm used is roughly as follows: 2195 * 2196 * * If free extent found is exactly as big as goal, then 2197 * stop the scan and use it immediately 2198 * 2199 * * If free extent found is smaller than goal, then keep retrying 2200 * upto a max of sbi->s_mb_max_to_scan times (default 200). After 2201 * that stop scanning and use whatever we have. 2202 * 2203 * * If free extent found is bigger than goal, then keep retrying 2204 * upto a max of sbi->s_mb_min_to_scan times (default 10) before 2205 * stopping the scan and using the extent. 2206 * 2207 * 2208 * FIXME: real allocation policy is to be designed yet! 2209 */ 2210 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 2211 struct ext4_free_extent *ex, 2212 struct ext4_buddy *e4b) 2213 { 2214 struct ext4_free_extent *bex = &ac->ac_b_ex; 2215 struct ext4_free_extent *gex = &ac->ac_g_ex; 2216 2217 BUG_ON(ex->fe_len <= 0); 2218 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 2219 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 2220 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 2221 2222 ac->ac_found++; 2223 ac->ac_cX_found[ac->ac_criteria]++; 2224 2225 /* 2226 * The special case - take what you catch first 2227 */ 2228 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2229 *bex = *ex; 2230 ext4_mb_use_best_found(ac, e4b); 2231 return; 2232 } 2233 2234 /* 2235 * Let's check whether the chuck is good enough 2236 */ 2237 if (ex->fe_len == gex->fe_len) { 2238 *bex = *ex; 2239 ext4_mb_use_best_found(ac, e4b); 2240 return; 2241 } 2242 2243 /* 2244 * If this is first found extent, just store it in the context 2245 */ 2246 if (bex->fe_len == 0) { 2247 *bex = *ex; 2248 return; 2249 } 2250 2251 /* 2252 * If new found extent is better, store it in the context 2253 */ 2254 if (bex->fe_len < gex->fe_len) { 2255 /* if the request isn't satisfied, any found extent 2256 * larger than previous best one is better */ 2257 if (ex->fe_len > bex->fe_len) 2258 *bex = *ex; 2259 } else if (ex->fe_len > gex->fe_len) { 2260 /* if the request is satisfied, then we try to find 2261 * an extent that still satisfy the request, but is 2262 * smaller than previous one */ 2263 if (ex->fe_len < bex->fe_len) 2264 *bex = *ex; 2265 } 2266 2267 ext4_mb_check_limits(ac, e4b, 0); 2268 } 2269 2270 static noinline_for_stack 2271 void ext4_mb_try_best_found(struct ext4_allocation_context *ac, 2272 struct ext4_buddy *e4b) 2273 { 2274 struct ext4_free_extent ex = ac->ac_b_ex; 2275 ext4_group_t group = ex.fe_group; 2276 int max; 2277 int err; 2278 2279 BUG_ON(ex.fe_len <= 0); 2280 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 2281 if (err) 2282 return; 2283 2284 ext4_lock_group(ac->ac_sb, group); 2285 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); 2286 2287 if (max > 0) { 2288 ac->ac_b_ex = ex; 2289 ext4_mb_use_best_found(ac, e4b); 2290 } 2291 2292 ext4_unlock_group(ac->ac_sb, group); 2293 ext4_mb_unload_buddy(e4b); 2294 } 2295 2296 static noinline_for_stack 2297 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 2298 struct ext4_buddy *e4b) 2299 { 2300 ext4_group_t group = ac->ac_g_ex.fe_group; 2301 int max; 2302 int err; 2303 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2304 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2305 struct ext4_free_extent ex; 2306 2307 if (!grp) 2308 return -EFSCORRUPTED; 2309 if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY))) 2310 return 0; 2311 if (grp->bb_free == 0) 2312 return 0; 2313 2314 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 2315 if (err) 2316 return err; 2317 2318 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { 2319 ext4_mb_unload_buddy(e4b); 2320 return 0; 2321 } 2322 2323 ext4_lock_group(ac->ac_sb, group); 2324 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, 2325 ac->ac_g_ex.fe_len, &ex); 2326 ex.fe_logical = 0xDEADFA11; /* debug value */ 2327 2328 if (max >= ac->ac_g_ex.fe_len && 2329 ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) { 2330 ext4_fsblk_t start; 2331 2332 start = ext4_grp_offs_to_block(ac->ac_sb, &ex); 2333 /* use do_div to get remainder (would be 64-bit modulo) */ 2334 if (do_div(start, sbi->s_stripe) == 0) { 2335 ac->ac_found++; 2336 ac->ac_b_ex = ex; 2337 ext4_mb_use_best_found(ac, e4b); 2338 } 2339 } else if (max >= ac->ac_g_ex.fe_len) { 2340 BUG_ON(ex.fe_len <= 0); 2341 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 2342 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 2343 ac->ac_found++; 2344 ac->ac_b_ex = ex; 2345 ext4_mb_use_best_found(ac, e4b); 2346 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 2347 /* Sometimes, caller may want to merge even small 2348 * number of blocks to an existing extent */ 2349 BUG_ON(ex.fe_len <= 0); 2350 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 2351 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 2352 ac->ac_found++; 2353 ac->ac_b_ex = ex; 2354 ext4_mb_use_best_found(ac, e4b); 2355 } 2356 ext4_unlock_group(ac->ac_sb, group); 2357 ext4_mb_unload_buddy(e4b); 2358 2359 return 0; 2360 } 2361 2362 /* 2363 * The routine scans buddy structures (not bitmap!) from given order 2364 * to max order and tries to find big enough chunk to satisfy the req 2365 */ 2366 static noinline_for_stack 2367 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 2368 struct ext4_buddy *e4b) 2369 { 2370 struct super_block *sb = ac->ac_sb; 2371 struct ext4_group_info *grp = e4b->bd_info; 2372 void *buddy; 2373 int i; 2374 int k; 2375 int max; 2376 2377 BUG_ON(ac->ac_2order <= 0); 2378 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { 2379 if (grp->bb_counters[i] == 0) 2380 continue; 2381 2382 buddy = mb_find_buddy(e4b, i, &max); 2383 if (WARN_RATELIMIT(buddy == NULL, 2384 "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i)) 2385 continue; 2386 2387 k = mb_find_next_zero_bit(buddy, max, 0); 2388 if (k >= max) { 2389 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, 2390 "%d free clusters of order %d. But found 0", 2391 grp->bb_counters[i], i); 2392 ext4_mark_group_bitmap_corrupted(ac->ac_sb, 2393 e4b->bd_group, 2394 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2395 break; 2396 } 2397 ac->ac_found++; 2398 ac->ac_cX_found[ac->ac_criteria]++; 2399 2400 ac->ac_b_ex.fe_len = 1 << i; 2401 ac->ac_b_ex.fe_start = k << i; 2402 ac->ac_b_ex.fe_group = e4b->bd_group; 2403 2404 ext4_mb_use_best_found(ac, e4b); 2405 2406 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); 2407 2408 if (EXT4_SB(sb)->s_mb_stats) 2409 atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 2410 2411 break; 2412 } 2413 } 2414 2415 /* 2416 * The routine scans the group and measures all found extents. 2417 * In order to optimize scanning, caller must pass number of 2418 * free blocks in the group, so the routine can know upper limit. 2419 */ 2420 static noinline_for_stack 2421 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 2422 struct ext4_buddy *e4b) 2423 { 2424 struct super_block *sb = ac->ac_sb; 2425 void *bitmap = e4b->bd_bitmap; 2426 struct ext4_free_extent ex; 2427 int i, j, freelen; 2428 int free; 2429 2430 free = e4b->bd_info->bb_free; 2431 if (WARN_ON(free <= 0)) 2432 return; 2433 2434 i = e4b->bd_info->bb_first_free; 2435 2436 while (free && ac->ac_status == AC_STATUS_CONTINUE) { 2437 i = mb_find_next_zero_bit(bitmap, 2438 EXT4_CLUSTERS_PER_GROUP(sb), i); 2439 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 2440 /* 2441 * IF we have corrupt bitmap, we won't find any 2442 * free blocks even though group info says we 2443 * have free blocks 2444 */ 2445 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 2446 "%d free clusters as per " 2447 "group info. But bitmap says 0", 2448 free); 2449 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 2450 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2451 break; 2452 } 2453 2454 if (!ext4_mb_cr_expensive(ac->ac_criteria)) { 2455 /* 2456 * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are 2457 * sure that this group will have a large enough 2458 * continuous free extent, so skip over the smaller free 2459 * extents 2460 */ 2461 j = mb_find_next_bit(bitmap, 2462 EXT4_CLUSTERS_PER_GROUP(sb), i); 2463 freelen = j - i; 2464 2465 if (freelen < ac->ac_g_ex.fe_len) { 2466 i = j; 2467 free -= freelen; 2468 continue; 2469 } 2470 } 2471 2472 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); 2473 if (WARN_ON(ex.fe_len <= 0)) 2474 break; 2475 if (free < ex.fe_len) { 2476 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 2477 "%d free clusters as per " 2478 "group info. But got %d blocks", 2479 free, ex.fe_len); 2480 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 2481 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2482 /* 2483 * The number of free blocks differs. This mostly 2484 * indicate that the bitmap is corrupt. So exit 2485 * without claiming the space. 2486 */ 2487 break; 2488 } 2489 ex.fe_logical = 0xDEADC0DE; /* debug value */ 2490 ext4_mb_measure_extent(ac, &ex, e4b); 2491 2492 i += ex.fe_len; 2493 free -= ex.fe_len; 2494 } 2495 2496 ext4_mb_check_limits(ac, e4b, 1); 2497 } 2498 2499 /* 2500 * This is a special case for storages like raid5 2501 * we try to find stripe-aligned chunks for stripe-size-multiple requests 2502 */ 2503 static noinline_for_stack 2504 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 2505 struct ext4_buddy *e4b) 2506 { 2507 struct super_block *sb = ac->ac_sb; 2508 struct ext4_sb_info *sbi = EXT4_SB(sb); 2509 void *bitmap = e4b->bd_bitmap; 2510 struct ext4_free_extent ex; 2511 ext4_fsblk_t first_group_block; 2512 ext4_fsblk_t a; 2513 ext4_grpblk_t i, stripe; 2514 int max; 2515 2516 BUG_ON(sbi->s_stripe == 0); 2517 2518 /* find first stripe-aligned block in group */ 2519 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 2520 2521 a = first_group_block + sbi->s_stripe - 1; 2522 do_div(a, sbi->s_stripe); 2523 i = (a * sbi->s_stripe) - first_group_block; 2524 2525 stripe = EXT4_B2C(sbi, sbi->s_stripe); 2526 i = EXT4_B2C(sbi, i); 2527 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 2528 if (!mb_test_bit(i, bitmap)) { 2529 max = mb_find_extent(e4b, i, stripe, &ex); 2530 if (max >= stripe) { 2531 ac->ac_found++; 2532 ac->ac_cX_found[ac->ac_criteria]++; 2533 ex.fe_logical = 0xDEADF00D; /* debug value */ 2534 ac->ac_b_ex = ex; 2535 ext4_mb_use_best_found(ac, e4b); 2536 break; 2537 } 2538 } 2539 i += stripe; 2540 } 2541 } 2542 2543 /* 2544 * This is also called BEFORE we load the buddy bitmap. 2545 * Returns either 1 or 0 indicating that the group is either suitable 2546 * for the allocation or not. 2547 */ 2548 static bool ext4_mb_good_group(struct ext4_allocation_context *ac, 2549 ext4_group_t group, enum criteria cr) 2550 { 2551 ext4_grpblk_t free, fragments; 2552 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 2553 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2554 2555 BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS); 2556 2557 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 2558 return false; 2559 2560 free = grp->bb_free; 2561 if (free == 0) 2562 return false; 2563 2564 fragments = grp->bb_fragments; 2565 if (fragments == 0) 2566 return false; 2567 2568 switch (cr) { 2569 case CR_POWER2_ALIGNED: 2570 BUG_ON(ac->ac_2order == 0); 2571 2572 /* Avoid using the first bg of a flexgroup for data files */ 2573 if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 2574 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 2575 ((group % flex_size) == 0)) 2576 return false; 2577 2578 if (free < ac->ac_g_ex.fe_len) 2579 return false; 2580 2581 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) 2582 return true; 2583 2584 if (grp->bb_largest_free_order < ac->ac_2order) 2585 return false; 2586 2587 return true; 2588 case CR_GOAL_LEN_FAST: 2589 case CR_BEST_AVAIL_LEN: 2590 if ((free / fragments) >= ac->ac_g_ex.fe_len) 2591 return true; 2592 break; 2593 case CR_GOAL_LEN_SLOW: 2594 if (free >= ac->ac_g_ex.fe_len) 2595 return true; 2596 break; 2597 case CR_ANY_FREE: 2598 return true; 2599 default: 2600 BUG(); 2601 } 2602 2603 return false; 2604 } 2605 2606 /* 2607 * This could return negative error code if something goes wrong 2608 * during ext4_mb_init_group(). This should not be called with 2609 * ext4_lock_group() held. 2610 * 2611 * Note: because we are conditionally operating with the group lock in 2612 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this 2613 * function using __acquire and __release. This means we need to be 2614 * super careful before messing with the error path handling via "goto 2615 * out"! 2616 */ 2617 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, 2618 ext4_group_t group, enum criteria cr) 2619 { 2620 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2621 struct super_block *sb = ac->ac_sb; 2622 struct ext4_sb_info *sbi = EXT4_SB(sb); 2623 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; 2624 ext4_grpblk_t free; 2625 int ret = 0; 2626 2627 if (!grp) 2628 return -EFSCORRUPTED; 2629 if (sbi->s_mb_stats) 2630 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); 2631 if (should_lock) { 2632 ext4_lock_group(sb, group); 2633 __release(ext4_group_lock_ptr(sb, group)); 2634 } 2635 free = grp->bb_free; 2636 if (free == 0) 2637 goto out; 2638 /* 2639 * In all criterias except CR_ANY_FREE we try to avoid groups that 2640 * can't possibly satisfy the full goal request due to insufficient 2641 * free blocks. 2642 */ 2643 if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len) 2644 goto out; 2645 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 2646 goto out; 2647 if (should_lock) { 2648 __acquire(ext4_group_lock_ptr(sb, group)); 2649 ext4_unlock_group(sb, group); 2650 } 2651 2652 /* We only do this if the grp has never been initialized */ 2653 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 2654 struct ext4_group_desc *gdp = 2655 ext4_get_group_desc(sb, group, NULL); 2656 int ret; 2657 2658 /* 2659 * cr=CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic 2660 * search to find large good chunks almost for free. If buddy 2661 * data is not ready, then this optimization makes no sense. But 2662 * we never skip the first block group in a flex_bg, since this 2663 * gets used for metadata block allocation, and we want to make 2664 * sure we locate metadata blocks in the first block group in 2665 * the flex_bg if possible. 2666 */ 2667 if (!ext4_mb_cr_expensive(cr) && 2668 (!sbi->s_log_groups_per_flex || 2669 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && 2670 !(ext4_has_group_desc_csum(sb) && 2671 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) 2672 return 0; 2673 ret = ext4_mb_init_group(sb, group, GFP_NOFS); 2674 if (ret) 2675 return ret; 2676 } 2677 2678 if (should_lock) { 2679 ext4_lock_group(sb, group); 2680 __release(ext4_group_lock_ptr(sb, group)); 2681 } 2682 ret = ext4_mb_good_group(ac, group, cr); 2683 out: 2684 if (should_lock) { 2685 __acquire(ext4_group_lock_ptr(sb, group)); 2686 ext4_unlock_group(sb, group); 2687 } 2688 return ret; 2689 } 2690 2691 /* 2692 * Start prefetching @nr block bitmaps starting at @group. 2693 * Return the next group which needs to be prefetched. 2694 */ 2695 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, 2696 unsigned int nr, int *cnt) 2697 { 2698 ext4_group_t ngroups = ext4_get_groups_count(sb); 2699 struct buffer_head *bh; 2700 struct blk_plug plug; 2701 2702 blk_start_plug(&plug); 2703 while (nr-- > 0) { 2704 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, 2705 NULL); 2706 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 2707 2708 /* 2709 * Prefetch block groups with free blocks; but don't 2710 * bother if it is marked uninitialized on disk, since 2711 * it won't require I/O to read. Also only try to 2712 * prefetch once, so we avoid getblk() call, which can 2713 * be expensive. 2714 */ 2715 if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) && 2716 EXT4_MB_GRP_NEED_INIT(grp) && 2717 ext4_free_group_clusters(sb, gdp) > 0 ) { 2718 bh = ext4_read_block_bitmap_nowait(sb, group, true); 2719 if (bh && !IS_ERR(bh)) { 2720 if (!buffer_uptodate(bh) && cnt) 2721 (*cnt)++; 2722 brelse(bh); 2723 } 2724 } 2725 if (++group >= ngroups) 2726 group = 0; 2727 } 2728 blk_finish_plug(&plug); 2729 return group; 2730 } 2731 2732 /* 2733 * Prefetching reads the block bitmap into the buffer cache; but we 2734 * need to make sure that the buddy bitmap in the page cache has been 2735 * initialized. Note that ext4_mb_init_group() will block if the I/O 2736 * is not yet completed, or indeed if it was not initiated by 2737 * ext4_mb_prefetch did not start the I/O. 2738 * 2739 * TODO: We should actually kick off the buddy bitmap setup in a work 2740 * queue when the buffer I/O is completed, so that we don't block 2741 * waiting for the block allocation bitmap read to finish when 2742 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). 2743 */ 2744 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, 2745 unsigned int nr) 2746 { 2747 struct ext4_group_desc *gdp; 2748 struct ext4_group_info *grp; 2749 2750 while (nr-- > 0) { 2751 if (!group) 2752 group = ext4_get_groups_count(sb); 2753 group--; 2754 gdp = ext4_get_group_desc(sb, group, NULL); 2755 grp = ext4_get_group_info(sb, group); 2756 2757 if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) && 2758 ext4_free_group_clusters(sb, gdp) > 0) { 2759 if (ext4_mb_init_group(sb, group, GFP_NOFS)) 2760 break; 2761 } 2762 } 2763 } 2764 2765 static noinline_for_stack int 2766 ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2767 { 2768 ext4_group_t prefetch_grp = 0, ngroups, group, i; 2769 enum criteria new_cr, cr = CR_GOAL_LEN_FAST; 2770 int err = 0, first_err = 0; 2771 unsigned int nr = 0, prefetch_ios = 0; 2772 struct ext4_sb_info *sbi; 2773 struct super_block *sb; 2774 struct ext4_buddy e4b; 2775 int lost; 2776 2777 sb = ac->ac_sb; 2778 sbi = EXT4_SB(sb); 2779 ngroups = ext4_get_groups_count(sb); 2780 /* non-extent files are limited to low blocks/groups */ 2781 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2782 ngroups = sbi->s_blockfile_groups; 2783 2784 BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2785 2786 /* first, try the goal */ 2787 err = ext4_mb_find_by_goal(ac, &e4b); 2788 if (err || ac->ac_status == AC_STATUS_FOUND) 2789 goto out; 2790 2791 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2792 goto out; 2793 2794 /* 2795 * ac->ac_2order is set only if the fe_len is a power of 2 2796 * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED 2797 * so that we try exact allocation using buddy. 2798 */ 2799 i = fls(ac->ac_g_ex.fe_len); 2800 ac->ac_2order = 0; 2801 /* 2802 * We search using buddy data only if the order of the request 2803 * is greater than equal to the sbi_s_mb_order2_reqs 2804 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2805 * We also support searching for power-of-two requests only for 2806 * requests upto maximum buddy size we have constructed. 2807 */ 2808 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { 2809 if (is_power_of_2(ac->ac_g_ex.fe_len)) 2810 ac->ac_2order = array_index_nospec(i - 1, 2811 MB_NUM_ORDERS(sb)); 2812 } 2813 2814 /* if stream allocation is enabled, use global goal */ 2815 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2816 /* TBD: may be hot point */ 2817 spin_lock(&sbi->s_md_lock); 2818 ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2819 ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2820 spin_unlock(&sbi->s_md_lock); 2821 } 2822 2823 /* 2824 * Let's just scan groups to find more-less suitable blocks We 2825 * start with CR_GOAL_LEN_FAST, unless it is power of 2 2826 * aligned, in which case let's do that faster approach first. 