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/backing-dev.h> 18 #include <trace/events/ext4.h> 19 20 #ifdef CONFIG_EXT4_DEBUG 21 ushort ext4_mballoc_debug __read_mostly; 22 23 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); 24 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); 25 #endif 26 27 /* 28 * MUSTDO: 29 * - test ext4_ext_search_left() and ext4_ext_search_right() 30 * - search for metadata in few groups 31 * 32 * TODO v4: 33 * - normalization should take into account whether file is still open 34 * - discard preallocations if no free space left (policy?) 35 * - don't normalize tails 36 * - quota 37 * - reservation for superuser 38 * 39 * TODO v3: 40 * - bitmap read-ahead (proposed by Oleg Drokin aka green) 41 * - track min/max extents in each group for better group selection 42 * - mb_mark_used() may allocate chunk right after splitting buddy 43 * - tree of groups sorted by number of free blocks 44 * - error handling 45 */ 46 47 /* 48 * The allocation request involve request for multiple number of blocks 49 * near to the goal(block) value specified. 50 * 51 * During initialization phase of the allocator we decide to use the 52 * group preallocation or inode preallocation depending on the size of 53 * the file. The size of the file could be the resulting file size we 54 * would have after allocation, or the current file size, which ever 55 * is larger. If the size is less than sbi->s_mb_stream_request we 56 * select to use the group preallocation. The default value of 57 * s_mb_stream_request is 16 blocks. This can also be tuned via 58 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 59 * terms of number of blocks. 60 * 61 * The main motivation for having small file use group preallocation is to 62 * ensure that we have small files closer together on the disk. 63 * 64 * First stage the allocator looks at the inode prealloc list, 65 * ext4_inode_info->i_prealloc_list, which contains list of prealloc 66 * spaces for this particular inode. The inode prealloc space is 67 * represented as: 68 * 69 * pa_lstart -> the logical start block for this prealloc space 70 * pa_pstart -> the physical start block for this prealloc space 71 * pa_len -> length for this prealloc space (in clusters) 72 * pa_free -> free space available in this prealloc space (in clusters) 73 * 74 * The inode preallocation space is used looking at the _logical_ start 75 * block. If only the logical file block falls within the range of prealloc 76 * space we will consume the particular prealloc space. This makes sure that 77 * we have contiguous physical blocks representing the file blocks 78 * 79 * The important thing to be noted in case of inode prealloc space is that 80 * we don't modify the values associated to inode prealloc space except 81 * pa_free. 82 * 83 * If we are not able to find blocks in the inode prealloc space and if we 84 * have the group allocation flag set then we look at the locality group 85 * prealloc space. These are per CPU prealloc list represented as 86 * 87 * ext4_sb_info.s_locality_groups[smp_processor_id()] 88 * 89 * The reason for having a per cpu locality group is to reduce the contention 90 * between CPUs. It is possible to get scheduled at this point. 91 * 92 * The locality group prealloc space is used looking at whether we have 93 * enough free space (pa_free) within the prealloc space. 94 * 95 * If we can't allocate blocks via inode prealloc or/and locality group 96 * prealloc then we look at the buddy cache. The buddy cache is represented 97 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 98 * mapped to the buddy and bitmap information regarding different 99 * groups. The buddy information is attached to buddy cache inode so that 100 * we can access them through the page cache. The information regarding 101 * each group is loaded via ext4_mb_load_buddy. The information involve 102 * block bitmap and buddy information. The information are stored in the 103 * inode as: 104 * 105 * { page } 106 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 107 * 108 * 109 * one block each for bitmap and buddy information. So for each group we 110 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / 111 * blocksize) blocks. So it can have information regarding groups_per_page 112 * which is blocks_per_page/2 113 * 114 * The buddy cache inode is not stored on disk. The inode is thrown 115 * away when the filesystem is unmounted. 