2827 */ 2828 if (ac->ac_2order) 2829 cr = CR_POWER2_ALIGNED; 2830 repeat: 2831 for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2832 ac->ac_criteria = cr; 2833 /* 2834 * searching for the right group start 2835 * from the goal value specified 2836 */ 2837 group = ac->ac_g_ex.fe_group; 2838 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; 2839 prefetch_grp = group; 2840 2841 for (i = 0, new_cr = cr; i < ngroups; i++, 2842 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { 2843 int ret = 0; 2844 2845 cond_resched(); 2846 if (new_cr != cr) { 2847 cr = new_cr; 2848 goto repeat; 2849 } 2850 2851 /* 2852 * Batch reads of the block allocation bitmaps 2853 * to get multiple READs in flight; limit 2854 * prefetching at inexpensive CR, otherwise mballoc 2855 * can spend a lot of time loading imperfect groups 2856 */ 2857 if ((prefetch_grp == group) && 2858 (ext4_mb_cr_expensive(cr) || 2859 prefetch_ios < sbi->s_mb_prefetch_limit)) { 2860 nr = sbi->s_mb_prefetch; 2861 if (ext4_has_feature_flex_bg(sb)) { 2862 nr = 1 << sbi->s_log_groups_per_flex; 2863 nr -= group & (nr - 1); 2864 nr = min(nr, sbi->s_mb_prefetch); 2865 } 2866 prefetch_grp = ext4_mb_prefetch(sb, group, 2867 nr, &prefetch_ios); 2868 } 2869 2870 /* This now checks without needing the buddy page */ 2871 ret = ext4_mb_good_group_nolock(ac, group, cr); 2872 if (ret <= 0) { 2873 if (!first_err) 2874 first_err = ret; 2875 continue; 2876 } 2877 2878 err = ext4_mb_load_buddy(sb, group, &e4b); 2879 if (err) 2880 goto out; 2881 2882 ext4_lock_group(sb, group); 2883 2884 /* 2885 * We need to check again after locking the 2886 * block group 2887 */ 2888 ret = ext4_mb_good_group(ac, group, cr); 2889 if (ret == 0) { 2890 ext4_unlock_group(sb, group); 2891 ext4_mb_unload_buddy(&e4b); 2892 continue; 2893 } 2894 2895 ac->ac_groups_scanned++; 2896 if (cr == CR_POWER2_ALIGNED) 2897 ext4_mb_simple_scan_group(ac, &e4b); 2898 else if ((cr == CR_GOAL_LEN_FAST || 2899 cr == CR_BEST_AVAIL_LEN) && 2900 sbi->s_stripe && 2901 !(ac->ac_g_ex.fe_len % 2902 EXT4_B2C(sbi, sbi->s_stripe))) 2903 ext4_mb_scan_aligned(ac, &e4b); 2904 else 2905 ext4_mb_complex_scan_group(ac, &e4b); 2906 2907 ext4_unlock_group(sb, group); 2908 ext4_mb_unload_buddy(&e4b); 2909 2910 if (ac->ac_status != AC_STATUS_CONTINUE) 2911 break; 2912 } 2913 /* Processed all groups and haven't found blocks */ 2914 if (sbi->s_mb_stats && i == ngroups) 2915 atomic64_inc(&sbi->s_bal_cX_failed[cr]); 2916 2917 if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN) 2918 /* Reset goal length to original goal length before 2919 * falling into CR_GOAL_LEN_SLOW */ 2920 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; 2921 } 2922 2923 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2924 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2925 /* 2926 * We've been searching too long. Let's try to allocate 2927 * the best chunk we've found so far 2928 */ 2929 ext4_mb_try_best_found(ac, &e4b); 2930 if (ac->ac_status != AC_STATUS_FOUND) { 2931 /* 2932 * Someone more lucky has already allocated it. 2933 * The only thing we can do is just take first 2934 * found block(s) 2935 */ 2936 lost = atomic_inc_return(&sbi->s_mb_lost_chunks); 2937 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", 2938 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, 2939 ac->ac_b_ex.fe_len, lost); 2940 2941 ac->ac_b_ex.fe_group = 0; 2942 ac->ac_b_ex.fe_start = 0; 2943 ac->ac_b_ex.fe_len = 0; 2944 ac->ac_status = AC_STATUS_CONTINUE; 2945 ac->ac_flags |= EXT4_MB_HINT_FIRST; 2946 cr = CR_ANY_FREE; 2947 goto repeat; 2948 } 2949 } 2950 2951 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) 2952 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); 2953 out: 2954 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) 2955 err = first_err; 2956 2957 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", 2958 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, 2959 ac->ac_flags, cr, err); 2960 2961 if (nr) 2962 ext4_mb_prefetch_fini(sb, prefetch_grp, nr); 2963 2964 return err; 2965 } 2966 2967 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2968 { 2969 struct super_block *sb = pde_data(file_inode(seq->file)); 2970 ext4_group_t group; 2971 2972 if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2973 return NULL; 2974 group = *pos + 1; 2975 return (void *) ((unsigned long) group); 2976 } 2977 2978 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2979 { 2980 struct super_block *sb = pde_data(file_inode(seq->file)); 2981 ext4_group_t group; 2982 2983 ++*pos; 2984 if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2985 return NULL; 2986 group = *pos + 1; 2987 return (void *) ((unsigned long) group); 2988 } 2989 2990 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2991 { 2992 struct super_block *sb = pde_data(file_inode(seq->file)); 2993 ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2994 int i; 2995 int err, buddy_loaded = 0; 2996 struct ext4_buddy e4b; 2997 struct ext4_group_info *grinfo; 2998 unsigned char blocksize_bits = min_t(unsigned char, 2999 sb->s_blocksize_bits, 3000 EXT4_MAX_BLOCK_LOG_SIZE); 3001 struct sg { 3002 struct ext4_group_info info; 3003 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; 3004 } sg; 3005 3006 group--; 3007 if (group == 0) 3008 seq_puts(seq, "#group: free frags first [" 3009 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " 3010 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); 3011 3012 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 3013 sizeof(struct ext4_group_info); 3014 3015 grinfo = ext4_get_group_info(sb, group); 3016 if (!grinfo) 3017 return 0; 3018 /* Load the group info in memory only if not already loaded. */ 3019 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { 3020 err = ext4_mb_load_buddy(sb, group, &e4b); 3021 if (err) { 3022 seq_printf(seq, "#%-5u: I/O error\n", group); 3023 return 0; 3024 } 3025 buddy_loaded = 1; 3026 } 3027 3028 memcpy(&sg, grinfo, i); 3029 3030 if (buddy_loaded) 3031 ext4_mb_unload_buddy(&e4b); 3032 3033 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 3034 sg.info.bb_fragments, sg.info.bb_first_free); 3035 for (i = 0; i <= 13; i++) 3036 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? 3037 sg.info.bb_counters[i] : 0); 3038 seq_puts(seq, " ]\n"); 3039 3040 return 0; 3041 } 3042 3043 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 3044 { 3045 } 3046 3047 const struct seq_operations ext4_mb_seq_groups_ops = { 3048 .start = ext4_mb_seq_groups_start, 3049 .next = ext4_mb_seq_groups_next, 3050 .stop = ext4_mb_seq_groups_stop, 3051 .show = ext4_mb_seq_groups_show, 3052 }; 3053 3054 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) 3055 { 3056 struct super_block *sb = seq->private; 3057 struct ext4_sb_info *sbi = EXT4_SB(sb); 3058 3059 seq_puts(seq, "mballoc:\n"); 3060 if (!sbi->s_mb_stats) { 3061 seq_puts(seq, "\tmb stats collection turned off.\n"); 3062 seq_puts( 3063 seq, 3064 "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); 3065 return 0; 3066 } 3067 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); 3068 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); 3069 3070 seq_printf(seq, "\tgroups_scanned: %u\n", 3071 atomic_read(&sbi->s_bal_groups_scanned)); 3072 3073 /* CR_POWER2_ALIGNED stats */ 3074 seq_puts(seq, "\tcr_p2_aligned_stats:\n"); 3075 seq_printf(seq, "\t\thits: %llu\n", 3076 atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED])); 3077 seq_printf( 3078 seq, "\t\tgroups_considered: %llu\n", 3079 atomic64_read( 3080 &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED])); 3081 seq_printf(seq, "\t\textents_scanned: %u\n", 3082 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED])); 3083 seq_printf(seq, "\t\tuseless_loops: %llu\n", 3084 atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED])); 3085 seq_printf(seq, "\t\tbad_suggestions: %u\n", 3086 atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions)); 3087 3088 /* CR_GOAL_LEN_FAST stats */ 3089 seq_puts(seq, "\tcr_goal_fast_stats:\n"); 3090 seq_printf(seq, "\t\thits: %llu\n", 3091 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST])); 3092 seq_printf(seq, "\t\tgroups_considered: %llu\n", 3093 atomic64_read( 3094 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST])); 3095 seq_printf(seq, "\t\textents_scanned: %u\n", 3096 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST])); 3097 seq_printf(seq, "\t\tuseless_loops: %llu\n", 3098 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST])); 3099 seq_printf(seq, "\t\tbad_suggestions: %u\n", 3100 atomic_read(&sbi->s_bal_goal_fast_bad_suggestions)); 3101 3102 /* CR_BEST_AVAIL_LEN stats */ 3103 seq_puts(seq, "\tcr_best_avail_stats:\n"); 3104 seq_printf(seq, "\t\thits: %llu\n", 3105 atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN])); 3106 seq_printf( 3107 seq, "\t\tgroups_considered: %llu\n", 3108 atomic64_read( 3109 &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN])); 3110 seq_printf(seq, "\t\textents_scanned: %u\n", 3111 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN])); 3112 seq_printf(seq, "\t\tuseless_loops: %llu\n", 3113 atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN])); 3114 seq_printf(seq, "\t\tbad_suggestions: %u\n", 3115 atomic_read(&sbi->s_bal_best_avail_bad_suggestions)); 3116 3117 /* CR_GOAL_LEN_SLOW stats */ 3118 seq_puts(seq, "\tcr_goal_slow_stats:\n"); 3119 seq_printf(seq, "\t\thits: %llu\n", 3120 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW])); 3121 seq_printf(seq, "\t\tgroups_considered: %llu\n", 3122 atomic64_read( 3123 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW])); 3124 seq_printf(seq, "\t\textents_scanned: %u\n", 3125 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW])); 3126 seq_printf(seq, "\t\tuseless_loops: %llu\n", 3127 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW])); 3128 3129 /* CR_ANY_FREE stats */ 3130 seq_puts(seq, "\tcr_any_free_stats:\n"); 3131 seq_printf(seq, "\t\thits: %llu\n", 3132 atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE])); 3133 seq_printf( 3134 seq, "\t\tgroups_considered: %llu\n", 3135 atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE])); 3136 seq_printf(seq, "\t\textents_scanned: %u\n", 3137 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE])); 3138 seq_printf(seq, "\t\tuseless_loops: %llu\n", 3139 atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE])); 3140 3141 /* Aggregates */ 3142 seq_printf(seq, "\textents_scanned: %u\n", 3143 atomic_read(&sbi->s_bal_ex_scanned)); 3144 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); 3145 seq_printf(seq, "\t\tlen_goal_hits: %u\n", 3146 atomic_read(&sbi->s_bal_len_goals)); 3147 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); 3148 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); 3149 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); 3150 seq_printf(seq, "\tbuddies_generated: %u/%u\n", 3151 atomic_read(&sbi->s_mb_buddies_generated), 3152 ext4_get_groups_count(sb)); 3153 seq_printf(seq, "\tbuddies_time_used: %llu\n", 3154 atomic64_read(&sbi->s_mb_generation_time)); 3155 seq_printf(seq, "\tpreallocated: %u\n", 3156 atomic_read(&sbi->s_mb_preallocated)); 3157 seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded)); 3158 return 0; 3159 } 3160 3161 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) 3162 __acquires(&EXT4_SB(sb)->s_mb_rb_lock) 3163 { 3164 struct super_block *sb = pde_data(file_inode(seq->file)); 3165 unsigned long position; 3166 3167 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) 3168 return NULL; 3169 position = *pos + 1; 3170 return (void *) ((unsigned long) position); 3171 } 3172 3173 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) 3174 { 3175 struct super_block *sb = pde_data(file_inode(seq->file)); 3176 unsigned long position; 3177 3178 ++*pos; 3179 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) 3180 return NULL; 3181 position = *pos + 1; 3182 return (void *) ((unsigned long) position); 3183 } 3184 3185 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) 3186 { 3187 struct super_block *sb = pde_data(file_inode(seq->file)); 3188 struct ext4_sb_info *sbi = EXT4_SB(sb); 3189 unsigned long position = ((unsigned long) v); 3190 struct ext4_group_info *grp; 3191 unsigned int count; 3192 3193 position--; 3194 if (position >= MB_NUM_ORDERS(sb)) { 3195 position -= MB_NUM_ORDERS(sb); 3196 if (position == 0) 3197 seq_puts(seq, "avg_fragment_size_lists:\n"); 3198 3199 count = 0; 3200 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]); 3201 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position], 3202 bb_avg_fragment_size_node) 3203 count++; 3204 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]); 3205 seq_printf(seq, "\tlist_order_%u_groups: %u\n", 3206 (unsigned int)position, count); 3207 return 0; 3208 } 3209 3210 if (position == 0) { 3211 seq_printf(seq, "optimize_scan: %d\n", 3212 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); 3213 seq_puts(seq, "max_free_order_lists:\n"); 3214 } 3215 count = 0; 3216 read_lock(&sbi->s_mb_largest_free_orders_locks[position]); 3217 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], 3218 bb_largest_free_order_node) 3219 count++; 3220 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]); 3221 seq_printf(seq, "\tlist_order_%u_groups: %u\n", 3222 (unsigned int)position, count); 3223 3224 return 0; 3225 } 3226 3227 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) 3228 { 3229 } 3230 3231 const struct seq_operations ext4_mb_seq_structs_summary_ops = { 3232 .start = ext4_mb_seq_structs_summary_start, 3233 .next = ext4_mb_seq_structs_summary_next, 3234 .stop = ext4_mb_seq_structs_summary_stop, 3235 .show = ext4_mb_seq_structs_summary_show, 3236 }; 3237 3238 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 3239 { 3240 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 3241 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 3242 3243 BUG_ON(!cachep); 3244 return cachep; 3245 } 3246 3247 /* 3248 * Allocate the top-level s_group_info array for the specified number 3249 * of groups 3250 */ 3251 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) 3252 { 3253 struct ext4_sb_info *sbi = EXT4_SB(sb); 3254 unsigned size; 3255 struct ext4_group_info ***old_groupinfo, ***new_groupinfo; 3256 3257 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> 3258 EXT4_DESC_PER_BLOCK_BITS(sb); 3259 if (size <= sbi->s_group_info_size) 3260 return 0; 3261 3262 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); 3263 new_groupinfo = kvzalloc(size, GFP_KERNEL); 3264 if (!new_groupinfo) { 3265 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 3266 return -ENOMEM; 3267 } 3268 rcu_read_lock(); 3269 old_groupinfo = rcu_dereference(sbi->s_group_info); 3270 if (old_groupinfo) 3271 memcpy(new_groupinfo, old_groupinfo, 3272 sbi->s_group_info_size * sizeof(*sbi->s_group_info)); 3273 rcu_read_unlock(); 3274 rcu_assign_pointer(sbi->s_group_info, new_groupinfo); 3275 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); 3276 if (old_groupinfo) 3277 ext4_kvfree_array_rcu(old_groupinfo); 3278 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 3279 sbi->s_group_info_size); 3280 return 0; 3281 } 3282 3283 /* Create and initialize ext4_group_info data for the given group. */ 3284 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 3285 struct ext4_group_desc *desc) 3286 { 3287 int i; 3288 int metalen = 0; 3289 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); 3290 struct ext4_sb_info *sbi = EXT4_SB(sb); 3291 struct ext4_group_info **meta_group_info; 3292 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 3293 3294 /* 3295 * First check if this group is the first of a reserved block. 3296 * If it's true, we have to allocate a new table of pointers 3297 * to ext4_group_info structures 3298 */ 3299 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 3300 metalen = sizeof(*meta_group_info) << 3301 EXT4_DESC_PER_BLOCK_BITS(sb); 3302 meta_group_info = kmalloc(metalen, GFP_NOFS); 3303 if (meta_group_info == NULL) { 3304 ext4_msg(sb, KERN_ERR, "can't allocate mem " 3305 "for a buddy group"); 3306 return -ENOMEM; 3307 } 3308 rcu_read_lock(); 3309 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; 3310 rcu_read_unlock(); 3311 } 3312 3313 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); 3314 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 3315 3316 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); 3317 if (meta_group_info[i] == NULL) { 3318 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 3319 goto exit_group_info; 3320 } 3321 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 3322 &(meta_group_info[i]->bb_state)); 3323 3324 /* 3325 * initialize bb_free to be able to skip 3326 * empty groups without initialization 3327 */ 3328 if (ext4_has_group_desc_csum(sb) && 3329 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 3330 meta_group_info[i]->bb_free = 3331 ext4_free_clusters_after_init(sb, group, desc); 3332 } else { 3333 meta_group_info[i]->bb_free = 3334 ext4_free_group_clusters(sb, desc); 3335 } 3336 3337 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 3338 init_rwsem(&meta_group_info[i]->alloc_sem); 3339 meta_group_info[i]->bb_free_root = RB_ROOT; 3340 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); 3341 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); 3342 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 3343 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ 3344 meta_group_info[i]->bb_group = group; 3345 3346 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); 3347 return 0; 3348 3349 exit_group_info: 3350 /* If a meta_group_info table has been allocated, release it now */ 3351 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 3352 struct ext4_group_info ***group_info; 3353 3354 rcu_read_lock(); 3355 group_info = rcu_dereference(sbi->s_group_info); 3356 kfree(group_info[idx]); 3357 group_info[idx] = NULL; 3358 rcu_read_unlock(); 3359 } 3360 return -ENOMEM; 3361 } /* ext4_mb_add_groupinfo */ 3362 3363 static int ext4_mb_init_backend(struct super_block *sb) 3364 { 3365 ext4_group_t ngroups = ext4_get_groups_count(sb); 3366 ext4_group_t i; 3367 struct ext4_sb_info *sbi = EXT4_SB(sb); 3368 int err; 3369 struct ext4_group_desc *desc; 3370 struct ext4_group_info ***group_info; 3371 struct kmem_cache *cachep; 3372 3373 err = ext4_mb_alloc_groupinfo(sb, ngroups); 3374 if (err) 3375 return err; 3376 3377 sbi->s_buddy_cache = new_inode(sb); 3378 if (sbi->s_buddy_cache == NULL) { 3379 ext4_msg(sb, KERN_ERR, "can't get new inode"); 3380 goto err_freesgi; 3381 } 3382 /* To avoid potentially colliding with an valid on-disk inode number, 3383 * use EXT4_BAD_INO for the buddy cache inode number. This inode is 3384 * not in the inode hash, so it should never be found by iget(), but 3385 * this will avoid confusion if it ever shows up during debugging. */ 3386 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 3387 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 3388 for (i = 0; i < ngroups; i++) { 3389 cond_resched(); 3390 desc = ext4_get_group_desc(sb, i, NULL); 3391 if (desc == NULL) { 3392 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 3393 goto err_freebuddy; 3394 } 3395 if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 3396 goto err_freebuddy; 3397 } 3398 3399 if (ext4_has_feature_flex_bg(sb)) { 3400 /* a single flex group is supposed to be read by a single IO. 3401 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is 3402 * unsigned integer, so the maximum shift is 32. 