116 * 117 * We look for count number of blocks in the buddy cache. If we were able 118 * to locate that many free blocks we return with additional information 119 * regarding rest of the contiguous physical block available 120 * 121 * Before allocating blocks via buddy cache we normalize the request 122 * blocks. This ensure we ask for more blocks that we needed. The extra 123 * blocks that we get after allocation is added to the respective prealloc 124 * list. In case of inode preallocation we follow a list of heuristics 125 * based on file size. This can be found in ext4_mb_normalize_request. If 126 * we are doing a group prealloc we try to normalize the request to 127 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 128 * dependent on the cluster size; for non-bigalloc file systems, it is 129 * 512 blocks. This can be tuned via 130 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 131 * terms of number of blocks. If we have mounted the file system with -O 132 * stripe=<value> option the group prealloc request is normalized to the 133 * the smallest multiple of the stripe value (sbi->s_stripe) which is 134 * greater than the default mb_group_prealloc. 135 * 136 * The regular allocator (using the buddy cache) supports a few tunables. 137 * 138 * /sys/fs/ext4/<partition>/mb_min_to_scan 139 * /sys/fs/ext4/<partition>/mb_max_to_scan 140 * /sys/fs/ext4/<partition>/mb_order2_req 141 * 142 * The regular allocator uses buddy scan only if the request len is power of 143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 144 * value of s_mb_order2_reqs can be tuned via 145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 146 * stripe size (sbi->s_stripe), we try to search for contiguous block in 147 * stripe size. This should result in better allocation on RAID setups. If 148 * not, we search in the specific group using bitmap for best extents. The 149 * tunable min_to_scan and max_to_scan control the behaviour here. 150 * min_to_scan indicate how long the mballoc __must__ look for a best 151 * extent and max_to_scan indicates how long the mballoc __can__ look for a 152 * best extent in the found extents. Searching for the blocks starts with 153 * the group specified as the goal value in allocation context via 154 * ac_g_ex. Each group is first checked based on the criteria whether it 155 * can be used for allocation. ext4_mb_good_group explains how the groups are 156 * checked. 157 * 158 * Both the prealloc space are getting populated as above. So for the first 159 * request we will hit the buddy cache which will result in this prealloc 160 * space getting filled. The prealloc space is then later used for the 161 * subsequent request. 162 */ 163 164 /* 165 * mballoc operates on the following data: 166 * - on-disk bitmap 167 * - in-core buddy (actually includes buddy and bitmap) 168 * - preallocation descriptors (PAs) 169 * 170 * there are two types of preallocations: 171 * - inode 172 * assiged to specific inode and can be used for this inode only. 173 * it describes part of inode's space preallocated to specific 174 * physical blocks. any block from that preallocated can be used 175 * independent. the descriptor just tracks number of blocks left 176 * unused. so, before taking some block from descriptor, one must 177 * make sure corresponded logical block isn't allocated yet. this 178 * also means that freeing any block within descriptor's range 179 * must discard all preallocated blocks. 180 * - locality group 181 * assigned to specific locality group which does not translate to 182 * permanent set of inodes: inode can join and leave group. space 183 * from this type of preallocation can be used for any inode. thus 184 * it's consumed from the beginning to the end. 185 * 186 * relation between them can be expressed as: 187 * in-core buddy = on-disk bitmap + preallocation descriptors 188 * 189 * this mean blocks mballoc considers used are: 190 * - allocated blocks (persistent) 191 * - preallocated blocks (non-persistent) 192 * 193 * consistency in mballoc world means that at any time a block is either 194 * free or used in ALL structures. notice: "any time" should not be read 195 * literally -- time is discrete and delimited by locks. 196 * 197 * to keep it simple, we don't use block numbers, instead we count number of 198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 199 * 200 * all operations can be expressed as: 201 * - init buddy: buddy = on-disk + PAs 202 * - new PA: buddy += N; PA = N 203 * - use inode PA: on-disk += N; PA -= N 204 * - discard inode PA buddy -= on-disk - PA; PA = 0 205 * - use locality group PA on-disk += N; PA -= N 206 * - discard locality group PA buddy -= PA; PA = 0 207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 208 * is used in real operation because we can't know actual used 209 * bits from PA, only from on-disk bitmap 210 * 211 * if we follow this strict logic, then all operations above should be atomic. 212 * given some of them can block, we'd have to use something like semaphores 213 * killing performance on high-end SMP hardware. let's try to relax it using 214 * the following knowledge: 215 * 1) if buddy is referenced, it's already initialized 216 * 2) while block is used in buddy and the buddy is referenced, 217 * nobody can re-allocate that block 218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 219 * bit set and PA claims same block, it's OK. IOW, one can set bit in 220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 221 * block 222 * 223 * so, now we're building a concurrency table: 224 * - init buddy vs. 225 * - new PA 226 * blocks for PA are allocated in the buddy, buddy must be referenced 227 * until PA is linked to allocation group to avoid concurrent buddy init 228 * - use inode PA 229 * we need to make sure that either on-disk bitmap or PA has uptodate data 230 * given (3) we care that PA-=N operation doesn't interfere with init 231 * - discard inode PA 232 * the simplest way would be to have buddy initialized by the discard 233 * - use locality group PA 234 * again PA-=N must be serialized with init 235 * - discard locality group PA 236 * the simplest way would be to have buddy initialized by the discard 237 * - new PA vs. 238 * - use inode PA 239 * i_data_sem serializes them 240 * - discard inode PA 241 * discard process must wait until PA isn't used by another process 242 * - use locality group PA 243 * some mutex should serialize them 244 * - discard locality group PA 245 * discard process must wait until PA isn't used by another process 246 * - use inode PA 247 * - use inode PA 248 * i_data_sem or another mutex should serializes them 249 * - discard inode PA 250 * discard process must wait until PA isn't used by another process 251 * - use locality group PA 252 * nothing wrong here -- they're different PAs covering different blocks 253 * - discard locality group PA 254 * discard process must wait until PA isn't used by another process 255 * 256 * now we're ready to make few consequences: 257 * - PA is referenced and while it is no discard is possible 258 * - PA is referenced until block isn't marked in on-disk bitmap 259 * - PA changes only after on-disk bitmap 260 * - discard must not compete with init. either init is done before 261 * any discard or they're serialized somehow 262 * - buddy init as sum of on-disk bitmap and PAs is done atomically 263 * 264 * a special case when we've used PA to emptiness. no need to modify buddy 265 * in this case, but we should care about concurrent init 266 * 267 */ 268 269 /* 270 * Logic in few words: 271 * 272 * - allocation: 273 * load group 274 * find blocks 275 * mark bits in on-disk bitmap 276 * release group 277 * 278 * - use preallocation: 279 * find proper PA (per-inode or group) 280 * load group 281 * mark bits in on-disk bitmap 282 * release group 283 * release PA 284 * 285 * - free: 286 * load group 287 * mark bits in on-disk bitmap 288 * release group 289 * 290 * - discard preallocations in group: 291 * mark PAs deleted 292 * move them onto local list 293 * load on-disk bitmap 294 * load group 295 * remove PA from object (inode or locality group) 296 * mark free blocks in-core 297 * 298 * - discard inode's preallocations: 299 */ 300 301 /* 302 * Locking rules 303 * 304 * Locks: 305 * - bitlock on a group (group) 306 * - object (inode/locality) (object) 307 * - per-pa lock (pa) 308 * 309 * Paths: 310 * - new pa 311 * object 312 * group 313 * 314 * - find and use pa: 315 * pa 316 * 317 * - release consumed pa: 318 * pa 319 * group 320 * object 321 * 322 * - generate in-core bitmap: 323 * group 324 * pa 325 * 326 * - discard all for given object (inode, locality group): 327 * object 328 * pa 329 * group 330 * 331 * - discard all for given group: 332 * group 333 * pa 334 * group 335 * object 336 * 337 */ 338 static struct