3403 */ 3404 if (sbi->s_es->s_log_groups_per_flex >= 32) { 3405 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); 3406 goto err_freebuddy; 3407 } 3408 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, 3409 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); 3410 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ 3411 } else { 3412 sbi->s_mb_prefetch = 32; 3413 } 3414 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) 3415 sbi->s_mb_prefetch = ext4_get_groups_count(sb); 3416 /* now many real IOs to prefetch within a single allocation at cr=0 3417 * given cr=0 is an CPU-related optimization we shouldn't try to 3418 * load too many groups, at some point we should start to use what 3419 * we've got in memory. 3420 * with an average random access time 5ms, it'd take a second to get 3421 * 200 groups (* N with flex_bg), so let's make this limit 4 3422 */ 3423 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; 3424 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) 3425 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); 3426 3427 return 0; 3428 3429 err_freebuddy: 3430 cachep = get_groupinfo_cache(sb->s_blocksize_bits); 3431 while (i-- > 0) { 3432 struct ext4_group_info *grp = ext4_get_group_info(sb, i); 3433 3434 if (grp) 3435 kmem_cache_free(cachep, grp); 3436 } 3437 i = sbi->s_group_info_size; 3438 rcu_read_lock(); 3439 group_info = rcu_dereference(sbi->s_group_info); 3440 while (i-- > 0) 3441 kfree(group_info[i]); 3442 rcu_read_unlock(); 3443 iput(sbi->s_buddy_cache); 3444 err_freesgi: 3445 rcu_read_lock(); 3446 kvfree(rcu_dereference(sbi->s_group_info)); 3447 rcu_read_unlock(); 3448 return -ENOMEM; 3449 } 3450 3451 static void ext4_groupinfo_destroy_slabs(void) 3452 { 3453 int i; 3454 3455 for (i = 0; i < NR_GRPINFO_CACHES; i++) { 3456 kmem_cache_destroy(ext4_groupinfo_caches[i]); 3457 ext4_groupinfo_caches[i] = NULL; 3458 } 3459 } 3460 3461 static int ext4_groupinfo_create_slab(size_t size) 3462 { 3463 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 3464 int slab_size; 3465 int blocksize_bits = order_base_2(size); 3466 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 3467 struct kmem_cache *cachep; 3468 3469 if (cache_index >= NR_GRPINFO_CACHES) 3470 return -EINVAL; 3471 3472 if (unlikely(cache_index < 0)) 3473 cache_index = 0; 3474 3475 mutex_lock(&ext4_grpinfo_slab_create_mutex); 3476 if (ext4_groupinfo_caches[cache_index]) { 3477 mutex_unlock(&ext4_grpinfo_slab_create_mutex); 3478 return 0; /* Already created */ 3479 } 3480 3481 slab_size = offsetof(struct ext4_group_info, 3482 bb_counters[blocksize_bits + 2]); 3483 3484 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 3485 slab_size, 0, SLAB_RECLAIM_ACCOUNT, 3486 NULL); 3487 3488 ext4_groupinfo_caches[cache_index] = cachep; 3489 3490 mutex_unlock(&ext4_grpinfo_slab_create_mutex); 3491 if (!cachep) { 3492 printk(KERN_EMERG 3493 "EXT4-fs: no memory for groupinfo slab cache\n"); 3494 return -ENOMEM; 3495 } 3496 3497 return 0; 3498 } 3499 3500 static void ext4_discard_work(struct work_struct *work) 3501 { 3502 struct ext4_sb_info *sbi = container_of(work, 3503 struct ext4_sb_info, s_discard_work); 3504 struct super_block *sb = sbi->s_sb; 3505 struct ext4_free_data *fd, *nfd; 3506 struct ext4_buddy e4b; 3507 struct list_head discard_list; 3508 ext4_group_t grp, load_grp; 3509 int err = 0; 3510 3511 INIT_LIST_HEAD(&discard_list); 3512 spin_lock(&sbi->s_md_lock); 3513 list_splice_init(&sbi->s_discard_list, &discard_list); 3514 spin_unlock(&sbi->s_md_lock); 3515 3516 load_grp = UINT_MAX; 3517 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { 3518 /* 3519 * If filesystem is umounting or no memory or suffering 3520 * from no space, give up the discard 3521 */ 3522 if ((sb->s_flags & SB_ACTIVE) && !err && 3523 !atomic_read(&sbi->s_retry_alloc_pending)) { 3524 grp = fd->efd_group; 3525 if (grp != load_grp) { 3526 if (load_grp != UINT_MAX) 3527 ext4_mb_unload_buddy(&e4b); 3528 3529 err = ext4_mb_load_buddy(sb, grp, &e4b); 3530 if (err) { 3531 kmem_cache_free(ext4_free_data_cachep, fd); 3532 load_grp = UINT_MAX; 3533 continue; 3534 } else { 3535 load_grp = grp; 3536 } 3537 } 3538 3539 ext4_lock_group(sb, grp); 3540 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, 3541 fd->efd_start_cluster + fd->efd_count - 1, 1); 3542 ext4_unlock_group(sb, grp); 3543 } 3544 kmem_cache_free(ext4_free_data_cachep, fd); 3545 } 3546 3547 if (load_grp != UINT_MAX) 3548 ext4_mb_unload_buddy(&e4b); 3549 } 3550 3551 int ext4_mb_init(struct super_block *sb) 3552 { 3553 struct ext4_sb_info *sbi = EXT4_SB(sb); 3554 unsigned i, j; 3555 unsigned offset, offset_incr; 3556 unsigned max; 3557 int ret; 3558 3559 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); 3560 3561 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 3562 if (sbi->s_mb_offsets == NULL) { 3563 ret = -ENOMEM; 3564 goto out; 3565 } 3566 3567 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); 3568 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 3569 if (sbi->s_mb_maxs == NULL) { 3570 ret = -ENOMEM; 3571 goto out; 3572 } 3573 3574 ret = ext4_groupinfo_create_slab(sb->s_blocksize); 3575 if (ret < 0) 3576 goto out; 3577 3578 /* order 0 is regular bitmap */ 3579 sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 3580 sbi->s_mb_offsets[0] = 0; 3581 3582 i = 1; 3583 offset = 0; 3584 offset_incr = 1 << (sb->s_blocksize_bits - 1); 3585 max = sb->s_blocksize << 2; 3586 do { 3587 sbi->s_mb_offsets[i] = offset; 3588 sbi->s_mb_maxs[i] = max; 3589 offset += offset_incr; 3590 offset_incr = offset_incr >> 1; 3591 max = max >> 1; 3592 i++; 3593 } while (i < MB_NUM_ORDERS(sb)); 3594 3595 sbi->s_mb_avg_fragment_size = 3596 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), 3597 GFP_KERNEL); 3598 if (!sbi->s_mb_avg_fragment_size) { 3599 ret = -ENOMEM; 3600 goto out; 3601 } 3602 sbi->s_mb_avg_fragment_size_locks = 3603 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), 3604 GFP_KERNEL); 3605 if (!sbi->s_mb_avg_fragment_size_locks) { 3606 ret = -ENOMEM; 3607 goto out; 3608 } 3609 for (i = 0; i < MB_NUM_ORDERS(sb); i++) { 3610 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); 3611 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); 3612 } 3613 sbi->s_mb_largest_free_orders = 3614 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), 3615 GFP_KERNEL); 3616 if (!sbi->s_mb_largest_free_orders) { 3617 ret = -ENOMEM; 3618 goto out; 3619 } 3620 sbi->s_mb_largest_free_orders_locks = 3621 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), 3622 GFP_KERNEL); 3623 if (!sbi->s_mb_largest_free_orders_locks) { 3624 ret = -ENOMEM; 3625 goto out; 3626 } 3627 for (i = 0; i < MB_NUM_ORDERS(sb); i++) { 3628 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); 3629 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); 3630 } 3631 3632 spin_lock_init(&sbi->s_md_lock); 3633 sbi->s_mb_free_pending = 0; 3634 INIT_LIST_HEAD(&sbi->s_freed_data_list); 3635 INIT_LIST_HEAD(&sbi->s_discard_list); 3636 INIT_WORK(&sbi->s_discard_work, ext4_discard_work); 3637 atomic_set(&sbi->s_retry_alloc_pending, 0); 3638 3639 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 3640 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 3641 sbi->s_mb_stats = MB_DEFAULT_STATS; 3642 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 3643 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 3644 sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER; 3645 3646 /* 3647 * The default group preallocation is 512, which for 4k block 3648 * sizes translates to 2 megabytes. However for bigalloc file 3649 * systems, this is probably too big (i.e, if the cluster size 3650 * is 1 megabyte, then group preallocation size becomes half a 3651 * gigabyte!). As a default, we will keep a two megabyte 3652 * group pralloc size for cluster sizes up to 64k, and after 3653 * that, we will force a minimum group preallocation size of 3654 * 32 clusters. This translates to 8 megs when the cluster 3655 * size is 256k, and 32 megs when the cluster size is 1 meg, 3656 * which seems reasonable as a default. 3657 */ 3658 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 3659 sbi->s_cluster_bits, 32); 3660 /* 3661 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 3662 * to the lowest multiple of s_stripe which is bigger than 3663 * the s_mb_group_prealloc as determined above. We want 3664 * the preallocation size to be an exact multiple of the 3665 * RAID stripe size so that preallocations don't fragment 3666 * the stripes. 3667 */ 3668 if (sbi->s_stripe > 1) { 3669 sbi->s_mb_group_prealloc = roundup( 3670 sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe)); 3671 } 3672 3673 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 3674 if (sbi->s_locality_groups == NULL) { 3675 ret = -ENOMEM; 3676 goto out; 3677 } 3678 for_each_possible_cpu(i) { 3679 struct ext4_locality_group *lg; 3680 lg = per_cpu_ptr(sbi->s_locality_groups, i); 3681 mutex_init(&lg->lg_mutex); 3682 for (j = 0; j < PREALLOC_TB_SIZE; j++) 3683 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 3684 spin_lock_init(&lg->lg_prealloc_lock); 3685 } 3686 3687 if (bdev_nonrot(sb->s_bdev)) 3688 sbi->s_mb_max_linear_groups = 0; 3689 else 3690 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; 3691 /* init file for buddy data */ 3692 ret = ext4_mb_init_backend(sb); 3693 if (ret != 0) 3694 goto out_free_locality_groups; 3695 3696 return 0; 3697 3698 out_free_locality_groups: 3699 free_percpu(sbi->s_locality_groups); 3700 sbi->s_locality_groups = NULL; 3701 out: 3702 kfree(sbi->s_mb_avg_fragment_size); 3703 kfree(sbi->s_mb_avg_fragment_size_locks); 3704 kfree(sbi->s_mb_largest_free_orders); 3705 kfree(sbi->s_mb_largest_free_orders_locks); 3706 kfree(sbi->s_mb_offsets); 3707 sbi->s_mb_offsets = NULL; 3708 kfree(sbi->s_mb_maxs); 3709 sbi->s_mb_maxs = NULL; 3710 return ret; 3711 } 3712 3713 /* need to called with the ext4 group lock held */ 3714 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) 3715 { 3716 struct ext4_prealloc_space *pa; 3717 struct list_head *cur, *tmp; 3718 int count = 0; 3719 3720 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 3721 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3722 list_del(&pa->pa_group_list); 3723 count++; 3724 kmem_cache_free(ext4_pspace_cachep, pa); 3725 } 3726 return count; 3727 } 3728 3729 int ext4_mb_release(struct super_block *sb) 3730 { 3731 ext4_group_t ngroups = ext4_get_groups_count(sb); 3732 ext4_group_t i; 3733 int num_meta_group_infos; 3734 struct ext4_group_info *grinfo, ***group_info; 3735 struct ext4_sb_info *sbi = EXT4_SB(sb); 3736 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 3737 int count; 3738 3739 if (test_opt(sb, DISCARD)) { 3740 /* 3741 * wait the discard work to drain all of ext4_free_data 3742 */ 3743 flush_work(&sbi->s_discard_work); 3744 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); 3745 } 3746 3747 if (sbi->s_group_info) { 3748 for (i = 0; i < ngroups; i++) { 3749 cond_resched(); 3750 grinfo = ext4_get_group_info(sb, i); 3751 if (!grinfo) 3752 continue; 3753 mb_group_bb_bitmap_free(grinfo); 3754 ext4_lock_group(sb, i); 3755 count = ext4_mb_cleanup_pa(grinfo); 3756 if (count) 3757 mb_debug(sb, "mballoc: %d PAs left\n", 3758 count); 3759 ext4_unlock_group(sb, i); 3760 kmem_cache_free(cachep, grinfo); 3761 } 3762 num_meta_group_infos = (ngroups + 3763 EXT4_DESC_PER_BLOCK(sb) - 1) >> 3764 EXT4_DESC_PER_BLOCK_BITS(sb); 3765 rcu_read_lock(); 3766 group_info = rcu_dereference(sbi->s_group_info); 3767 for (i = 0; i < num_meta_group_infos; i++) 3768 kfree(group_info[i]); 3769 kvfree(group_info); 3770 rcu_read_unlock(); 3771 } 3772 kfree(sbi->s_mb_avg_fragment_size); 3773 kfree(sbi->s_mb_avg_fragment_size_locks); 3774 kfree(sbi->s_mb_largest_free_orders); 3775 kfree(sbi->s_mb_largest_free_orders_locks); 3776 kfree(sbi->s_mb_offsets); 3777 kfree(sbi->s_mb_maxs); 3778 iput(sbi->s_buddy_cache); 3779 if (sbi->s_mb_stats) { 3780 ext4_msg(sb, KERN_INFO, 3781 "mballoc: %u blocks %u reqs (%u success)", 3782 atomic_read(&sbi->s_bal_allocated), 3783 atomic_read(&sbi->s_bal_reqs), 3784 atomic_read(&sbi->s_bal_success)); 3785 ext4_msg(sb, KERN_INFO, 3786 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " 3787 "%u 2^N hits, %u breaks, %u lost", 3788 atomic_read(&sbi->s_bal_ex_scanned), 3789 atomic_read(&sbi->s_bal_groups_scanned), 3790 atomic_read(&sbi->s_bal_goals), 3791 atomic_read(&sbi->s_bal_2orders), 3792 atomic_read(&sbi->s_bal_breaks), 3793 atomic_read(&sbi->s_mb_lost_chunks)); 3794 ext4_msg(sb, KERN_INFO, 3795 "mballoc: %u generated and it took %llu", 3796 atomic_read(&sbi->s_mb_buddies_generated), 3797 atomic64_read(&sbi->s_mb_generation_time)); 3798 ext4_msg(sb, KERN_INFO, 3799 "mballoc: %u preallocated, %u discarded", 3800 atomic_read(&sbi->s_mb_preallocated), 3801 atomic_read(&sbi->s_mb_discarded)); 3802 } 3803 3804 free_percpu(sbi->s_locality_groups); 3805 3806 return 0; 3807 } 3808 3809 static inline int ext4_issue_discard(struct super_block *sb, 3810 ext4_group_t block_group, ext4_grpblk_t cluster, int count, 3811 struct bio **biop) 3812 { 3813 ext4_fsblk_t discard_block; 3814 3815 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 3816 ext4_group_first_block_no(sb, block_group)); 3817 count = EXT4_C2B(EXT4_SB(sb), count); 3818 trace_ext4_discard_blocks(sb, 3819 (unsigned long long) discard_block, count); 3820 if (biop) { 3821 return __blkdev_issue_discard(sb->s_bdev, 3822 (sector_t)discard_block << (sb->s_blocksize_bits - 9), 3823 (sector_t)count << (sb->s_blocksize_bits - 9), 3824 GFP_NOFS, biop); 3825 } else 3826 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 3827 } 3828 3829 static void ext4_free_data_in_buddy(struct super_block *sb, 3830 struct ext4_free_data *entry) 3831 { 3832 struct ext4_buddy e4b; 3833 struct ext4_group_info *db; 3834 int err, count = 0; 3835 3836 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", 3837 entry->efd_count, entry->efd_group, entry); 3838 3839 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 3840 /* we expect to find existing buddy because it's pinned */ 3841 BUG_ON(err != 0); 3842 3843 spin_lock(&EXT4_SB(sb)->s_md_lock); 3844 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; 3845 spin_unlock(&EXT4_SB(sb)->s_md_lock); 3846 3847 db = e4b.bd_info; 3848 /* there are blocks to put in buddy to make them really free */ 3849 count += entry->efd_count; 3850 ext4_lock_group(sb, entry->efd_group); 3851 /* Take it out of per group rb tree */ 3852 rb_erase(&entry->efd_node, &(db->bb_free_root)); 3853 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 3854 3855 /* 3856 * Clear the trimmed flag for the group so that the next 3857 * ext4_trim_fs can trim it. 3858 * If the volume is mounted with -o discard, online discard 3859 * is supported and the free blocks will be trimmed online. 3860 */ 3861 if (!test_opt(sb, DISCARD)) 3862 EXT4_MB_GRP_CLEAR_TRIMMED(db); 3863 3864 if (!db->bb_free_root.rb_node) { 3865 /* No more items in the per group rb tree 3866 * balance refcounts from ext4_mb_free_metadata() 3867 */ 3868 put_page(e4b.bd_buddy_page); 3869 put_page(e4b.bd_bitmap_page); 3870 } 3871 ext4_unlock_group(sb, entry->efd_group); 3872 ext4_mb_unload_buddy(&e4b); 3873 3874 mb_debug(sb, "freed %d blocks in 1 structures\n", count); 3875 } 3876 3877 /* 3878 * This function is called by the jbd2 layer once the commit has finished, 3879 * so we know we can free the blocks that were released with that commit. 3880 */ 3881 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) 3882 { 3883 struct ext4_sb_info *sbi = EXT4_SB(sb); 3884 struct ext4_free_data *entry, *tmp; 3885 struct list_head freed_data_list; 3886 struct list_head *cut_pos = NULL; 3887 bool wake; 3888 3889 INIT_LIST_HEAD(&freed_data_list); 3890 3891 spin_lock(&sbi->s_md_lock); 3892 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { 3893 if (entry->efd_tid != commit_tid) 3894 break; 3895 cut_pos = &entry->efd_list; 3896 } 3897 if (cut_pos) 3898 list_cut_position(&freed_data_list, &sbi->s_freed_data_list, 3899 cut_pos); 3900 spin_unlock(&sbi->s_md_lock); 3901 3902 list_for_each_entry(entry, &freed_data_list, efd_list) 3903 ext4_free_data_in_buddy(sb, entry); 3904 3905 if (test_opt(sb, DISCARD)) { 3906 spin_lock(&sbi->s_md_lock); 3907 wake = list_empty(&sbi->s_discard_list); 3908 list_splice_tail(&freed_data_list, &sbi->s_discard_list); 3909 spin_unlock(&sbi->s_md_lock); 3910 if (wake) 3911 queue_work(system_unbound_wq, &sbi->s_discard_work); 3912 } else { 3913 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) 3914 kmem_cache_free(ext4_free_data_cachep, entry); 3915 } 3916 } 3917 3918 int __init ext4_init_mballoc(void) 3919 { 3920 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 3921 SLAB_RECLAIM_ACCOUNT); 3922 if (ext4_pspace_cachep == NULL) 3923 goto out; 3924 3925 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 3926 SLAB_RECLAIM_ACCOUNT); 3927 if (ext4_ac_cachep == NULL) 3928 goto out_pa_free; 3929 3930 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 3931 SLAB_RECLAIM_ACCOUNT); 3932 if (ext4_free_data_cachep == NULL) 3933 goto out_ac_free; 3934 3935 return 0; 3936 3937 out_ac_free: 3938 kmem_cache_destroy(ext4_ac_cachep); 3939 out_pa_free: 3940 kmem_cache_destroy(ext4_pspace_cachep); 3941 out: 3942 return -ENOMEM; 3943 } 3944 3945 void ext4_exit_mballoc(void) 3946 { 3947 /* 3948 * Wait for completion of call_rcu()'s on ext4_pspace_cachep 3949 * before destroying the slab cache. 3950 */ 3951 rcu_barrier(); 3952 kmem_cache_destroy(ext4_pspace_cachep); 3953 kmem_cache_destroy(ext4_ac_cachep); 3954 kmem_cache_destroy(ext4_free_data_cachep); 3955 ext4_groupinfo_destroy_slabs(); 3956 } 3957 3958 3959 /* 3960 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 3961 * Returns 0 if success or error code 3962 */ 3963 static noinline_for_stack int 3964 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 3965 handle_t *handle, unsigned int reserv_clstrs) 3966 { 3967 struct buffer_head *bitmap_bh = NULL; 3968 struct ext4_group_desc *gdp; 3969 struct buffer_head *gdp_bh; 3970 struct ext4_sb_info *sbi; 3971 struct super_block *sb; 3972 ext4_fsblk_t block; 3973 int err, len; 3974 3975 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3976 BUG_ON(ac->ac_b_ex.fe_len <= 0); 3977 3978 sb = ac->ac_sb; 3979 sbi = EXT4_SB(sb); 3980 3981 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 3982 if (IS_ERR(bitmap_bh)) { 3983 return PTR_ERR(bitmap_bh); 3984 } 3985 3986 BUFFER_TRACE(bitmap_bh, "getting write access"); 3987 err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 3988 EXT4_JTR_NONE); 3989 if (err) 3990 goto out_err; 3991 3992 err = -EIO; 3993 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 3994 if (!gdp) 3995 goto out_err; 3996 3997 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 3998 ext4_free_group_clusters(sb, gdp)); 3999 4000 BUFFER_TRACE(gdp_bh, "get_write_access"); 4001 err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE); 4002 if (err) 4003 goto out_err; 4004 4005 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4006 4007 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 4008 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { 4009 ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 4010 "fs metadata", block, block+len); 4011 /* File system mounted not to panic on error 4012 * Fix the bitmap and return EFSCORRUPTED 4013 * We leak some of the blocks here. 4014 */ 4015 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 4016 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 4017 ac->ac_b_ex.fe_len); 4018 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 4019 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4020 if (!err) 4021 err = -EFSCORRUPTED; 4022 goto out_err; 4023 } 4024 4025 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 4026 #ifdef AGGRESSIVE_CHECK 4027 { 4028 int i; 4029 for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 4030 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 4031 bitmap_bh->b_data)); 4032 } 4033 } 4034 #endif 4035 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 4036 ac->ac_b_ex.fe_len); 4037 if (ext4_has_group_desc_csum(sb) && 4038 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 4039 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 4040 ext4_free_group_clusters_set(sb, gdp, 4041 ext4_free_clusters_after_init(sb, 4042 ac->ac_b_ex.fe_group, gdp)); 4043 } 4044 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; 4045 ext4_free_group_clusters_set(sb, gdp, len); 4046 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 4047 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); 4048 4049 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 4050 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 4051 /* 4052 * Now reduce the dirty block count also. Should not go negative 4053 */ 4054 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 4055 /* release all the reserved blocks if non delalloc */ 4056 percpu_counter_sub(&sbi->s_dirtyclusters_counter, 4057 reserv_clstrs); 4058 4059 if (sbi->s_log_groups_per_flex) { 4060 ext4_group_t flex_group = ext4_flex_group(sbi, 4061 ac->ac_b_ex.fe_group); 4062 atomic64_sub(ac->ac_b_ex.fe_len, 4063 &sbi_array_rcu_deref(sbi, s_flex_groups, 4064 flex_group)->free_clusters); 4065 } 4066 4067 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4068 if (err) 4069 goto out_err; 4070 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 4071 4072 out_err: 4073 brelse(bitmap_bh); 4074 return err; 4075 } 4076 4077 /* 4078 * Idempotent helper for Ext4 fast commit replay path to set the state of 4079 * blocks in bitmaps and update counters. 4080 */ 4081 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, 4082 int len, int state) 4083 { 4084 struct buffer_head *bitmap_bh = NULL; 4085 struct ext4_group_desc *gdp; 4086 struct buffer_head *gdp_bh; 4087 struct ext4_sb_info *sbi = EXT4_SB(sb); 4088 ext4_group_t group; 4089 ext4_grpblk_t blkoff; 4090 int i, err; 4091 int already; 4092 unsigned int clen, clen_changed, thisgrp_len; 4093 4094 while (len > 0) { 4095 ext4_get_group_no_and_offset(sb, block, &group, &blkoff); 4096 4097 /* 4098 * Check to see if we are freeing blocks across a group 4099 * boundary. 4100 * In case of flex_bg, this can happen that (block, len) may 4101 * span across more than one group. In that case we need to 4102 * get the corresponding group metadata to work with. 4103 * For this we have goto again loop. 4104 */ 4105 thisgrp_len = min_t(unsigned int, (unsigned int)len, 4106 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); 4107 clen = EXT4_NUM_B2C(sbi, thisgrp_len); 4108 4109 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { 4110 ext4_error(sb, "Marking blocks in system zone - " 4111 "Block = %llu, len = %u", 4112 block, thisgrp_len); 4113 bitmap_bh = NULL; 4114 break; 4115 } 4116 4117 bitmap_bh = ext4_read_block_bitmap(sb, group); 4118 if (IS_ERR(bitmap_bh)) { 4119 err = PTR_ERR(bitmap_bh); 4120 bitmap_bh = NULL; 4121 break; 4122 } 4123 4124 err = -EIO; 4125 gdp = ext4_get_group_desc(sb, group, &gdp_bh); 4126 if (!gdp) 4127 break; 4128 4129 ext4_lock_group(sb, group); 4130 already = 0; 4131 for (i = 0; i < clen; i++) 4132 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == 4133 !state) 4134 already++; 4135 4136 clen_changed = clen - already; 4137 if (state) 4138 mb_set_bits(bitmap_bh->b_data, blkoff, clen); 4139 else 4140 mb_clear_bits(bitmap_bh->b_data, blkoff, clen); 4141 if (ext4_has_group_desc_csum(sb) && 4142 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 4143 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 4144 ext4_free_group_clusters_set(sb, gdp, 4145 ext4_free_clusters_after_init(sb, group, gdp)); 4146 } 4147 if (state) 4148 clen = ext4_free_group_clusters(sb, gdp) - clen_changed; 4149 else 4150 clen = ext4_free_group_clusters(sb, gdp) + clen_changed; 4151 4152 ext4_free_group_clusters_set(sb, gdp, clen); 4153 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 4154 ext4_group_desc_csum_set(sb, group, gdp); 4155 4156 ext4_unlock_group(sb, group); 4157 4158 if (sbi->s_log_groups_per_flex) { 4159 ext4_group_t flex_group = ext4_flex_group(sbi, group); 4160 struct flex_groups *fg = sbi_array_rcu_deref(sbi, 4161 s_flex_groups, flex_group); 4162 4163 if (state) 4164 atomic64_sub(clen_changed, &fg->free_clusters); 4165 else 4166 atomic64_add(clen_changed, &fg->free_clusters); 4167 4168 } 4169 4170 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); 4171 if (err) 4172 break; 4173 sync_dirty_buffer(bitmap_bh); 4174 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); 4175 sync_dirty_buffer(gdp_bh); 4176 if (err) 4177 break; 4178 4179 block += thisgrp_len; 4180 len -= thisgrp_len; 4181 brelse(bitmap_bh); 4182 BUG_ON(len < 0); 4183 } 4184 4185 if (err) 4186 brelse(bitmap_bh); 4187 } 4188 4189 /* 4190 * here we normalize request for locality group 4191 * Group request are normalized to s_mb_group_prealloc, which goes to 4192 * s_strip if we set the same via mount option. 4193 * s_mb_group_prealloc can be configured via 4194 * /sys/fs/ext4/<partition>/mb_group_prealloc 4195 * 4196 * XXX: should we try to preallocate more than the group has now? 4197 */ 4198 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 4199 { 4200 struct super_block *sb = ac->ac_sb; 4201 struct ext4_locality_group *lg = ac->ac_lg; 4202 4203 BUG_ON(lg == NULL); 4204 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 4205 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); 4206 } 4207 4208 /* 4209 * This function returns the next element to look at during inode 4210 * PA rbtree walk. We assume that we have held the inode PA rbtree lock 4211 * (ei->i_prealloc_lock) 4212 * 4213 * new_start The start of the range we want to compare 4214 * cur_start The existing start that we are comparing against 4215 * node The node of the rb_tree 4216 */ 4217 static inline struct rb_node* 4218 ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node) 4219 { 4220 if (new_start < cur_start) 4221 return node->rb_left; 4222 else 4223 return node->rb_right; 4224 } 4225 4226 static inline void 4227 ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac, 4228 ext4_lblk_t start, loff_t end) 4229 { 4230 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4231 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 4232 struct ext4_prealloc_space *tmp_pa; 4233 ext4_lblk_t tmp_pa_start; 4234 loff_t tmp_pa_end; 4235 struct rb_node *iter; 4236 4237 read_lock(&ei->i_prealloc_lock); 4238 for (iter = ei->i_prealloc_node.rb_node; iter; 4239 iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) { 4240 tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 4241 pa_node.inode_node); 4242 tmp_pa_start = tmp_pa->pa_lstart; 4243 tmp_pa_end = pa_logical_end(sbi, tmp_pa); 4244 4245 spin_lock(&tmp_pa->pa_lock); 4246 if (tmp_pa->pa_deleted == 0) 4247 BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start)); 4248 spin_unlock(&tmp_pa->pa_lock); 4249 } 4250 read_unlock(&ei->i_prealloc_lock); 4251 } 4252 4253 /* 4254 * Given an allocation context "ac" and a range "start", "end", check 4255 * and adjust boundaries if the range overlaps with any of the existing 4256 * preallocatoins stored in the corresponding inode of the allocation context. 4257 * 4258 * Parameters: 4259 * ac allocation context 4260 * start start of the new range 4261 * end end of the new range 4262 */ 4263 static inline void 4264 ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac, 4265 ext4_lblk_t *start, loff_t *end) 4266 { 4267 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 4268 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4269 struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL; 4270 struct rb_node *iter; 4271 ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1; 4272 loff_t new_end, tmp_pa_end, left_pa_end = -1; 4273 4274 new_start = *start; 4275 new_end = *end; 4276 4277 /* 4278 * Adjust the normalized range so that it doesn't overlap with any 4279 * existing preallocated blocks(PAs). Make sure to hold the rbtree lock 4280 * so it doesn't change underneath us. 4281 */ 4282 read_lock(&ei->i_prealloc_lock); 4283 4284 /* Step 1: find any one immediate neighboring PA of the normalized range */ 4285 for (iter = ei->i_prealloc_node.rb_node; iter; 4286 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, 4287 tmp_pa_start, iter)) { 4288 tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 4289 pa_node.inode_node); 4290 tmp_pa_start = tmp_pa->pa_lstart; 4291 tmp_pa_end = pa_logical_end(sbi, tmp_pa); 4292 4293 /* PA must not overlap original request */ 4294 spin_lock(&tmp_pa->pa_lock); 4295 if (tmp_pa->pa_deleted == 0) 4296 BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end || 4297 ac->ac_o_ex.fe_logical < tmp_pa_start)); 4298 spin_unlock(&tmp_pa->pa_lock); 4299 } 4300 4301 /* 4302 * Step 2: check if the found PA is left or right neighbor and 4303 * get the other neighbor 4304 */ 4305 if (tmp_pa) { 4306 if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) { 4307 struct rb_node *tmp; 4308 4309 left_pa = tmp_pa; 4310 tmp = rb_next(&left_pa->pa_node.inode_node); 4311 if (tmp) { 4312 right_pa = rb_entry(tmp, 4313 struct ext4_prealloc_space, 4314 pa_node.inode_node); 4315 } 4316 } else { 4317 struct rb_node *tmp; 4318 4319 right_pa = tmp_pa; 4320 tmp = rb_prev(&right_pa->pa_node.inode_node); 4321 if (tmp) { 4322 left_pa = rb_entry(tmp, 4323 struct ext4_prealloc_space, 4324 pa_node.inode_node); 4325 } 4326 } 4327 } 4328 4329 /* Step 3: get the non deleted neighbors */ 4330 if (left_pa) { 4331 for (iter = &left_pa->pa_node.inode_node;; 4332 iter = rb_prev(iter)) { 4333 if (!iter) { 4334 left_pa = NULL; 4335 break; 4336 } 4337 4338 tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 4339 pa_node.inode_node); 4340 left_pa = tmp_pa; 4341 spin_lock(&tmp_pa->pa_lock); 4342 if (tmp_pa->pa_deleted == 0) { 4343 spin_unlock(&tmp_pa->pa_lock); 4344 break; 4345 } 4346 spin_unlock(&tmp_pa->pa_lock); 4347 } 4348 } 4349 4350 if (right_pa) { 4351 for (iter = &right_pa->pa_node.inode_node;; 4352 iter = rb_next(iter)) { 4353 if (!iter) { 4354 right_pa = NULL; 4355 break; 4356 } 4357 4358 tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 4359 pa_node.inode_node); 4360 right_pa = tmp_pa; 4361 spin_lock(&tmp_pa->pa_lock); 4362 if (tmp_pa->pa_deleted == 0) { 4363 spin_unlock(&tmp_pa->pa_lock); 4364 break; 4365 } 4366 spin_unlock(&tmp_pa->pa_lock); 4367 } 4368 } 4369 4370 if (left_pa) { 4371 left_pa_end = pa_logical_end(sbi, left_pa); 4372 BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical); 4373 } 4374 4375 if (right_pa) { 4376 right_pa_start = right_pa->pa_lstart; 4377 BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical); 4378 } 4379 4380 /* Step 4: trim our normalized range to not overlap with the neighbors */ 4381 if (left_pa) { 4382 if (left_pa_end > new_start) 4383 new_start = left_pa_end; 4384 } 4385 4386 if (right_pa) { 4387 if (right_pa_start < new_end) 4388 new_end = right_pa_start; 4389 } 4390 read_unlock(&ei->i_prealloc_lock); 4391 4392 /* XXX: extra loop to check we really don't overlap preallocations */ 4393 ext4_mb_pa_assert_overlap(ac, new_start, new_end); 4394 4395 *start = new_start; 4396 *end = new_end; 4397 } 4398 4399 /* 4400 * Normalization means making request better in terms of 4401 * size and alignment 4402 */ 4403 static noinline_for_stack void 4404 ext4_mb_normalize_request(struct ext4_allocation_context *ac, 4405 struct ext4_allocation_request *ar) 4406 { 4407 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4408 struct ext4_super_block *es = sbi->s_es; 4409 int bsbits, max; 4410 loff_t size, start_off, end; 4411 loff_t orig_size __maybe_unused; 4412 ext4_lblk_t start; 4413 4414 /* do normalize only data requests, metadata requests 4415 do not need preallocation */ 4416 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4417 return; 4418 4419 /* sometime caller may want exact blocks */ 4420 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 4421 return; 4422 4423 /* caller may indicate that preallocation isn't 4424 * required (it's a tail, for example) */ 4425 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 4426 return; 4427 4428 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 4429 ext4_mb_normalize_group_request(ac); 4430 return ; 4431 } 4432 4433 bsbits = ac->ac_sb->s_blocksize_bits; 4434 4435 /* first, let's learn actual file size 4436 * given current request is allocated */ 4437 size = extent_logical_end(sbi, &ac->ac_o_ex); 4438 size = size << bsbits; 4439 if (size < i_size_read(ac->ac_inode)) 4440 size = i_size_read(ac->ac_inode); 4441 orig_size = size; 4442 4443 /* max size of free chunks */ 4444 max = 2 << bsbits; 4445 4446 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 4447 (req <= (size) || max <= (chunk_size)) 4448 4449 /* first, try to predict filesize */ 4450 /* XXX: should this table be tunable? */ 4451 start_off = 0; 4452 if (size <= 16 * 1024) { 4453 size = 16 * 1024; 4454 } else if (size <= 32 * 1024) { 4455 size = 32 * 1024; 4456 } else if (size <= 64 * 1024) { 4457 size = 64 * 1024; 4458 } else if (size <= 128 * 1024) { 4459 size = 128 * 1024; 4460 } else if (size <= 256 * 1024) { 4461 size = 256 * 1024; 4462 } else if (size <= 512 * 1024) { 4463 size = 512 * 1024; 4464 } else if (size <= 1024 * 1024) { 4465 size = 1024 * 1024; 4466 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 4467 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4468 (21 - bsbits)) << 21; 4469 size = 2 * 1024 * 1024; 4470 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 4471 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4472 (22 - bsbits)) << 22; 4473 size = 4 * 1024 * 1024; 4474 } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len), 4475 (8<<20)>>bsbits, max, 8 * 1024)) { 4476 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4477 (23 - bsbits)) << 23; 4478 size = 8 * 1024 * 1024; 4479 } else { 4480 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; 4481 size = (loff_t) EXT4_C2B(sbi, 4482 ac->ac_o_ex.fe_len) << bsbits; 4483 } 4484 size = size >> bsbits; 4485 start = start_off >> bsbits; 4486 4487 /* 4488 * For tiny groups (smaller than 8MB) the chosen allocation 4489 * alignment may be larger than group size. Make sure the 4490 * alignment does not move allocation to a different group which 4491 * makes mballoc fail assertions later. 4492 */ 4493 start = max(start, rounddown(ac->ac_o_ex.fe_logical, 4494 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); 4495 4496 /* don't cover already allocated blocks in selected range */ 4497 if (ar->pleft && start <= ar->lleft) { 4498 size -= ar->lleft + 1 - start; 4499 start = ar->lleft + 1; 4500 } 4501 if (ar->pright && start + size - 1 >= ar->lright) 4502 size -= start + size - ar->lright; 4503 4504 /* 4505 * Trim allocation request for filesystems with artificially small 4506 * groups. 4507 */ 4508 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) 4509 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); 4510 4511 end = start + size; 4512 4513 ext4_mb_pa_adjust_overlap(ac, &start, &end); 4514 4515 size = end - start; 4516 4517 /* 4518 * In this function "start" and "size" are normalized for better 4519 * alignment and length such that we could preallocate more blocks. 4520 * This normalization is done such that original request of 4521 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and 4522 * "size" boundaries. 4523 * (Note fe_len can be relaxed since FS block allocation API does not 4524 * provide gurantee on number of contiguous blocks allocation since that 4525 * depends upon free space left, etc). 4526 * In case of inode pa, later we use the allocated blocks 4527 * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated 4528 * range of goal/best blocks [start, size] to put it at the 4529 * ac_o_ex.fe_logical extent of this inode. 4530 * (See ext4_mb_use_inode_pa() for more details) 4531 */ 4532 if (start + size <= ac->ac_o_ex.fe_logical || 4533 start > ac->ac_o_ex.fe_logical) { 4534 ext4_msg(ac->ac_sb, KERN_ERR, 4535 "start %lu, size %lu, fe_logical %lu", 4536 (unsigned long) start, (unsigned long) size, 4537 (unsigned long) ac->ac_o_ex.fe_logical); 4538 BUG(); 4539 } 4540 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 4541 4542 /* now prepare goal request */ 4543 4544 /* XXX: is it better to align blocks WRT to logical 4545 * placement or satisfy big request as is */ 4546 ac->ac_g_ex.fe_logical = start; 4547 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 4548 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; 4549 4550 /* define goal start in order to merge */ 4551 if (ar->pright && (ar->lright == (start + size)) && 4552 ar->pright >= size && 4553 ar->pright - size >= le32_to_cpu(es->s_first_data_block)) { 4554 /* merge to the right */ 4555 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 4556 &ac->ac_g_ex.fe_group, 4557 &ac->ac_g_ex.fe_start); 4558 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 4559 } 4560 if (ar->pleft && (ar->lleft + 1 == start) && 4561 ar->pleft + 1 < ext4_blocks_count(es)) { 4562 /* merge to the left */ 4563 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 4564 &ac->ac_g_ex.fe_group, 4565 &ac->ac_g_ex.fe_start); 4566 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 4567 } 4568 4569 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, 4570 orig_size, start); 4571 } 4572 4573 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 4574 { 4575 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4576 4577 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { 4578 atomic_inc(&sbi->s_bal_reqs); 4579 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 4580 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 4581 atomic_inc(&sbi->s_bal_success); 4582 4583 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 4584 for (int i=0; i<EXT4_MB_NUM_CRS; i++) { 4585 atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]); 4586 } 4587 4588 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); 4589 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 4590 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 4591 atomic_inc(&sbi->s_bal_goals); 4592 /* did we allocate as much as normalizer originally wanted? */ 4593 if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len) 4594 atomic_inc(&sbi->s_bal_len_goals); 4595 4596 if (ac->ac_found > sbi->s_mb_max_to_scan) 4597 atomic_inc(&sbi->s_bal_breaks); 4598 } 4599 4600 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 4601 trace_ext4_mballoc_alloc(ac); 4602 else 4603 trace_ext4_mballoc_prealloc(ac); 4604 } 4605 4606 /* 4607 * Called on failure; free up any blocks from the inode PA for this 4608 * context. We don't need this for MB_GROUP_PA because we only change 4609 * pa_free in ext4_mb_release_context(), but on failure, we've already 4610 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 4611 */ 4612 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 4613 { 4614 struct ext4_prealloc_space *pa = ac->ac_pa; 4615 struct ext4_buddy e4b; 4616 int err; 4617 4618 if (pa == NULL) { 4619 if (ac->ac_f_ex.fe_len == 0) 4620 return; 4621 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); 4622 if (WARN_RATELIMIT(err, 4623 "ext4: mb_load_buddy failed (%d)", err)) 4624 /* 4625 * This should never happen since we pin the 4626 * pages in the ext4_allocation_context so 4627 * ext4_mb_load_buddy() should never fail. 4628 */ 4629 return; 4630 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 4631 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, 4632 ac->ac_f_ex.fe_len); 4633 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 4634 ext4_mb_unload_buddy(&e4b); 4635 return; 4636 } 4637 if (pa->pa_type == MB_INODE_PA) { 4638 spin_lock(&pa->pa_lock); 4639 pa->pa_free += ac->ac_b_ex.fe_len; 4640 spin_unlock(&pa->pa_lock); 4641 } 4642 } 4643 4644 /* 4645 * use blocks preallocated to inode 4646 */ 4647 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 4648 struct ext4_prealloc_space *pa) 4649 { 4650 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4651 ext4_fsblk_t start; 4652 ext4_fsblk_t end; 4653 int len; 4654 4655 /* found preallocated blocks, use them */ 4656 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 4657 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 4658 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 4659 len = EXT4_NUM_B2C(sbi, end - start); 4660 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 4661 &ac->ac_b_ex.fe_start); 4662 ac->ac_b_ex.fe_len = len; 4663 ac->ac_status = AC_STATUS_FOUND; 4664 ac->ac_pa = pa; 4665 4666 BUG_ON(start < pa->pa_pstart); 4667 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 4668 BUG_ON(pa->pa_free < len); 4669 BUG_ON(ac->ac_b_ex.fe_len <= 0); 4670 pa->pa_free -= len; 4671 4672 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); 4673 } 4674 4675 /* 4676 * use blocks preallocated to locality group 4677 */ 4678 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 4679 struct ext4_prealloc_space *pa) 4680 { 4681 unsigned int len = ac->ac_o_ex.fe_len; 4682 4683 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 4684 &ac->ac_b_ex.fe_group, 4685 &ac->ac_b_ex.fe_start); 4686 ac->ac_b_ex.fe_len = len; 4687 ac->ac_status = AC_STATUS_FOUND; 4688 ac->ac_pa = pa; 4689 4690 /* we don't correct pa_pstart or pa_len here to avoid 4691 * possible race when the group is being loaded concurrently 4692 * instead we correct pa later, after blocks are marked 4693 * in on-disk bitmap -- see ext4_mb_release_context() 4694 * Other CPUs are prevented from allocating from this pa by lg_mutex 4695 */ 4696 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", 4697 pa->pa_lstart, len, pa); 4698 } 4699 4700 /* 4701 * Return the prealloc space that have minimal distance 4702 * from the goal block. @cpa is the prealloc 4703 * space that is having currently known minimal distance 4704 * from the goal block. 