kmem_cache *ext4_pspace_cachep; 339 static struct kmem_cache *ext4_ac_cachep; 340 static struct kmem_cache *ext4_free_data_cachep; 341 342 /* We create slab caches for groupinfo data structures based on the 343 * superblock block size. There will be one per mounted filesystem for 344 * each unique s_blocksize_bits */ 345 #define NR_GRPINFO_CACHES 8 346 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 347 348 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 349 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 350 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 351 "ext4_groupinfo_64k", "ext4_groupinfo_128k" 352 }; 353 354 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 355 ext4_group_t group); 356 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 357 ext4_group_t group); 358 359 static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 360 { 361 #if BITS_PER_LONG == 64 362 *bit += ((unsigned long) addr & 7UL) << 3; 363 addr = (void *) ((unsigned long) addr & ~7UL); 364 #elif BITS_PER_LONG == 32 365 *bit += ((unsigned long) addr & 3UL) << 3; 366 addr = (void *) ((unsigned long) addr & ~3UL); 367 #else 368 #error "how many bits you are?!" 369 #endif 370 return addr; 371 } 372 373 static inline int mb_test_bit(int bit, void *addr) 374 { 375 /* 376 * ext4_test_bit on architecture like powerpc 377 * needs unsigned long aligned address 378 */ 379 addr = mb_correct_addr_and_bit(&bit, addr); 380 return ext4_test_bit(bit, addr); 381 } 382 383 static inline void mb_set_bit(int bit, void *addr) 384 { 385 addr = mb_correct_addr_and_bit(&bit, addr); 386 ext4_set_bit(bit, addr); 387 } 388 389 static inline void mb_clear_bit(int bit, void *addr) 390 { 391 addr = mb_correct_addr_and_bit(&bit, addr); 392 ext4_clear_bit(bit, addr); 393 } 394 395 static inline int mb_test_and_clear_bit(int bit, void *addr) 396 { 397 addr = mb_correct_addr_and_bit(&bit, addr); 398 return ext4_test_and_clear_bit(bit, addr); 399 } 400 401 static inline int mb_find_next_zero_bit(void *addr, int max, int start) 402 { 403 int fix = 0, ret, tmpmax; 404 addr = mb_correct_addr_and_bit(&fix, addr); 405 tmpmax = max + fix; 406 start += fix; 407 408 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 409 if (ret > max) 410 return max; 411 return ret; 412 } 413 414 static inline int mb_find_next_bit(void *addr, int max, int start) 415 { 416 int fix = 0, ret, tmpmax; 417 addr = mb_correct_addr_and_bit(&fix, addr); 418 tmpmax = max + fix; 419 start += fix; 420 421 ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 422 if (ret > max) 423 return max; 424 return ret; 425 } 426 427 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 428 { 429 char *bb; 430 431 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 432 BUG_ON(max == NULL); 433 434 if (order > e4b->bd_blkbits + 1) { 435 *max = 0; 436 return NULL; 437 } 438 439 /* at order 0 we see each particular block */ 440 if (order == 0) { 441 *max = 1 << (e4b->bd_blkbits + 3); 442 return e4b->bd_bitmap; 443 } 444 445 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 446 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 447 448 return bb; 449 } 450 451 #ifdef DOUBLE_CHECK 452 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 453 int first, int count) 454 { 455 int i; 456 struct super_block *sb = e4b->bd_sb; 457 458 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 459 return; 460 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 461 for (i = 0; i < count; i++) { 462 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 463 ext4_fsblk_t blocknr; 464 465 blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 466 blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 467 ext4_grp_locked_error(sb, e4b->bd_group, 468 inode ? inode->i_ino : 0, 469 blocknr, 470 "freeing block already freed " 471 "(bit %u)", 472 first + i); 473 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 474 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 475 } 476 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 477 } 478 } 479 480 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 481 { 482 int i; 483 484 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 485 return; 486 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 487 for (i = 0; i < count; i++) { 488 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 489 mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 490 } 491 } 492 493 