4705 */ 4706 static struct ext4_prealloc_space * 4707 ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 4708 struct ext4_prealloc_space *pa, 4709 struct ext4_prealloc_space *cpa) 4710 { 4711 ext4_fsblk_t cur_distance, new_distance; 4712 4713 if (cpa == NULL) { 4714 atomic_inc(&pa->pa_count); 4715 return pa; 4716 } 4717 cur_distance = abs(goal_block - cpa->pa_pstart); 4718 new_distance = abs(goal_block - pa->pa_pstart); 4719 4720 if (cur_distance <= new_distance) 4721 return cpa; 4722 4723 /* drop the previous reference */ 4724 atomic_dec(&cpa->pa_count); 4725 atomic_inc(&pa->pa_count); 4726 return pa; 4727 } 4728 4729 /* 4730 * check if found pa meets EXT4_MB_HINT_GOAL_ONLY 4731 */ 4732 static bool 4733 ext4_mb_pa_goal_check(struct ext4_allocation_context *ac, 4734 struct ext4_prealloc_space *pa) 4735 { 4736 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4737 ext4_fsblk_t start; 4738 4739 if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))) 4740 return true; 4741 4742 /* 4743 * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted 4744 * in ext4_mb_normalize_request and will keep same with ac_o_ex 4745 * from ext4_mb_initialize_context. Choose ac_g_ex here to keep 4746 * consistent with ext4_mb_find_by_goal. 4747 */ 4748 start = pa->pa_pstart + 4749 (ac->ac_g_ex.fe_logical - pa->pa_lstart); 4750 if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start) 4751 return false; 4752 4753 if (ac->ac_g_ex.fe_len > pa->pa_len - 4754 EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart)) 4755 return false; 4756 4757 return true; 4758 } 4759 4760 /* 4761 * search goal blocks in preallocated space 4762 */ 4763 static noinline_for_stack bool 4764 ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 4765 { 4766 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4767 int order, i; 4768 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 4769 struct ext4_locality_group *lg; 4770 struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL; 4771 struct rb_node *iter; 4772 ext4_fsblk_t goal_block; 4773 4774 /* only data can be preallocated */ 4775 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4776 return false; 4777 4778 /* 4779 * first, try per-file preallocation by searching the inode pa rbtree. 4780 * 4781 * Here, we can't do a direct traversal of the tree because 4782 * ext4_mb_discard_group_preallocation() can paralelly mark the pa 4783 * deleted and that can cause direct traversal to skip some entries. 4784 */ 4785 read_lock(&ei->i_prealloc_lock); 4786 4787 if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) { 4788 goto try_group_pa; 4789 } 4790 4791 /* 4792 * Step 1: Find a pa with logical start immediately adjacent to the 4793 * original logical start. This could be on the left or right. 4794 * 4795 * (tmp_pa->pa_lstart never changes so we can skip locking for it). 4796 */ 4797 for (iter = ei->i_prealloc_node.rb_node; iter; 4798 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, 4799 tmp_pa->pa_lstart, iter)) { 4800 tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 4801 pa_node.inode_node); 4802 } 4803 4804 /* 4805 * Step 2: The adjacent pa might be to the right of logical start, find 4806 * the left adjacent pa. After this step we'd have a valid tmp_pa whose 4807 * logical start is towards the left of original request's logical start 4808 */ 4809 if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) { 4810 struct rb_node *tmp; 4811 tmp = rb_prev(&tmp_pa->pa_node.inode_node); 4812 4813 if (tmp) { 4814 tmp_pa = rb_entry(tmp, struct ext4_prealloc_space, 4815 pa_node.inode_node); 4816 } else { 4817 /* 4818 * If there is no adjacent pa to the left then finding 4819 * an overlapping pa is not possible hence stop searching 4820 * inode pa tree 4821 */ 4822 goto try_group_pa; 4823 } 4824 } 4825 4826 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); 4827 4828 /* 4829 * Step 3: If the left adjacent pa is deleted, keep moving left to find 4830 * the first non deleted adjacent pa. After this step we should have a 4831 * valid tmp_pa which is guaranteed to be non deleted. 4832 */ 4833 for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) { 4834 if (!iter) { 4835 /* 4836 * no non deleted left adjacent pa, so stop searching 4837 * inode pa tree 4838 */ 4839 goto try_group_pa; 4840 } 4841 tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 4842 pa_node.inode_node); 4843 spin_lock(&tmp_pa->pa_lock); 4844 if (tmp_pa->pa_deleted == 0) { 4845 /* 4846 * We will keep holding the pa_lock from 4847 * this point on because we don't want group discard 4848 * to delete this pa underneath us. Since group 4849 * discard is anyways an ENOSPC operation it 4850 * should be okay for it to wait a few more cycles. 4851 */ 4852 break; 4853 } else { 4854 spin_unlock(&tmp_pa->pa_lock); 4855 } 4856 } 4857 4858 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); 4859 BUG_ON(tmp_pa->pa_deleted == 1); 4860 4861 /* 4862 * Step 4: We now have the non deleted left adjacent pa. Only this 4863 * pa can possibly satisfy the request hence check if it overlaps 4864 * original logical start and stop searching if it doesn't. 4865 */ 4866 if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) { 4867 spin_unlock(&tmp_pa->pa_lock); 4868 goto try_group_pa; 4869 } 4870 4871 /* non-extent files can't have physical blocks past 2^32 */ 4872 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 4873 (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) > 4874 EXT4_MAX_BLOCK_FILE_PHYS)) { 4875 /* 4876 * Since PAs don't overlap, we won't find any other PA to 4877 * satisfy this. 4878 */ 4879 spin_unlock(&tmp_pa->pa_lock); 4880 goto try_group_pa; 4881 } 4882 4883 if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) { 4884 atomic_inc(&tmp_pa->pa_count); 4885 ext4_mb_use_inode_pa(ac, tmp_pa); 4886 spin_unlock(&tmp_pa->pa_lock); 4887 read_unlock(&ei->i_prealloc_lock); 4888 return true; 4889 } else { 4890 /* 4891 * We found a valid overlapping pa but couldn't use it because 4892 * it had no free blocks. This should ideally never happen 4893 * because: 4894 * 4895 * 1. When a new inode pa is added to rbtree it must have 4896 * pa_free > 0 since otherwise we won't actually need 4897 * preallocation. 4898 * 4899 * 2. An inode pa that is in the rbtree can only have it's 4900 * pa_free become zero when another thread calls: 4901 * ext4_mb_new_blocks 4902 * ext4_mb_use_preallocated 4903 * ext4_mb_use_inode_pa 4904 * 4905 * 3. Further, after the above calls make pa_free == 0, we will 4906 * immediately remove it from the rbtree in: 4907 * ext4_mb_new_blocks 4908 * ext4_mb_release_context 4909 * ext4_mb_put_pa 4910 * 4911 * 4. Since the pa_free becoming 0 and pa_free getting removed 4912 * from tree both happen in ext4_mb_new_blocks, which is always 4913 * called with i_data_sem held for data allocations, we can be 4914 * sure that another process will never see a pa in rbtree with 4915 * pa_free == 0. 4916 */ 4917 WARN_ON_ONCE(tmp_pa->pa_free == 0); 4918 } 4919 spin_unlock(&tmp_pa->pa_lock); 4920 try_group_pa: 4921 read_unlock(&ei->i_prealloc_lock); 4922 4923 /* can we use group allocation? */ 4924 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 4925 return false; 4926 4927 /* inode may have no locality group for some reason */ 4928 lg = ac->ac_lg; 4929 if (lg == NULL) 4930 return false; 4931 order = fls(ac->ac_o_ex.fe_len) - 1; 4932 if (order > PREALLOC_TB_SIZE - 1) 4933 /* The max size of hash table is PREALLOC_TB_SIZE */ 4934 order = PREALLOC_TB_SIZE - 1; 4935 4936 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 4937 /* 4938 * search for the prealloc space that is having 4939 * minimal distance from the goal block. 4940 */ 4941 for (i = order; i < PREALLOC_TB_SIZE; i++) { 4942 rcu_read_lock(); 4943 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i], 4944 pa_node.lg_list) { 4945 spin_lock(&tmp_pa->pa_lock); 4946 if (tmp_pa->pa_deleted == 0 && 4947 tmp_pa->pa_free >= ac->ac_o_ex.fe_len) { 4948 4949 cpa = ext4_mb_check_group_pa(goal_block, 4950 tmp_pa, cpa); 4951 } 4952 spin_unlock(&tmp_pa->pa_lock); 4953 } 4954 rcu_read_unlock(); 4955 } 4956 if (cpa) { 4957 ext4_mb_use_group_pa(ac, cpa); 4958 return true; 4959 } 4960 return false; 4961 } 4962 4963 /* 4964 * the function goes through all block freed in the group 4965 * but not yet committed and marks them used in in-core bitmap. 4966 * buddy must be generated from this bitmap 4967 * Need to be called with the ext4 group lock held 4968 */ 4969 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 4970 ext4_group_t group) 4971 { 4972 struct rb_node *n; 4973 struct ext4_group_info *grp; 4974 struct ext4_free_data *entry; 4975 4976 grp = ext4_get_group_info(sb, group); 4977 if (!grp) 4978 return; 4979 n = rb_first(&(grp->bb_free_root)); 4980 4981 while (n) { 4982 entry = rb_entry(n, struct ext4_free_data, efd_node); 4983 mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); 4984 n = rb_next(n); 4985 } 4986 } 4987 4988 /* 4989 * the function goes through all preallocation in this group and marks them 4990 * used in in-core bitmap. buddy must be generated from this bitmap 4991 * Need to be called with ext4 group lock held 4992 */ 4993 static noinline_for_stack 4994 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 4995 ext4_group_t group) 4996 { 4997 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 4998 struct ext4_prealloc_space *pa; 4999 struct list_head *cur; 5000 ext4_group_t groupnr; 5001 ext4_grpblk_t start; 5002 int preallocated = 0; 5003 int len; 5004 5005 if (!grp) 5006 return; 5007 5008 /* all form of preallocation discards first load group, 5009 * so the only competing code is preallocation use. 5010 * we don't need any locking here 5011 * notice we do NOT ignore preallocations with pa_deleted 5012 * otherwise we could leave used blocks available for 5013 * allocation in buddy when concurrent ext4_mb_put_pa() 5014 * is dropping preallocation 5015 */ 5016 list_for_each(cur, &grp->bb_prealloc_list) { 5017 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 5018 spin_lock(&pa->pa_lock); 5019 ext4_get_group_no_and_offset(sb, pa->pa_pstart, 5020 &groupnr, &start); 5021 len = pa->pa_len; 5022 spin_unlock(&pa->pa_lock); 5023 if (unlikely(len == 0)) 5024 continue; 5025 BUG_ON(groupnr != group); 5026 mb_set_bits(bitmap, start, len); 5027 preallocated += len; 5028 } 5029 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); 5030 } 5031 5032 static void ext4_mb_mark_pa_deleted(struct super_block *sb, 5033 struct ext4_prealloc_space *pa) 5034 { 5035 struct ext4_inode_info *ei; 5036 5037 if (pa->pa_deleted) { 5038 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", 5039 pa->pa_type, pa->pa_pstart, pa->pa_lstart, 5040 pa->pa_len); 5041 return; 5042 } 5043 5044 pa->pa_deleted = 1; 5045 5046 if (pa->pa_type == MB_INODE_PA) { 5047 ei = EXT4_I(pa->pa_inode); 5048 atomic_dec(&ei->i_prealloc_active); 5049 } 5050 } 5051 5052 static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa) 5053 { 5054 BUG_ON(!pa); 5055 BUG_ON(atomic_read(&pa->pa_count)); 5056 BUG_ON(pa->pa_deleted == 0); 5057 kmem_cache_free(ext4_pspace_cachep, pa); 5058 } 5059 5060 static void ext4_mb_pa_callback(struct rcu_head *head) 5061 { 5062 struct ext4_prealloc_space *pa; 5063 5064 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 5065 ext4_mb_pa_free(pa); 5066 } 5067 5068 /* 5069 * drops a reference to preallocated space descriptor 5070 * if this was the last reference and the space is consumed 5071 */ 5072 static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 5073 struct super_block *sb, struct ext4_prealloc_space *pa) 5074 { 5075 ext4_group_t grp; 5076 ext4_fsblk_t grp_blk; 5077 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 5078 5079 /* in this short window concurrent discard can set pa_deleted */ 5080 spin_lock(&pa->pa_lock); 5081 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { 5082 spin_unlock(&pa->pa_lock); 5083 return; 5084 } 5085 5086 if (pa->pa_deleted == 1) { 5087 spin_unlock(&pa->pa_lock); 5088 return; 5089 } 5090 5091 ext4_mb_mark_pa_deleted(sb, pa); 5092 spin_unlock(&pa->pa_lock); 5093 5094 grp_blk = pa->pa_pstart; 5095 /* 5096 * If doing group-based preallocation, pa_pstart may be in the 5097 * next group when pa is used up 5098 */ 5099 if (pa->pa_type == MB_GROUP_PA) 5100 grp_blk--; 5101 5102 grp = ext4_get_group_number(sb, grp_blk); 5103 5104 /* 5105 * possible race: 5106 * 5107 * P1 (buddy init) P2 (regular allocation) 5108 * find block B in PA 5109 * copy on-disk bitmap to buddy 5110 * mark B in on-disk bitmap 5111 * drop PA from group 5112 * mark all PAs in buddy 5113 * 5114 * thus, P1 initializes buddy with B available. to prevent this 5115 * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 5116 * against that pair 5117 */ 5118 ext4_lock_group(sb, grp); 5119 list_del(&pa->pa_group_list); 5120 ext4_unlock_group(sb, grp); 5121 5122 if (pa->pa_type == MB_INODE_PA) { 5123 write_lock(pa->pa_node_lock.inode_lock); 5124 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 5125 write_unlock(pa->pa_node_lock.inode_lock); 5126 ext4_mb_pa_free(pa); 5127 } else { 5128 spin_lock(pa->pa_node_lock.lg_lock); 5129 list_del_rcu(&pa->pa_node.lg_list); 5130 spin_unlock(pa->pa_node_lock.lg_lock); 5131 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 5132 } 5133 } 5134 5135 static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new) 5136 { 5137 struct rb_node **iter = &root->rb_node, *parent = NULL; 5138 struct ext4_prealloc_space *iter_pa, *new_pa; 5139 ext4_lblk_t iter_start, new_start; 5140 5141 while (*iter) { 5142 iter_pa = rb_entry(*iter, struct ext4_prealloc_space, 5143 pa_node.inode_node); 5144 new_pa = rb_entry(new, struct ext4_prealloc_space, 5145 pa_node.inode_node); 5146 iter_start = iter_pa->pa_lstart; 5147 new_start = new_pa->pa_lstart; 5148 5149 parent = *iter; 5150 if (new_start < iter_start) 5151 iter = &((*iter)->rb_left); 5152 else 5153 iter = &((*iter)->rb_right); 5154 } 5155 5156 rb_link_node(new, parent, iter); 5157 rb_insert_color(new, root); 5158 } 5159 5160 /* 5161 * creates new preallocated space for given inode 5162 */ 5163 static noinline_for_stack void 5164 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 5165 { 5166 struct super_block *sb = ac->ac_sb; 5167 struct ext4_sb_info *sbi = EXT4_SB(sb); 5168 struct ext4_prealloc_space *pa; 5169 struct ext4_group_info *grp; 5170 struct ext4_inode_info *ei; 5171 5172 /* preallocate only when found space is larger then requested */ 5173 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 5174 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 5175 BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 5176 BUG_ON(ac->ac_pa == NULL); 5177 5178 pa = ac->ac_pa; 5179 5180 if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) { 5181 struct ext4_free_extent ex = { 5182 .fe_logical = ac->ac_g_ex.fe_logical, 5183 .fe_len = ac->ac_orig_goal_len, 5184 }; 5185 loff_t orig_goal_end = extent_logical_end(sbi, &ex); 5186 5187 /* we can't allocate as much as normalizer wants. 5188 * so, found space must get proper lstart 5189 * to cover original request */ 5190 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 5191 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 5192 5193 /* 5194 * Use the below logic for adjusting best extent as it keeps 5195 * fragmentation in check while ensuring logical range of best 5196 * extent doesn't overflow out of goal extent: 5197 * 5198 * 1. Check if best ex can be kept at end of goal (before 5199 * cr_best_avail trimmed it) and still cover original start 5200 * 2. Else, check if best ex can be kept at start of goal and 5201 * still cover original start 5202 * 3. Else, keep the best ex at start of original request. 5203 */ 5204 ex.fe_len = ac->ac_b_ex.fe_len; 5205 5206 ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len); 5207 if (ac->ac_o_ex.fe_logical >= ex.fe_logical) 5208 goto adjust_bex; 5209 5210 ex.fe_logical = ac->ac_g_ex.fe_logical; 5211 if (ac->ac_o_ex.fe_logical < extent_logical_end(sbi, &ex)) 5212 goto adjust_bex; 5213 5214 ex.fe_logical = ac->ac_o_ex.fe_logical; 5215 adjust_bex: 5216 ac->ac_b_ex.fe_logical = ex.fe_logical; 5217 5218 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 5219 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 5220 BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end); 5221 } 5222 5223 pa->pa_lstart = ac->ac_b_ex.fe_logical; 5224 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 5225 pa->pa_len = ac->ac_b_ex.fe_len; 5226 pa->pa_free = pa->pa_len; 5227 spin_lock_init(&pa->pa_lock); 5228 INIT_LIST_HEAD(&pa->pa_group_list); 5229 pa->pa_deleted = 0; 5230 pa->pa_type = MB_INODE_PA; 5231 5232 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, 5233 pa->pa_len, pa->pa_lstart); 5234 trace_ext4_mb_new_inode_pa(ac, pa); 5235 5236 atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 5237 ext4_mb_use_inode_pa(ac, pa); 5238 5239 ei = EXT4_I(ac->ac_inode); 5240 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 5241 if (!grp) 5242 return; 5243 5244 pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock; 5245 pa->pa_inode = ac->ac_inode; 5246 5247 list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 5248 5249 write_lock(pa->pa_node_lock.inode_lock); 5250 ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node); 5251 write_unlock(pa->pa_node_lock.inode_lock); 5252 atomic_inc(&ei->i_prealloc_active); 5253 } 5254 5255 /* 5256 * creates new preallocated space for locality group inodes belongs to 5257 */ 5258 static noinline_for_stack void 5259 ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 5260 { 5261 struct super_block *sb = ac->ac_sb; 5262 struct ext4_locality_group *lg; 5263 struct ext4_prealloc_space *pa; 5264 struct ext4_group_info *grp; 5265 5266 /* preallocate only when found space is larger then requested */ 5267 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 5268 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 5269 BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 5270 BUG_ON(ac->ac_pa == NULL); 5271 5272 pa = ac->ac_pa; 5273 5274 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 5275 pa->pa_lstart = pa->pa_pstart; 5276 pa->pa_len = ac->ac_b_ex.fe_len; 5277 pa->pa_free = pa->pa_len; 5278 spin_lock_init(&pa->pa_lock); 5279 INIT_LIST_HEAD(&pa->pa_node.lg_list); 5280 INIT_LIST_HEAD(&pa->pa_group_list); 5281 pa->pa_deleted = 0; 5282 pa->pa_type = MB_GROUP_PA; 5283 5284 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, 5285 pa->pa_len, pa->pa_lstart); 5286 trace_ext4_mb_new_group_pa(ac, pa); 5287 5288 ext4_mb_use_group_pa(ac, pa); 5289 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 5290 5291 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 5292 if (!grp) 5293 return; 5294 lg = ac->ac_lg; 5295 BUG_ON(lg == NULL); 5296 5297 pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock; 5298 pa->pa_inode = NULL; 5299 5300 list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 5301 5302 /* 5303 * We will later add the new pa to the right bucket 5304 * after updating the pa_free in ext4_mb_release_context 5305 */ 5306 } 5307 5308 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 5309 { 5310 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 5311 ext4_mb_new_group_pa(ac); 5312 else 5313 ext4_mb_new_inode_pa(ac); 5314 } 5315 5316 /* 5317 * finds all unused blocks in on-disk bitmap, frees them in 5318 * in-core bitmap and buddy. 5319 * @pa must be unlinked from inode and group lists, so that 5320 * nobody else can find/use it. 5321 * the caller MUST hold group/inode locks. 5322 * TODO: optimize the case when there are no in-core structures yet 5323 */ 5324 static noinline_for_stack int 5325 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 5326 struct ext4_prealloc_space *pa) 5327 { 5328 struct super_block *sb = e4b->bd_sb; 5329 struct ext4_sb_info *sbi = EXT4_SB(sb); 5330 unsigned int end; 5331 unsigned int next; 5332 ext4_group_t group; 5333 ext4_grpblk_t bit; 5334 unsigned long long grp_blk_start; 5335 int free = 0; 5336 5337 BUG_ON(pa->pa_deleted == 0); 5338 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 5339 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 5340 BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 5341 end = bit + pa->pa_len; 5342 5343 while (bit < end) { 5344 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 5345 if (bit >= end) 5346 break; 5347 next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 5348 mb_debug(sb, "free preallocated %u/%u in group %u\n", 5349 (unsigned) ext4_group_first_block_no(sb, group) + bit, 5350 (unsigned) next - bit, (unsigned) group); 5351 free += next - bit; 5352 5353 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 5354 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 5355 EXT4_C2B(sbi, bit)), 5356 next - bit); 5357 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 5358 bit = next + 1; 5359 } 5360 if (free != pa->pa_free) { 5361 ext4_msg(e4b->bd_sb, KERN_CRIT, 5362 "pa %p: logic %lu, phys. %lu, len %d", 5363 pa, (unsigned long) pa->pa_lstart, 5364 (unsigned long) pa->pa_pstart, 5365 pa->pa_len); 5366 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 5367 free, pa->pa_free); 5368 /* 5369 * pa is already deleted so we use the value obtained 5370 * from the bitmap and continue. 5371 */ 5372 } 5373 atomic_add(free, &sbi->s_mb_discarded); 5374 5375 return 0; 5376 } 5377 5378 static noinline_for_stack int 5379 ext4_mb_release_group_pa(struct ext4_buddy *e4b, 5380 struct ext4_prealloc_space *pa) 5381 { 5382 struct super_block *sb = e4b->bd_sb; 5383 ext4_group_t group; 5384 ext4_grpblk_t bit; 5385 5386 trace_ext4_mb_release_group_pa(sb, pa); 5387 BUG_ON(pa->pa_deleted == 0); 5388 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 5389 if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) { 5390 ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu", 5391 e4b->bd_group, group, pa->pa_pstart); 5392 return 0; 5393 } 5394 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 5395 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 5396 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 5397 5398 return 0; 5399 } 5400 5401 /* 5402 * releases all preallocations in given group 5403 * 5404 * first, we need to decide discard policy: 5405 * - when do we discard 5406 * 1) ENOSPC 5407 * - how many do we discard 5408 * 1) how many requested 5409 */ 5410 static noinline_for_stack int 5411 ext4_mb_discard_group_preallocations(struct super_block *sb, 5412 ext4_group_t group, int *busy) 5413 { 5414 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 5415 struct buffer_head *bitmap_bh = NULL; 5416 struct ext4_prealloc_space *pa, *tmp; 5417 struct list_head list; 5418 struct ext4_buddy e4b; 5419 struct ext4_inode_info *ei; 5420 int err; 5421 int free = 0; 5422 5423 if (!grp) 5424 return 0; 5425 mb_debug(sb, "discard preallocation for group %u\n", group); 5426 if (list_empty(&grp->bb_prealloc_list)) 5427 goto out_dbg; 5428 5429 bitmap_bh = ext4_read_block_bitmap(sb, group); 5430 if (IS_ERR(bitmap_bh)) { 5431 err = PTR_ERR(bitmap_bh); 5432 ext4_error_err(sb, -err, 5433 "Error %d reading block bitmap for %u", 5434 err, group); 5435 goto out_dbg; 5436 } 5437 5438 err = ext4_mb_load_buddy(sb, group, &e4b); 5439 if (err) { 5440 ext4_warning(sb, "Error %d loading buddy information for %u", 5441 err, group); 5442 put_bh(bitmap_bh); 5443 goto out_dbg; 5444 } 5445 5446 INIT_LIST_HEAD(&list); 5447 ext4_lock_group(sb, group); 5448 list_for_each_entry_safe(pa, tmp, 5449 &grp->bb_prealloc_list, pa_group_list) { 5450 spin_lock(&pa->pa_lock); 5451 if (atomic_read(&pa->pa_count)) { 5452 spin_unlock(&pa->pa_lock); 5453 *busy = 1; 5454 continue; 5455 } 5456 if (pa->pa_deleted) { 5457 spin_unlock(&pa->pa_lock); 5458 continue; 5459 } 5460 5461 /* seems this one can be freed ... */ 5462 ext4_mb_mark_pa_deleted(sb, pa); 5463 5464 if (!free) 5465 this_cpu_inc(discard_pa_seq); 5466 5467 /* we can trust pa_free ... */ 5468 free += pa->pa_free; 5469 5470 spin_unlock(&pa->pa_lock); 5471 5472 list_del(&pa->pa_group_list); 5473 list_add(&pa->u.pa_tmp_list, &list); 5474 } 5475 5476 /* now free all selected PAs */ 5477 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 5478 5479 /* remove from object (inode or locality group) */ 5480 if (pa->pa_type == MB_GROUP_PA) { 5481 spin_lock(pa->pa_node_lock.lg_lock); 5482 list_del_rcu(&pa->pa_node.lg_list); 5483 spin_unlock(pa->pa_node_lock.lg_lock); 5484 } else { 5485 write_lock(pa->pa_node_lock.inode_lock); 5486 ei = EXT4_I(pa->pa_inode); 5487 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 5488 write_unlock(pa->pa_node_lock.inode_lock); 5489 } 5490 5491 list_del(&pa->u.pa_tmp_list); 5492 5493 if (pa->pa_type == MB_GROUP_PA) { 5494 ext4_mb_release_group_pa(&e4b, pa); 5495 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 5496 } else { 5497 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 5498 ext4_mb_pa_free(pa); 5499 } 5500 } 5501 5502 ext4_unlock_group(sb, group); 5503 ext4_mb_unload_buddy(&e4b); 5504 put_bh(bitmap_bh); 5505 out_dbg: 5506 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", 5507 free, group, grp->bb_free); 5508 return free; 5509 } 5510 5511 /* 5512 * releases all non-used preallocated blocks for given inode 5513 * 5514 * It's important to discard preallocations under i_data_sem 5515 * We don't want another block to be served from the prealloc 5516 * space when we are discarding the inode prealloc space. 5517 * 5518 * FIXME!! Make sure it is valid at all the call sites 5519 */ 5520 void ext4_discard_preallocations(struct inode *inode, unsigned int needed) 5521 { 5522 struct ext4_inode_info *ei = EXT4_I(inode); 5523 struct super_block *sb = inode->i_sb; 5524 struct buffer_head *bitmap_bh = NULL; 5525 struct ext4_prealloc_space *pa, *tmp; 5526 ext4_group_t group = 0; 5527 struct list_head list; 5528 struct ext4_buddy e4b; 5529 struct rb_node *iter; 5530 int err; 5531 5532 if (!S_ISREG(inode->i_mode)) { 5533 return; 5534 } 5535 5536 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) 5537 return; 5538 5539 mb_debug(sb, "discard preallocation for inode %lu\n", 5540 inode->i_ino); 5541 trace_ext4_discard_preallocations(inode, 5542 atomic_read(&ei->i_prealloc_active), needed); 5543 5544 INIT_LIST_HEAD(&list); 5545 5546 if (needed == 0) 5547 needed = UINT_MAX; 5548 5549 repeat: 5550 /* first, collect all pa's in the inode */ 5551 write_lock(&ei->i_prealloc_lock); 5552 for (iter = rb_first(&ei->i_prealloc_node); iter && needed; 5553 iter = rb_next(iter)) { 5554 pa = rb_entry(iter, struct ext4_prealloc_space, 5555 pa_node.inode_node); 5556 BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock); 5557 5558 spin_lock(&pa->pa_lock); 5559 if (atomic_read(&pa->pa_count)) { 5560 /* this shouldn't happen often - nobody should 5561 * use preallocation while we're discarding it */ 5562 spin_unlock(&pa->pa_lock); 5563 write_unlock(&ei->i_prealloc_lock); 5564 ext4_msg(sb, KERN_ERR, 5565 "uh-oh! used pa while discarding"); 5566 WARN_ON(1); 5567 schedule_timeout_uninterruptible(HZ); 5568 goto repeat; 5569 5570 } 5571 if (pa->pa_deleted == 0) { 5572 ext4_mb_mark_pa_deleted(sb, pa); 5573 spin_unlock(&pa->pa_lock); 5574 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 5575 list_add(&pa->u.pa_tmp_list, &list); 5576 needed--; 5577 continue; 5578 } 5579 5580 /* someone is deleting pa right now */ 5581 spin_unlock(&pa->pa_lock); 5582 write_unlock(&ei->i_prealloc_lock); 5583 5584 /* we have to wait here because pa_deleted 5585 * doesn't mean pa is already unlinked from 5586 * the list. as we might be called from 5587 * ->clear_inode() the inode will get freed 5588 * and concurrent thread which is unlinking 5589 * pa from inode's list may access already 5590 * freed memory, bad-bad-bad */ 5591 5592 /* XXX: if this happens too often, we can 5593 * add a flag to force wait only in case 5594 * of ->clear_inode(), but not in case of 5595 * regular truncate */ 5596 schedule_timeout_uninterruptible(HZ); 5597 goto repeat; 5598 } 5599 write_unlock(&ei->i_prealloc_lock); 5600 5601 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 5602 BUG_ON(pa->pa_type != MB_INODE_PA); 5603 group = ext4_get_group_number(sb, pa->pa_pstart); 5604 5605 err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 5606 GFP_NOFS|__GFP_NOFAIL); 5607 if (err) { 5608 ext4_error_err(sb, -err, "Error %d loading buddy information for %u", 5609 err, group); 5610 continue; 5611 } 5612 5613 bitmap_bh = ext4_read_block_bitmap(sb, group); 5614 if (IS_ERR(bitmap_bh)) { 5615 err = PTR_ERR(bitmap_bh); 5616 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", 5617 err, group); 5618 ext4_mb_unload_buddy(&e4b); 5619 continue; 5620 } 5621 5622 ext4_lock_group(sb, group); 5623 list_del(&pa->pa_group_list); 5624 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 5625 ext4_unlock_group(sb, group); 5626 5627 ext4_mb_unload_buddy(&e4b); 5628 put_bh(bitmap_bh); 5629 5630 list_del(&pa->u.pa_tmp_list); 5631 ext4_mb_pa_free(pa); 5632 } 5633 } 5634 5635 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) 5636 { 5637 struct ext4_prealloc_space *pa; 5638 5639 BUG_ON(ext4_pspace_cachep == NULL); 5640 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); 5641 if (!pa) 5642 return -ENOMEM; 5643 atomic_set(&pa->pa_count, 1); 5644 ac->ac_pa = pa; 5645 return 0; 5646 } 5647 5648 static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac) 5649 { 5650 struct ext4_prealloc_space *pa = ac->ac_pa; 5651 5652 BUG_ON(!pa); 5653 ac->ac_pa = NULL; 5654 WARN_ON(!atomic_dec_and_test(&pa->pa_count)); 5655 /* 5656 * current function is only called due to an error or due to 5657 * len of found blocks < len of requested blocks hence the PA has not 5658 * been added to grp->bb_prealloc_list. So we don't need to lock it 5659 */ 5660 pa->pa_deleted = 1; 5661 ext4_mb_pa_free(pa); 5662 } 5663 5664 #ifdef CONFIG_EXT4_DEBUG 5665 static inline void ext4_mb_show_pa(struct super_block *sb) 5666 { 5667 ext4_group_t i, ngroups; 5668 5669 if (ext4_forced_shutdown(sb)) 5670 return; 5671 5672 ngroups = ext4_get_groups_count(sb); 5673 mb_debug(sb, "groups: "); 5674 for (i = 0; i < ngroups; i++) { 5675 struct ext4_group_info *grp = ext4_get_group_info(sb, i); 5676 struct ext4_prealloc_space *pa; 5677 ext4_grpblk_t start; 5678 struct list_head *cur; 5679 5680 if (!grp) 5681 continue; 5682 ext4_lock_group(sb, i); 5683 list_for_each(cur, &grp->bb_prealloc_list) { 5684 pa = list_entry(cur, struct ext4_prealloc_space, 5685 pa_group_list); 5686 spin_lock(&pa->pa_lock); 5687 ext4_get_group_no_and_offset(sb, pa->pa_pstart, 5688 NULL, &start); 5689 spin_unlock(&pa->pa_lock); 5690 mb_debug(sb, "PA:%u:%d:%d\n", i, start, 5691 pa->pa_len); 5692 } 5693 ext4_unlock_group(sb, i); 5694 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, 5695 grp->bb_fragments); 5696 } 5697 } 5698 5699 static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 5700 { 5701 struct super_block *sb = ac->ac_sb; 5702 5703 if (ext4_forced_shutdown(sb)) 5704 return; 5705 5706 mb_debug(sb, "Can't allocate:" 5707 " Allocation context details:"); 5708 mb_debug(sb, "status %u flags 0x%x", 5709 ac->ac_status, ac->ac_flags); 5710 mb_debug(sb, "orig %lu/%lu/%lu@%lu, " 5711 "goal %lu/%lu/%lu@%lu, " 5712 "best %lu/%lu/%lu@%lu cr %d", 5713 (unsigned long)ac->ac_o_ex.fe_group, 5714 (unsigned long)ac->ac_o_ex.fe_start, 5715 (unsigned long)ac->ac_o_ex.fe_len, 5716 (unsigned long)ac->ac_o_ex.fe_logical, 5717 (unsigned long)ac->ac_g_ex.fe_group, 5718 (unsigned long)ac->ac_g_ex.fe_start, 5719 (unsigned long)ac->ac_g_ex.fe_len, 5720 (unsigned long)ac->ac_g_ex.fe_logical, 5721 (unsigned long)ac->ac_b_ex.fe_group, 5722 (unsigned long)ac->ac_b_ex.fe_start, 5723 (unsigned long)ac->ac_b_ex.fe_len, 5724 (unsigned long)ac->ac_b_ex.fe_logical, 5725 (int)ac->ac_criteria); 5726 mb_debug(sb, "%u found", ac->ac_found); 5727 mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no"); 5728 if (ac->ac_pa) 5729 mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ? 5730 "group pa" : "inode pa"); 5731 ext4_mb_show_pa(sb); 5732 } 5733 #else 5734 static inline void ext4_mb_show_pa(struct super_block *sb) 5735 { 5736 } 5737 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 5738 { 5739 ext4_mb_show_pa(ac->ac_sb); 5740 } 5741 #endif 5742 5743 /* 5744 * We use locality group preallocation for small size file. The size of the 5745 * file is determined by the current size or the resulting size after 5746 * allocation which ever is larger 5747 * 5748 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 5749 */ 5750 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 5751 { 5752 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 5753 int bsbits = ac->ac_sb->s_blocksize_bits; 5754 loff_t size, isize; 5755 bool inode_pa_eligible, group_pa_eligible; 5756 5757 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 5758 return; 5759 5760 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 5761 return; 5762 5763 group_pa_eligible = sbi->s_mb_group_prealloc > 0; 5764 inode_pa_eligible = true; 5765 size = extent_logical_end(sbi, &ac->ac_o_ex); 5766 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 5767 >> bsbits; 5768 5769 /* No point in using inode preallocation for closed files */ 5770 if ((size == isize) && !ext4_fs_is_busy(sbi) && 5771 !inode_is_open_for_write(ac->ac_inode)) 5772 inode_pa_eligible = false; 5773 5774 size = max(size, isize); 5775 /* Don't use group allocation for large files */ 5776 if (size > sbi->s_mb_stream_request) 5777 group_pa_eligible = false; 5778 5779 if (!group_pa_eligible) { 5780 if (inode_pa_eligible) 5781 ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 5782 else 5783 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 5784 return; 5785 } 5786 5787 BUG_ON(ac->ac_lg != NULL); 5788 /* 5789 * locality group prealloc space are per cpu. The reason for having 5790 * per cpu locality group is to reduce the contention between block 5791 * request from multiple CPUs. 5792 */ 5793 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); 5794 5795 /* we're going to use group allocation */ 5796 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 5797 5798 /* serialize all allocations in the group */ 5799 mutex_lock(&ac->ac_lg->lg_mutex); 5800 } 5801 5802 static noinline_for_stack void 5803 ext4_mb_initialize_context(struct ext4_allocation_context *ac, 5804 struct ext4_allocation_request *ar) 5805 { 5806 struct super_block *sb = ar->inode->i_sb; 5807 struct ext4_sb_info *sbi = EXT4_SB(sb); 5808 struct ext4_super_block *es = sbi->s_es; 5809 ext4_group_t group; 5810 unsigned int len; 5811 ext4_fsblk_t goal; 5812 ext4_grpblk_t block; 5813 5814 /* we can't allocate > group size */ 5815 len = ar->len; 5816 5817 /* just a dirty hack to filter too big requests */ 5818 if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) 5819 len = EXT4_CLUSTERS_PER_GROUP(sb); 5820 5821 /* start searching from the goal */ 5822 goal = ar->goal; 5823 if (goal < le32_to_cpu(es->s_first_data_block) || 5824 goal >= ext4_blocks_count(es)) 5825 goal = le32_to_cpu(es->s_first_data_block); 5826 ext4_get_group_no_and_offset(sb, goal, &group, &block); 5827 5828 /* set up allocation goals */ 5829 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); 5830 ac->ac_status = AC_STATUS_CONTINUE; 5831 ac->ac_sb = sb; 5832 ac->ac_inode = ar->inode; 5833 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 5834 ac->ac_o_ex.fe_group = group; 5835 ac->ac_o_ex.fe_start = block; 5836 ac->ac_o_ex.fe_len = len; 5837 ac->ac_g_ex = ac->ac_o_ex; 5838 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; 5839 ac->ac_flags = ar->flags; 5840 5841 /* we have to define context: we'll work with a file or 5842 * locality group. this is a policy, actually */ 5843 ext4_mb_group_or_file(ac); 5844 5845 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " 5846 "left: %u/%u, right %u/%u to %swritable\n", 5847 (unsigned) ar->len, (unsigned) ar->logical, 5848 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 5849 (unsigned) ar->lleft, (unsigned) ar->pleft, 5850 (unsigned) ar->lright, (unsigned) ar->pright, 5851 inode_is_open_for_write(ar->inode) ? "" : "non-"); 5852 } 5853 5854 static noinline_for_stack void 5855 ext4_mb_discard_lg_preallocations(struct super_block *sb, 5856 struct ext4_locality_group *lg, 5857 int order, int total_entries) 5858 { 5859 ext4_group_t group = 0; 5860 struct ext4_buddy e4b; 5861 struct list_head discard_list; 5862 struct ext4_prealloc_space *pa, *tmp; 5863 5864 mb_debug(sb, "discard locality group preallocation\n"); 5865 5866 INIT_LIST_HEAD(&discard_list); 5867 5868 spin_lock(&lg->lg_prealloc_lock); 5869 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 5870 pa_node.lg_list, 5871 lockdep_is_held(&lg->lg_prealloc_lock)) { 5872 spin_lock(&pa->pa_lock); 5873 if (atomic_read(&pa->pa_count)) { 5874 /* 5875 * This is the pa that we just used 5876 * for block allocation. So don't 5877 * free that 5878 */ 5879 spin_unlock(&pa->pa_lock); 5880 continue; 5881 } 5882 if (pa->pa_deleted) { 5883 spin_unlock(&pa->pa_lock); 5884 continue; 5885 } 5886 /* only lg prealloc space */ 5887 BUG_ON(pa->pa_type != MB_GROUP_PA); 5888 5889 /* seems this one can be freed ... */ 5890 ext4_mb_mark_pa_deleted(sb, pa); 5891 spin_unlock(&pa->pa_lock); 5892 5893 list_del_rcu(&pa->pa_node.lg_list); 5894 list_add(&pa->u.pa_tmp_list, &discard_list); 5895 5896 total_entries--; 5897 if (total_entries <= 5) { 5898 /* 5899 * we want to keep only 5 entries 5900 * allowing it to grow to 8. This 5901 * mak sure we don't call discard 5902 * soon for this list. 5903 */ 5904 break; 5905 } 5906 } 5907 spin_unlock(&lg->lg_prealloc_lock); 5908 5909 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 5910 int err; 5911 5912 group = ext4_get_group_number(sb, pa->pa_pstart); 5913 err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 5914 GFP_NOFS|__GFP_NOFAIL); 5915 if (err) { 5916 ext4_error_err(sb, -err, "Error %d loading buddy information for %u", 5917 err, group); 5918 continue; 5919 } 5920 ext4_lock_group(sb, group); 5921 list_del(&pa->pa_group_list); 5922 ext4_mb_release_group_pa(&e4b, pa); 5923 ext4_unlock_group(sb, group); 5924 5925 ext4_mb_unload_buddy(&e4b); 5926 list_del(&pa->u.pa_tmp_list); 5927 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 5928 } 5929 } 5930 5931 /* 5932 * We have incremented pa_count. So it cannot be freed at this 5933 * point. Also we hold lg_mutex. So no parallel allocation is 5934 * possible from this lg. That means pa_free cannot be updated. 