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 494 { 495 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 496 unsigned char *b1, *b2; 497 int i; 498 b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 499 b2 = (unsigned char *) bitmap; 500 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 501 if (b1[i] != b2[i]) { 502 ext4_msg(e4b->bd_sb, KERN_ERR, 503 "corruption in group %u " 504 "at byte %u(%u): %x in copy != %x " 505 "on disk/prealloc", 506 e4b->bd_group, i, i * 8, b1[i], b2[i]); 507 BUG(); 508 } 509 } 510 } 511 } 512 513 #else 514 static inline void mb_free_blocks_double(struct inode *inode, 515 struct ext4_buddy *e4b, int first, int count) 516 { 517 return; 518 } 519 static inline void mb_mark_used_double(struct ext4_buddy *e4b, 520 int first, int count) 521 { 522 return; 523 } 524 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 525 { 526 return; 527 } 528 #endif 529 530 #ifdef AGGRESSIVE_CHECK 531 532 #define MB_CHECK_ASSERT(assert) \ 533 do { \ 534 if (!(assert)) { \ 535 printk(KERN_EMERG \ 536 "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 537 function, file, line, # assert); \ 538 BUG(); \ 539 } \ 540 } while (0) 541 542 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 543 const char *function, int line) 544 { 545 struct super_block *sb = e4b->bd_sb; 546 int order = e4b->bd_blkbits + 1; 547 int max; 548 int max2; 549 int i; 550 int j; 551 int k; 552 int count; 553 struct ext4_group_info *grp; 554 int fragments = 0; 555 int fstart; 556 struct list_head *cur; 557 void *buddy; 558 void *buddy2; 559 560 { 561 static int mb_check_counter; 562 if (mb_check_counter++ % 100 != 0) 563 return 0; 564 } 565 566 while (order > 1) { 567 buddy = mb_find_buddy(e4b, order, &max); 568 MB_CHECK_ASSERT(buddy); 569 buddy2 = mb_find_buddy(e4b, order - 1, &max2); 570 MB_CHECK_ASSERT(buddy2); 571 MB_CHECK_ASSERT(buddy != buddy2); 572 MB_CHECK_ASSERT(max * 2 == max2); 573 574 count = 0; 575 for (i = 0; i < max; i++) { 576 577 if (mb_test_bit(i, buddy)) { 578 /* only single bit in buddy2 may be 1 */ 579 if (!mb_test_bit(i << 1, buddy2)) { 580 MB_CHECK_ASSERT( 581 mb_test_bit((i<<1)+1, buddy2)); 582 } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 583 MB_CHECK_ASSERT( 584 mb_test_bit(i << 1, buddy2)); 585 } 586 continue; 587 } 588 589 /* both bits in buddy2 must be 1 */ 590 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 591 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 592 593 for (j = 0; j < (1 << order); j++) { 594 k = (i * (1 << order)) + j; 595 MB_CHECK_ASSERT( 596 !mb_test_bit(k, e4b->bd_bitmap)); 597 } 598 count++; 599 } 600 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 601 order--; 602 } 603 604 fstart = -1; 605 buddy = mb_find_buddy(e4b, 0, &max); 606 for (i = 0; i < max; i++) { 607 if (!mb_test_bit(i, buddy)) { 608 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 609 if (fstart == -1) { 610 fragments++; 611 fstart = i; 612 } 613 continue; 614 } 615 fstart = -1; 616 /* check used bits only */ 617 for (j = 0; j < e4b->bd_blkbits + 1; j++) { 618 buddy2 = mb_find_buddy(e4b, j, &max2); 619 k = i >> j; 620 MB_CHECK_ASSERT(k < max2); 621 MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 622 } 623 } 624 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 625 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 626 627 grp = ext4_get_group_info(sb, e4b->bd_group); 628 list_for_each(cur, &grp->bb_prealloc_list) { 629 ext4_group_t groupnr; 630 struct ext4_prealloc_space *pa; 631 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 632 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 633 MB_CHECK_ASSERT(groupnr == e4b->bd_group); 634 for (i = 0; i < pa->pa_len; i++) 635 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 636 } 637 return 0; 638 } 639 #undef MB_CHECK_ASSERT 640 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 641 __FILE__, __func__, __LINE__) 642 #else 643 #define mb_check_buddy(e4b) 644 #endif 645 646 /* 647 * Divide blocks started from @first with length @len into 648 * smaller chunks with power of 2 blocks. 