5935 * 5936 * A parallel ext4_mb_discard_group_preallocations is possible. 5937 * which can cause the lg_prealloc_list to be updated. 5938 */ 5939 5940 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 5941 { 5942 int order, added = 0, lg_prealloc_count = 1; 5943 struct super_block *sb = ac->ac_sb; 5944 struct ext4_locality_group *lg = ac->ac_lg; 5945 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 5946 5947 order = fls(pa->pa_free) - 1; 5948 if (order > PREALLOC_TB_SIZE - 1) 5949 /* The max size of hash table is PREALLOC_TB_SIZE */ 5950 order = PREALLOC_TB_SIZE - 1; 5951 /* Add the prealloc space to lg */ 5952 spin_lock(&lg->lg_prealloc_lock); 5953 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 5954 pa_node.lg_list, 5955 lockdep_is_held(&lg->lg_prealloc_lock)) { 5956 spin_lock(&tmp_pa->pa_lock); 5957 if (tmp_pa->pa_deleted) { 5958 spin_unlock(&tmp_pa->pa_lock); 5959 continue; 5960 } 5961 if (!added && pa->pa_free < tmp_pa->pa_free) { 5962 /* Add to the tail of the previous entry */ 5963 list_add_tail_rcu(&pa->pa_node.lg_list, 5964 &tmp_pa->pa_node.lg_list); 5965 added = 1; 5966 /* 5967 * we want to count the total 5968 * number of entries in the list 5969 */ 5970 } 5971 spin_unlock(&tmp_pa->pa_lock); 5972 lg_prealloc_count++; 5973 } 5974 if (!added) 5975 list_add_tail_rcu(&pa->pa_node.lg_list, 5976 &lg->lg_prealloc_list[order]); 5977 spin_unlock(&lg->lg_prealloc_lock); 5978 5979 /* Now trim the list to be not more than 8 elements */ 5980 if (lg_prealloc_count > 8) 5981 ext4_mb_discard_lg_preallocations(sb, lg, 5982 order, lg_prealloc_count); 5983 } 5984 5985 /* 5986 * release all resource we used in allocation 5987 */ 5988 static int ext4_mb_release_context(struct ext4_allocation_context *ac) 5989 { 5990 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 5991 struct ext4_prealloc_space *pa = ac->ac_pa; 5992 if (pa) { 5993 if (pa->pa_type == MB_GROUP_PA) { 5994 /* see comment in ext4_mb_use_group_pa() */ 5995 spin_lock(&pa->pa_lock); 5996 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 5997 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 5998 pa->pa_free -= ac->ac_b_ex.fe_len; 5999 pa->pa_len -= ac->ac_b_ex.fe_len; 6000 spin_unlock(&pa->pa_lock); 6001 6002 /* 6003 * We want to add the pa to the right bucket. 6004 * Remove it from the list and while adding 6005 * make sure the list to which we are adding 6006 * doesn't grow big. 6007 */ 6008 if (likely(pa->pa_free)) { 6009 spin_lock(pa->pa_node_lock.lg_lock); 6010 list_del_rcu(&pa->pa_node.lg_list); 6011 spin_unlock(pa->pa_node_lock.lg_lock); 6012 ext4_mb_add_n_trim(ac); 6013 } 6014 } 6015 6016 ext4_mb_put_pa(ac, ac->ac_sb, pa); 6017 } 6018 if (ac->ac_bitmap_page) 6019 put_page(ac->ac_bitmap_page); 6020 if (ac->ac_buddy_page) 6021 put_page(ac->ac_buddy_page); 6022 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 6023 mutex_unlock(&ac->ac_lg->lg_mutex); 6024 ext4_mb_collect_stats(ac); 6025 return 0; 6026 } 6027 6028 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 6029 { 6030 ext4_group_t i, ngroups = ext4_get_groups_count(sb); 6031 int ret; 6032 int freed = 0, busy = 0; 6033 int retry = 0; 6034 6035 trace_ext4_mb_discard_preallocations(sb, needed); 6036 6037 if (needed == 0) 6038 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 6039 repeat: 6040 for (i = 0; i < ngroups && needed > 0; i++) { 6041 ret = ext4_mb_discard_group_preallocations(sb, i, &busy); 6042 freed += ret; 6043 needed -= ret; 6044 cond_resched(); 6045 } 6046 6047 if (needed > 0 && busy && ++retry < 3) { 6048 busy = 0; 6049 goto repeat; 6050 } 6051 6052 return freed; 6053 } 6054 6055 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, 6056 struct ext4_allocation_context *ac, u64 *seq) 6057 { 6058 int freed; 6059 u64 seq_retry = 0; 6060 bool ret = false; 6061 6062 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 6063 if (freed) { 6064 ret = true; 6065 goto out_dbg; 6066 } 6067 seq_retry = ext4_get_discard_pa_seq_sum(); 6068 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { 6069 ac->ac_flags |= EXT4_MB_STRICT_CHECK; 6070 *seq = seq_retry; 6071 ret = true; 6072 } 6073 6074 out_dbg: 6075 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); 6076 return ret; 6077 } 6078 6079 /* 6080 * Simple allocator for Ext4 fast commit replay path. It searches for blocks 6081 * linearly starting at the goal block and also excludes the blocks which 6082 * are going to be in use after fast commit replay. 6083 */ 6084 static ext4_fsblk_t 6085 ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp) 6086 { 6087 struct buffer_head *bitmap_bh; 6088 struct super_block *sb = ar->inode->i_sb; 6089 struct ext4_sb_info *sbi = EXT4_SB(sb); 6090 ext4_group_t group, nr; 6091 ext4_grpblk_t blkoff; 6092 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 6093 ext4_grpblk_t i = 0; 6094 ext4_fsblk_t goal, block; 6095 struct ext4_super_block *es = EXT4_SB(sb)->s_es; 6096 6097 goal = ar->goal; 6098 if (goal < le32_to_cpu(es->s_first_data_block) || 6099 goal >= ext4_blocks_count(es)) 6100 goal = le32_to_cpu(es->s_first_data_block); 6101 6102 ar->len = 0; 6103 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); 6104 for (nr = ext4_get_groups_count(sb); nr > 0; nr--) { 6105 bitmap_bh = ext4_read_block_bitmap(sb, group); 6106 if (IS_ERR(bitmap_bh)) { 6107 *errp = PTR_ERR(bitmap_bh); 6108 pr_warn("Failed to read block bitmap\n"); 6109 return 0; 6110 } 6111 6112 while (1) { 6113 i = mb_find_next_zero_bit(bitmap_bh->b_data, max, 6114 blkoff); 6115 if (i >= max) 6116 break; 6117 if (ext4_fc_replay_check_excluded(sb, 6118 ext4_group_first_block_no(sb, group) + 6119 EXT4_C2B(sbi, i))) { 6120 blkoff = i + 1; 6121 } else 6122 break; 6123 } 6124 brelse(bitmap_bh); 6125 if (i < max) 6126 break; 6127 6128 if (++group >= ext4_get_groups_count(sb)) 6129 group = 0; 6130 6131 blkoff = 0; 6132 } 6133 6134 if (i >= max) { 6135 *errp = -ENOSPC; 6136 return 0; 6137 } 6138 6139 block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i); 6140 ext4_mb_mark_bb(sb, block, 1, 1); 6141 ar->len = 1; 6142 6143 return block; 6144 } 6145 6146 /* 6147 * Main entry point into mballoc to allocate blocks 6148 * it tries to use preallocation first, then falls back 6149 * to usual allocation 6150 */ 6151 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 6152 struct ext4_allocation_request *ar, int *errp) 6153 { 6154 struct ext4_allocation_context *ac = NULL; 6155 struct ext4_sb_info *sbi; 6156 struct super_block *sb; 6157 ext4_fsblk_t block = 0; 6158 unsigned int inquota = 0; 6159 unsigned int reserv_clstrs = 0; 6160 int retries = 0; 6161 u64 seq; 6162 6163 might_sleep(); 6164 sb = ar->inode->i_sb; 6165 sbi = EXT4_SB(sb); 6166 6167 trace_ext4_request_blocks(ar); 6168 if (sbi->s_mount_state & EXT4_FC_REPLAY) 6169 return ext4_mb_new_blocks_simple(ar, errp); 6170 6171 /* Allow to use superuser reservation for quota file */ 6172 if (ext4_is_quota_file(ar->inode)) 6173 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 6174 6175 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { 6176 /* Without delayed allocation we need to verify 6177 * there is enough free blocks to do block allocation 6178 * and verify allocation doesn't exceed the quota limits. 6179 */ 6180 while (ar->len && 6181 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 6182 6183 /* let others to free the space */ 6184 cond_resched(); 6185 ar->len = ar->len >> 1; 6186 } 6187 if (!ar->len) { 6188 ext4_mb_show_pa(sb); 6189 *errp = -ENOSPC; 6190 return 0; 6191 } 6192 reserv_clstrs = ar->len; 6193 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 6194 dquot_alloc_block_nofail(ar->inode, 6195 EXT4_C2B(sbi, ar->len)); 6196 } else { 6197 while (ar->len && 6198 dquot_alloc_block(ar->inode, 6199 EXT4_C2B(sbi, ar->len))) { 6200 6201 ar->flags |= EXT4_MB_HINT_NOPREALLOC; 6202 ar->len--; 6203 } 6204 } 6205 inquota = ar->len; 6206 if (ar->len == 0) { 6207 *errp = -EDQUOT; 6208 goto out; 6209 } 6210 } 6211 6212 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); 6213 if (!ac) { 6214 ar->len = 0; 6215 *errp = -ENOMEM; 6216 goto out; 6217 } 6218 6219 ext4_mb_initialize_context(ac, ar); 6220 6221 ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 6222 seq = this_cpu_read(discard_pa_seq); 6223 if (!ext4_mb_use_preallocated(ac)) { 6224 ac->ac_op = EXT4_MB_HISTORY_ALLOC; 6225 ext4_mb_normalize_request(ac, ar); 6226 6227 *errp = ext4_mb_pa_alloc(ac); 6228 if (*errp) 6229 goto errout; 6230 repeat: 6231 /* allocate space in core */ 6232 *errp = ext4_mb_regular_allocator(ac); 6233 /* 6234 * pa allocated above is added to grp->bb_prealloc_list only 6235 * when we were able to allocate some block i.e. when 6236 * ac->ac_status == AC_STATUS_FOUND. 6237 * And error from above mean ac->ac_status != AC_STATUS_FOUND 6238 * So we have to free this pa here itself. 6239 */ 6240 if (*errp) { 6241 ext4_mb_pa_put_free(ac); 6242 ext4_discard_allocated_blocks(ac); 6243 goto errout; 6244 } 6245 if (ac->ac_status == AC_STATUS_FOUND && 6246 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) 6247 ext4_mb_pa_put_free(ac); 6248 } 6249 if (likely(ac->ac_status == AC_STATUS_FOUND)) { 6250 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 6251 if (*errp) { 6252 ext4_discard_allocated_blocks(ac); 6253 goto errout; 6254 } else { 6255 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 6256 ar->len = ac->ac_b_ex.fe_len; 6257 } 6258 } else { 6259 if (++retries < 3 && 6260 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) 6261 goto repeat; 6262 /* 6263 * If block allocation fails then the pa allocated above 6264 * needs to be freed here itself. 6265 */ 6266 ext4_mb_pa_put_free(ac); 6267 *errp = -ENOSPC; 6268 } 6269 6270 if (*errp) { 6271 errout: 6272 ac->ac_b_ex.fe_len = 0; 6273 ar->len = 0; 6274 ext4_mb_show_ac(ac); 6275 } 6276 ext4_mb_release_context(ac); 6277 kmem_cache_free(ext4_ac_cachep, ac); 6278 out: 6279 if (inquota && ar->len < inquota) 6280 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 6281 if (!ar->len) { 6282 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) 6283 /* release all the reserved blocks if non delalloc */ 6284 percpu_counter_sub(&sbi->s_dirtyclusters_counter, 6285 reserv_clstrs); 6286 } 6287 6288 trace_ext4_allocate_blocks(ar, (unsigned long long)block); 6289 6290 return block; 6291 } 6292 6293 /* 6294 * We can merge two free data extents only if the physical blocks 6295 * are contiguous, AND the extents were freed by the same transaction, 6296 * AND the blocks are associated with the same group. 6297 */ 6298 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, 6299 struct ext4_free_data *entry, 6300 struct ext4_free_data *new_entry, 6301 struct rb_root *entry_rb_root) 6302 { 6303 if ((entry->efd_tid != new_entry->efd_tid) || 6304 (entry->efd_group != new_entry->efd_group)) 6305 return; 6306 if (entry->efd_start_cluster + entry->efd_count == 6307 new_entry->efd_start_cluster) { 6308 new_entry->efd_start_cluster = entry->efd_start_cluster; 6309 new_entry->efd_count += entry->efd_count; 6310 } else if (new_entry->efd_start_cluster + new_entry->efd_count == 6311 entry->efd_start_cluster) { 6312 new_entry->efd_count += entry->efd_count; 6313 } else 6314 return; 6315 spin_lock(&sbi->s_md_lock); 6316 list_del(&entry->efd_list); 6317 spin_unlock(&sbi->s_md_lock); 6318 rb_erase(&entry->efd_node, entry_rb_root); 6319 kmem_cache_free(ext4_free_data_cachep, entry); 6320 } 6321 6322 static noinline_for_stack void 6323 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 6324 struct ext4_free_data *new_entry) 6325 { 6326 ext4_group_t group = e4b->bd_group; 6327 ext4_grpblk_t cluster; 6328 ext4_grpblk_t clusters = new_entry->efd_count; 6329 struct ext4_free_data *entry; 6330 struct ext4_group_info *db = e4b->bd_info; 6331 struct super_block *sb = e4b->bd_sb; 6332 struct ext4_sb_info *sbi = EXT4_SB(sb); 6333 struct rb_node **n = &db->bb_free_root.rb_node, *node; 6334 struct rb_node *parent = NULL, *new_node; 6335 6336 BUG_ON(!ext4_handle_valid(handle)); 6337 BUG_ON(e4b->bd_bitmap_page == NULL); 6338 BUG_ON(e4b->bd_buddy_page == NULL); 6339 6340 new_node = &new_entry->efd_node; 6341 cluster = new_entry->efd_start_cluster; 6342 6343 if (!*n) { 6344 /* first free block exent. We need to 6345 protect buddy cache from being freed, 6346 * otherwise we'll refresh it from 6347 * on-disk bitmap and lose not-yet-available 6348 * blocks */ 6349 get_page(e4b->bd_buddy_page); 6350 get_page(e4b->bd_bitmap_page); 6351 } 6352 while (*n) { 6353 parent = *n; 6354 entry = rb_entry(parent, struct ext4_free_data, efd_node); 6355 if (cluster < entry->efd_start_cluster) 6356 n = &(*n)->rb_left; 6357 else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 6358 n = &(*n)->rb_right; 6359 else { 6360 ext4_grp_locked_error(sb, group, 0, 6361 ext4_group_first_block_no(sb, group) + 6362 EXT4_C2B(sbi, cluster), 6363 "Block already on to-be-freed list"); 6364 kmem_cache_free(ext4_free_data_cachep, new_entry); 6365 return; 6366 } 6367 } 6368 6369 rb_link_node(new_node, parent, n); 6370 rb_insert_color(new_node, &db->bb_free_root); 6371 6372 /* Now try to see the extent can be merged to left and right */ 6373 node = rb_prev(new_node); 6374 if (node) { 6375 entry = rb_entry(node, struct ext4_free_data, efd_node); 6376 ext4_try_merge_freed_extent(sbi, entry, new_entry, 6377 &(db->bb_free_root)); 6378 } 6379 6380 node = rb_next(new_node); 6381 if (node) { 6382 entry = rb_entry(node, struct ext4_free_data, efd_node); 6383 ext4_try_merge_freed_extent(sbi, entry, new_entry, 6384 &(db->bb_free_root)); 6385 } 6386 6387 spin_lock(&sbi->s_md_lock); 6388 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); 6389 sbi->s_mb_free_pending += clusters; 6390 spin_unlock(&sbi->s_md_lock); 6391 } 6392 6393 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, 6394 unsigned long count) 6395 { 6396 struct buffer_head *bitmap_bh; 6397 struct super_block *sb = inode->i_sb; 6398 struct ext4_group_desc *gdp; 6399 struct buffer_head *gdp_bh; 6400 ext4_group_t group; 6401 ext4_grpblk_t blkoff; 6402 int already_freed = 0, err, i; 6403 6404 ext4_get_group_no_and_offset(sb, block, &group, &blkoff); 6405 bitmap_bh = ext4_read_block_bitmap(sb, group); 6406 if (IS_ERR(bitmap_bh)) { 6407 pr_warn("Failed to read block bitmap\n"); 6408 return; 6409 } 6410 gdp = ext4_get_group_desc(sb, group, &gdp_bh); 6411 if (!gdp) 6412 goto err_out; 6413 6414 for (i = 0; i < count; i++) { 6415 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data)) 6416 already_freed++; 6417 } 6418 mb_clear_bits(bitmap_bh->b_data, blkoff, count); 6419 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); 6420 if (err) 6421 goto err_out; 6422 ext4_free_group_clusters_set( 6423 sb, gdp, ext4_free_group_clusters(sb, gdp) + 6424 count - already_freed); 6425 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 6426 ext4_group_desc_csum_set(sb, group, gdp); 6427 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); 6428 sync_dirty_buffer(bitmap_bh); 6429 sync_dirty_buffer(gdp_bh); 6430 6431 err_out: 6432 brelse(bitmap_bh); 6433 } 6434 6435 /** 6436 * ext4_mb_clear_bb() -- helper function for freeing blocks. 6437 * Used by ext4_free_blocks() 6438 * @handle: handle for this transaction 6439 * @inode: inode 6440 * @block: starting physical block to be freed 6441 * @count: number of blocks to be freed 6442 * @flags: flags used by ext4_free_blocks 6443 */ 6444 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, 6445 ext4_fsblk_t block, unsigned long count, 6446 int flags) 6447 { 6448 struct buffer_head *bitmap_bh = NULL; 6449 struct super_block *sb = inode->i_sb; 6450 struct ext4_group_desc *gdp; 6451 struct ext4_group_info *grp; 6452 unsigned int overflow; 6453 ext4_grpblk_t bit; 6454 struct buffer_head *gd_bh; 6455 ext4_group_t block_group; 6456 struct ext4_sb_info *sbi; 6457 struct ext4_buddy e4b; 6458 unsigned int count_clusters; 6459 int err = 0; 6460 int ret; 6461 6462 sbi = EXT4_SB(sb); 6463 6464 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 6465 !ext4_inode_block_valid(inode, block, count)) { 6466 ext4_error(sb, "Freeing blocks in system zone - " 6467 "Block = %llu, count = %lu", block, count); 6468 /* err = 0. ext4_std_error should be a no op */ 6469 goto error_return; 6470 } 6471 flags |= EXT4_FREE_BLOCKS_VALIDATED; 6472 6473 do_more: 6474 overflow = 0; 6475 ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 6476 6477 grp = ext4_get_group_info(sb, block_group); 6478 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 6479 return; 6480 6481 /* 6482 * Check to see if we are freeing blocks across a group 6483 * boundary. 6484 */ 6485 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 6486 overflow = EXT4_C2B(sbi, bit) + count - 6487 EXT4_BLOCKS_PER_GROUP(sb); 6488 count -= overflow; 6489 /* The range changed so it's no longer validated */ 6490 flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6491 } 6492 count_clusters = EXT4_NUM_B2C(sbi, count); 6493 bitmap_bh = ext4_read_block_bitmap(sb, block_group); 6494 if (IS_ERR(bitmap_bh)) { 6495 err = PTR_ERR(bitmap_bh); 6496 bitmap_bh = NULL; 6497 goto error_return; 6498 } 6499 gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 6500 if (!gdp) { 6501 err = -EIO; 6502 goto error_return; 6503 } 6504 6505 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 6506 !ext4_inode_block_valid(inode, block, count)) { 6507 ext4_error(sb, "Freeing blocks in system zone - " 6508 "Block = %llu, count = %lu", block, count); 6509 /* err = 0. ext4_std_error should be a no op */ 6510 goto error_return; 6511 } 6512 6513 BUFFER_TRACE(bitmap_bh, "getting write access"); 6514 err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 6515 EXT4_JTR_NONE); 6516 if (err) 6517 goto error_return; 6518 6519 /* 6520 * We are about to modify some metadata. Call the journal APIs 6521 * to unshare ->b_data if a currently-committing transaction is 6522 * using it 6523 */ 6524 BUFFER_TRACE(gd_bh, "get_write_access"); 6525 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); 6526 if (err) 6527 goto error_return; 6528 #ifdef AGGRESSIVE_CHECK 6529 { 6530 int i; 6531 for (i = 0; i < count_clusters; i++) 6532 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 6533 } 6534 #endif 6535 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 6536 6537 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ 6538 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, 6539 GFP_NOFS|__GFP_NOFAIL); 6540 if (err) 6541 goto error_return; 6542 6543 /* 6544 * We need to make sure we don't reuse the freed block until after the 6545 * transaction is committed. We make an exception if the inode is to be 6546 * written in writeback mode since writeback mode has weak data 6547 * consistency guarantees. 6548 */ 6549 if (ext4_handle_valid(handle) && 6550 ((flags & EXT4_FREE_BLOCKS_METADATA) || 6551 !ext4_should_writeback_data(inode))) { 6552 struct ext4_free_data *new_entry; 6553 /* 6554 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed 6555 * to fail. 6556 */ 6557 new_entry = kmem_cache_alloc(ext4_free_data_cachep, 6558 GFP_NOFS|__GFP_NOFAIL); 6559 new_entry->efd_start_cluster = bit; 6560 new_entry->efd_group = block_group; 6561 new_entry->efd_count = count_clusters; 6562 new_entry->efd_tid = handle->h_transaction->t_tid; 6563 6564 ext4_lock_group(sb, block_group); 6565 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 6566 ext4_mb_free_metadata(handle, &e4b, new_entry); 6567 } else { 6568 /* need to update group_info->bb_free and bitmap 6569 * with group lock held. generate_buddy look at 6570 * them with group lock_held 6571 */ 6572 if (test_opt(sb, DISCARD)) { 6573 err = ext4_issue_discard(sb, block_group, bit, 6574 count_clusters, NULL); 6575 if (err && err != -EOPNOTSUPP) 6576 ext4_msg(sb, KERN_WARNING, "discard request in" 6577 " group:%u block:%d count:%lu failed" 6578 " with %d", block_group, bit, count, 6579 err); 6580 } else 6581 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); 6582 6583 ext4_lock_group(sb, block_group); 6584 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 6585 mb_free_blocks(inode, &e4b, bit, count_clusters); 6586 } 6587 6588 ret = ext4_free_group_clusters(sb, gdp) + count_clusters; 6589 ext4_free_group_clusters_set(sb, gdp, ret); 6590 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 6591 ext4_group_desc_csum_set(sb, block_group, gdp); 6592 ext4_unlock_group(sb, block_group); 6593 6594 if (sbi->s_log_groups_per_flex) { 6595 ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 6596 atomic64_add(count_clusters, 6597 &sbi_array_rcu_deref(sbi, s_flex_groups, 6598 flex_group)->free_clusters); 6599 } 6600 6601 /* 6602 * on a bigalloc file system, defer the s_freeclusters_counter 6603 * update to the caller (ext4_remove_space and friends) so they 6604 * can determine if a cluster freed here should be rereserved 6605 */ 6606 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { 6607 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 6608 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 6609 percpu_counter_add(&sbi->s_freeclusters_counter, 6610 count_clusters); 6611 } 6612 6613 ext4_mb_unload_buddy(&e4b); 6614 6615 /* We dirtied the bitmap block */ 6616 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 6617 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 6618 6619 /* And the group descriptor block */ 6620 BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 6621 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 6622 if (!err) 6623 err = ret; 6624 6625 if (overflow && !err) { 6626 block += count; 6627 count = overflow; 6628 put_bh(bitmap_bh); 6629 /* The range changed so it's no longer validated */ 6630 flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6631 goto do_more; 6632 } 6633 error_return: 6634 brelse(bitmap_bh); 6635 ext4_std_error(sb, err); 6636 } 6637 6638 /** 6639 * ext4_free_blocks() -- Free given blocks and update quota 6640 * @handle: handle for this transaction 6641 * @inode: inode 6642 * @bh: optional buffer of the block to be freed 6643 * @block: starting physical block to be freed 6644 * @count: number of blocks to be freed 6645 * @flags: flags used by ext4_free_blocks 6646 */ 6647 void ext4_free_blocks(handle_t *handle, struct inode *inode, 6648 struct buffer_head *bh, ext4_fsblk_t block, 6649 unsigned long count, int flags) 6650 { 6651 struct super_block *sb = inode->i_sb; 6652 unsigned int overflow; 6653 struct ext4_sb_info *sbi; 6654 6655 sbi = EXT4_SB(sb); 6656 6657 if (bh) { 6658 if (block) 6659 BUG_ON(block != bh->b_blocknr); 6660 else 6661 block = bh->b_blocknr; 6662 } 6663 6664 if (sbi->s_mount_state & EXT4_FC_REPLAY) { 6665 ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count)); 6666 return; 6667 } 6668 6669 might_sleep(); 6670 6671 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 6672 !ext4_inode_block_valid(inode, block, count)) { 6673 ext4_error(sb, "Freeing blocks not in datazone - " 6674 "block = %llu, count = %lu", block, count); 6675 return; 6676 } 6677 flags |= EXT4_FREE_BLOCKS_VALIDATED; 6678 6679 ext4_debug("freeing block %llu\n", block); 6680 trace_ext4_free_blocks(inode, block, count, flags); 6681 6682 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 6683 BUG_ON(count > 1); 6684 6685 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 6686 inode, bh, block); 6687 } 6688 6689 /* 6690 * If the extent to be freed does not begin on a cluster 6691 * boundary, we need to deal with partial clusters at the 6692 * beginning and end of the extent. Normally we will free 6693 * blocks at the beginning or the end unless we are explicitly 6694 * requested to avoid doing so. 6695 */ 6696 overflow = EXT4_PBLK_COFF(sbi, block); 6697 if (overflow) { 6698 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 6699 overflow = sbi->s_cluster_ratio - overflow; 6700 block += overflow; 6701 if (count > overflow) 6702 count -= overflow; 6703 else 6704 return; 6705 } else { 6706 block -= overflow; 6707 count += overflow; 6708 } 6709 /* The range changed so it's no longer validated */ 6710 flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6711 } 6712 overflow = EXT4_LBLK_COFF(sbi, count); 6713 if (overflow) { 6714 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 6715 if (count > overflow) 6716 count -= overflow; 6717 else 6718 return; 6719 } else 6720 count += sbi->s_cluster_ratio - overflow; 6721 /* The range changed so it's no longer validated */ 6722 flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6723 } 6724 6725 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 6726 int i; 6727 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; 6728 6729 for (i = 0; i < count; i++) { 6730 cond_resched(); 6731 if (is_metadata) 6732 bh = sb_find_get_block(inode->i_sb, block + i); 6733 ext4_forget(handle, is_metadata, inode, bh, block + i); 6734 } 6735 } 6736 6737 ext4_mb_clear_bb(handle, inode, block, count, flags); 6738 } 6739 6740 /** 6741 * ext4_group_add_blocks() -- Add given blocks to an existing group 6742 * @handle: handle to this transaction 6743 * @sb: super block 6744 * @block: start physical block to add to the block group 6745 * @count: number of blocks to free 6746 * 6747 * This marks the blocks as free in the bitmap and buddy. 6748 */ 6749 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 6750 ext4_fsblk_t block, unsigned long count) 6751 { 6752 struct buffer_head *bitmap_bh = NULL; 6753 struct buffer_head *gd_bh; 6754 ext4_group_t block_group; 6755 ext4_grpblk_t bit; 6756 unsigned int i; 6757 struct ext4_group_desc *desc; 6758 struct ext4_sb_info *sbi = EXT4_SB(sb); 6759 struct ext4_buddy e4b; 6760 int err = 0, ret, free_clusters_count; 6761 ext4_grpblk_t clusters_freed; 6762 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); 6763 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); 6764 unsigned long cluster_count = last_cluster - first_cluster + 1; 6765 6766 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 6767 6768 if (count == 0) 6769 return 0; 6770 6771 ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 6772 /* 6773 * Check to see if we are freeing blocks across a group 6774 * boundary. 6775 */ 6776 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { 6777 ext4_warning(sb, "too many blocks added to group %u", 6778 block_group); 6779 err = -EINVAL; 6780 goto error_return; 6781 } 6782 6783 bitmap_bh = ext4_read_block_bitmap(sb, block_group); 6784 if (IS_ERR(bitmap_bh)) { 6785 err = PTR_ERR(bitmap_bh); 6786 bitmap_bh = NULL; 6787 goto error_return; 6788 } 6789 6790 desc = ext4_get_group_desc(sb, block_group, &gd_bh); 6791 if (!desc) { 6792 err = -EIO; 6793 goto error_return; 6794 } 6795 6796 if (!ext4_sb_block_valid(sb, NULL, block, count)) { 6797 ext4_error(sb, "Adding blocks in system zones - " 6798 "Block = %llu, count = %lu", 6799 block, count); 6800 err = -EINVAL; 6801 goto error_return; 6802 } 6803 6804 BUFFER_TRACE(bitmap_bh, "getting write access"); 6805 err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 6806 EXT4_JTR_NONE); 6807 if (err) 6808 goto error_return; 6809 6810 /* 6811 * We are about to modify some metadata. Call the journal APIs 6812 * to unshare ->b_data if a currently-committing transaction is 6813 * using it 6814 */ 6815 BUFFER_TRACE(gd_bh, "get_write_access"); 6816 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); 6817 if (err) 6818 goto error_return; 6819 6820 for (i = 0, clusters_freed = 0; i < cluster_count; i++) { 6821 BUFFER_TRACE(bitmap_bh, "clear bit"); 6822 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { 6823 ext4_error(sb, "bit already cleared for block %llu", 6824 (ext4_fsblk_t)(block + i)); 6825 BUFFER_TRACE(bitmap_bh, "bit already cleared"); 6826 } else { 6827 clusters_freed++; 6828 } 6829 } 6830 6831 err = ext4_mb_load_buddy(sb, block_group, &e4b); 6832 if (err) 6833 goto error_return; 6834 6835 /* 6836 * need to update group_info->bb_free and bitmap 6837 * with group lock held. generate_buddy look at 6838 * them with group lock_held 6839 */ 6840 ext4_lock_group(sb, block_group); 6841 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); 6842 mb_free_blocks(NULL, &e4b, bit, cluster_count); 6843 free_clusters_count = clusters_freed + 6844 ext4_free_group_clusters(sb, desc); 6845 ext4_free_group_clusters_set(sb, desc, free_clusters_count); 6846 ext4_block_bitmap_csum_set(sb, desc, bitmap_bh); 6847 ext4_group_desc_csum_set(sb, block_group, desc); 6848 ext4_unlock_group(sb, block_group); 6849 percpu_counter_add(&sbi->s_freeclusters_counter, 6850 clusters_freed); 6851 6852 if (sbi->s_log_groups_per_flex) { 6853 ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 6854 atomic64_add(clusters_freed, 6855 &sbi_array_rcu_deref(sbi, s_flex_groups, 6856 flex_group)->free_clusters); 6857 } 6858 6859 ext4_mb_unload_buddy(&e4b); 6860 6861 /* We dirtied the bitmap block */ 6862 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 6863 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 6864 6865 /* And the group descriptor block */ 6866 BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 6867 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 6868 if (!err) 6869 err = ret; 6870 6871 error_return: 6872 brelse(bitmap_bh); 6873 ext4_std_error(sb, err); 6874 return err; 6875 } 6876 6877 /** 6878 * ext4_trim_extent -- function to TRIM one single free extent in the group 6879 * @sb: super block for the file system 6880 * @start: starting block of the free extent in the alloc. group 6881 * @count: number of blocks to TRIM 6882 * @e4b: ext4 buddy for the group 6883 * 6884 * Trim "count" blocks starting at "start" in the "group". To assure that no 6885 * one will allocate those blocks, mark it as used in buddy bitmap. This must 6886 * be called with under the group lock. 6887 */ 6888 static int ext4_trim_extent(struct super_block *sb, 6889 int start, int count, struct ext4_buddy *e4b) 6890 __releases(bitlock) 6891 __acquires(bitlock) 6892 { 6893 struct ext4_free_extent ex; 6894 ext4_group_t group = e4b->bd_group; 6895 int ret = 0; 6896 6897 trace_ext4_trim_extent(sb, group, start, count); 6898 6899 assert_spin_locked(ext4_group_lock_ptr(sb, group)); 6900 6901 ex.fe_start = start; 6902 ex.fe_group = group; 6903 ex.fe_len = count; 6904 6905 /* 6906 * Mark blocks used, so no one can reuse them while 6907 * being trimmed. 6908 */ 6909 mb_mark_used(e4b, &ex); 6910 ext4_unlock_group(sb, group); 6911 ret = ext4_issue_discard(sb, group, start, count, NULL); 6912 ext4_lock_group(sb, group); 6913 mb_free_blocks(NULL, e4b, start, ex.fe_len); 6914 return ret; 6915 } 6916 6917 static int ext4_try_to_trim_range(struct super_block *sb, 6918 struct ext4_buddy *e4b, ext4_grpblk_t start, 6919 ext4_grpblk_t max, ext4_grpblk_t minblocks) 6920 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) 6921 __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) 6922 { 6923 ext4_grpblk_t next, count, free_count; 6924 void *bitmap; 6925 6926 bitmap = e4b->bd_bitmap; 6927 start = max(e4b->bd_info->bb_first_free, start); 6928 count = 0; 6929 free_count = 0; 6930 6931 while (start <= max) { 6932 start = mb_find_next_zero_bit(bitmap, max + 1, start); 6933 if (start > max) 6934 break; 6935 next = mb_find_next_bit(bitmap, max + 1, start); 6936 6937 if ((next - start) >= minblocks) { 6938 int ret = ext4_trim_extent(sb, start, next - start, e4b); 6939 6940 if (ret && ret != -EOPNOTSUPP) 6941 break; 6942 count += next - start; 6943 } 6944 free_count += next - start; 6945 start = next + 1; 6946 6947 if (fatal_signal_pending(current)) { 6948 count = -ERESTARTSYS; 6949 break; 6950 } 6951 6952 if (need_resched()) { 6953 ext4_unlock_group(sb, e4b->bd_group); 6954 cond_resched(); 6955 ext4_lock_group(sb, e4b->bd_group); 6956 } 6957 6958 if ((e4b->bd_info->bb_free - free_count) < minblocks) 6959 break; 6960 } 6961 6962 return count; 6963 } 6964 6965 /** 6966 * ext4_trim_all_free -- function to trim all free space in alloc. group 6967 * @sb: super block for file system 6968 * @group: group to be trimmed 6969 * @start: first group block to examine 6970 * @max: last group block to examine 6971 * @minblocks: minimum extent block count 6972 * @set_trimmed: set the trimmed flag if at least one block is trimmed 6973 * 6974 * ext4_trim_all_free walks through group's block bitmap searching for free 6975 * extents. When the free extent is found, mark it as used in group buddy 6976 * bitmap. Then issue a TRIM command on this extent and free the extent in 6977 * the group buddy bitmap. 6978 */ 6979 static ext4_grpblk_t 6980 ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 6981 ext4_grpblk_t start, ext4_grpblk_t max, 6982 ext4_grpblk_t minblocks, bool set_trimmed) 6983 { 6984 struct ext4_buddy e4b; 6985 int ret; 6986 6987 trace_ext4_trim_all_free(sb, group, start, max); 6988 6989 ret = ext4_mb_load_buddy(sb, group, &e4b); 6990 if (ret) { 6991 ext4_warning(sb, "Error %d loading buddy information for %u", 6992 ret, group); 6993 return ret; 6994 } 6995 6996 ext4_lock_group(sb, group); 6997 6998 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || 6999 minblocks < EXT4_SB(sb)->s_last_trim_minblks) { 7000 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); 7001 if (ret >= 0 && set_trimmed) 7002 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); 7003 } else { 7004 ret = 0; 7005 } 7006 7007 ext4_unlock_group(sb, group); 7008 ext4_mb_unload_buddy(&e4b); 7009 7010 ext4_debug("trimmed %d blocks in the group %d\n", 7011 ret, group); 7012 7013 return ret; 7014 } 7015 7016 /** 7017 * ext4_trim_fs() -- trim ioctl handle function 7018 * @sb: superblock for filesystem 7019 * @range: fstrim_range structure 7020 * 7021 * start: First Byte to trim 7022 * len: number of Bytes to trim from start 7023 * minlen: minimum extent length in Bytes 7024 * ext4_trim_fs goes through all allocation groups containing Bytes from 7025 * start to start+len. For each such a group ext4_trim_all_free function 7026 * is invoked to trim all free space. 7027 */ 7028 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 7029 { 7030 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); 7031 struct ext4_group_info *grp; 7032 ext4_group_t group, first_group, last_group; 7033 ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 7034 uint64_t start, end, minlen, trimmed = 0; 7035 ext4_fsblk_t first_data_blk = 7036 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 7037 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 7038 bool whole_group, eof = false; 7039 int ret = 0; 7040 7041 start = range->start >> sb->s_blocksize_bits; 7042 end = start + (range->len >> sb->s_blocksize_bits) - 1; 7043 minlen = EXT4_NUM_B2C(EXT4_SB(sb), 7044 range->minlen >> sb->s_blocksize_bits); 7045 7046 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || 7047 start >= max_blks || 7048 range->len < sb->s_blocksize) 7049 return -EINVAL; 7050 /* No point to try to trim less than discard granularity */ 7051 if (range->minlen < discard_granularity) { 7052 minlen = EXT4_NUM_B2C(EXT4_SB(sb), 7053 discard_granularity >> sb->s_blocksize_bits); 7054 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) 7055 goto out; 7056 } 7057 if (end >= max_blks - 1) { 7058 end = max_blks - 1; 7059 eof = true; 7060 } 7061 if (end <= first_data_blk) 7062 goto out; 7063 if (start < first_data_blk) 7064 start = first_data_blk; 7065 7066 /* Determine first and last group to examine based on start and end */ 7067 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 7068 &first_group, &first_cluster); 7069 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 7070 &last_group, &last_cluster); 7071 7072 /* end now represents the last cluster to discard in this group */ 7073 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 7074 whole_group = true; 7075 7076 for (group = first_group; group <= last_group; group++) { 7077 grp = ext4_get_group_info(sb, group); 7078 if (!grp) 7079 continue; 7080 /* We only do this if the grp has never been initialized */ 7081 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 7082 ret = ext4_mb_init_group(sb, group, GFP_NOFS); 7083 if (ret) 7084 break; 7085 } 7086 7087 /* 7088 * For all the groups except the last one, last cluster will 7089 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 7090 * change it for the last group, note that last_cluster is 7091 * already computed earlier by ext4_get_group_no_and_offset() 7092 */ 7093 if (group == last_group) { 7094 end = last_cluster; 7095 whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1; 7096 } 7097 if (grp->bb_free >= minlen) { 7098 cnt = ext4_trim_all_free(sb, group, first_cluster, 7099 end, minlen, whole_group); 7100 if (cnt < 0) { 7101 ret = cnt; 7102 break; 7103 } 7104 trimmed += cnt; 7105 } 7106 7107 /* 7108 * For every group except the first one, we are sure 7109 * that the first cluster to discard will be cluster #0. 7110 */ 7111 first_cluster = 0; 7112 } 7113 7114 if (!ret) 7115 EXT4_SB(sb)->s_last_trim_minblks = minlen; 7116 7117 out: 7118 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; 7119 return ret; 7120 } 7121 7122 /* Iterate all the free extents in the group. */ 7123 int 7124 ext4_mballoc_query_range( 7125 struct super_block *sb, 7126 ext4_group_t group, 7127 ext4_grpblk_t start, 7128 ext4_grpblk_t end, 7129 ext4_mballoc_query_range_fn formatter, 7130 void *priv) 7131 { 7132 void *bitmap; 7133 ext4_grpblk_t next; 7134 struct ext4_buddy e4b; 7135 int error; 7136 7137 error = ext4_mb_load_buddy(sb, group, &e4b); 7138 if (error) 7139 return error; 7140 bitmap = e4b.bd_bitmap; 7141 7142 ext4_lock_group(sb, group); 7143 7144 start = max(e4b.bd_info->bb_first_free, start); 7145 if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) 7146 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 7147 7148 while (start <= end) { 7149 start = mb_find_next_zero_bit(bitmap, end + 1, start); 7150 if (start > end) 7151 break; 7152 next = mb_find_next_bit(bitmap, end + 1, start); 7153 7154 ext4_unlock_group(sb, group); 7155 error = formatter(sb, group, start, next - start, priv); 7156 if (error) 7157 goto out_unload; 7158 ext4_lock_group(sb, group); 7159 7160 start = next + 1; 7161 } 7162 7163 ext4_unlock_group(sb, group); 7164 out_unload: 7165 ext4_mb_unload_buddy(&e4b); 7166 7167 return error; 7168 } 7169