649 * Clear the bits in bitmap which the blocks of the chunk(s) covered, 650 * then increase bb_counters[] for corresponded chunk size. 651 */ 652 static void ext4_mb_mark_free_simple(struct super_block *sb, 653 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 654 struct ext4_group_info *grp) 655 { 656 struct ext4_sb_info *sbi = EXT4_SB(sb); 657 ext4_grpblk_t min; 658 ext4_grpblk_t max; 659 ext4_grpblk_t chunk; 660 unsigned int border; 661 662 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 663 664 border = 2 << sb->s_blocksize_bits; 665 666 while (len > 0) { 667 /* find how many blocks can be covered since this position */ 668 max = ffs(first | border) - 1; 669 670 /* find how many blocks of power 2 we need to mark */ 671 min = fls(len) - 1; 672 673 if (max < min) 674 min = max; 675 chunk = 1 << min; 676 677 /* mark multiblock chunks only */ 678 grp->bb_counters[min]++; 679 if (min > 0) 680 mb_clear_bit(first >> min, 681 buddy + sbi->s_mb_offsets[min]); 682 683 len -= chunk; 684 first += chunk; 685 } 686 } 687 688 /* 689 * Cache the order of the largest free extent we have available in this block 690 * group. 691 */ 692 static void 693 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 694 { 695 int i; 696 int bits; 697 698 grp->bb_largest_free_order = -1; /* uninit */ 699 700 bits = sb->s_blocksize_bits + 1; 701 for (i = bits; i >= 0; i--) { 702 if (grp->bb_counters[i] > 0) { 703 grp->bb_largest_free_order = i; 704 break; 705 } 706 } 707 } 708 709 static noinline_for_stack 710 void ext4_mb_generate_buddy(struct super_block *sb, 711 void *buddy, void *bitmap, ext4_group_t group) 712 { 713 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 714 struct ext4_sb_info *sbi = EXT4_SB(sb); 715 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 716 ext4_grpblk_t i = 0; 717 ext4_grpblk_t first; 718 ext4_grpblk_t len; 719 unsigned free = 0; 720 unsigned fragments = 0; 721 unsigned long long period = get_cycles(); 722 723 /* initialize buddy from bitmap which is aggregation 724 * of on-disk bitmap and preallocations */ 725 i = mb_find_next_zero_bit(bitmap, max, 0); 726 grp->bb_first_free = i; 727 while (i < max) { 728 fragments++; 729 first = i; 730 i = mb_find_next_bit(bitmap, max, i); 731 len = i - first; 732 free += len; 733 if (len > 1) 734 ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 735 else 736 grp->bb_counters[0]++; 737 if (i < max) 738 i = mb_find_next_zero_bit(bitmap, max, i); 739 } 740 grp->bb_fragments = fragments; 741 742 if (free != grp->bb_free) { 743 ext4_grp_locked_error(sb, group, 0, 0, 744 "block bitmap and bg descriptor " 745 "inconsistent: %u vs %u free clusters", 746 free, grp->bb_free); 747 /* 748 * If we intend to continue, we consider group descriptor 749 * corrupt and update bb_free using bitmap value 750 */ 751 grp->bb_free = free; 752 ext4_mark_group_bitmap_corrupted(sb, group, 753 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 754 } 755 mb_set_largest_free_order(sb, grp); 756 757 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 758 759 period = get_cycles() - period; 760 spin_lock(&sbi->s_bal_lock); 761 sbi->s_mb_buddies_generated++; 762 sbi->s_mb_generation_time += period; 763 spin_unlock(&sbi->s_bal_lock); 764 } 765 766 static void mb_regenerate_buddy(struct ext4_buddy *e4b) 767 { 768 int count; 769 int order = 1; 770 void *buddy; 771 772 while ((buddy = mb_find_buddy(e4b, order++, &count))) { 773 ext4_set_bits(buddy, 0, count); 774 } 775 e4b->bd_info->bb_fragments = 0; 776 memset(e4b->bd_info->bb_counters, 0, 777 sizeof(*e4b->bd_info->bb_counters) * 778 (e4b->bd_sb->s_blocksize_bits + 2)); 779 780 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, 781 e4b->bd_bitmap, e4b->bd_group); 782 } 783 784 /* The buddy information is attached the buddy cache inode 785 * for convenience. The information regarding each group 786 * is loaded via ext4_mb_load_buddy. The information involve 787 * block bitmap and buddy information. The information are 788 * stored in the inode as 789 * 790 * { page } 791 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 792 * 793 * 794 * one block each for bitmap and buddy information. 795 * So for each group we take up 2 blocks. A page can 796 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. 797 * So it can have information regarding groups_per_page which 798 * is blocks_per_page/2 799 * 800 * Locking note: This routine takes the block group lock of all groups 801 * for this page; do not hold this lock when calling this routine! 802 */ 803 804 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) 805 { 806 ext4_group_t ngroups; 807 int blocksize; 808 int blocks_per_page; 809 int groups_per_page; 810 int err = 0; 811 int i; 812 ext4_group_t first_group, group; 813 int first_block; 814 struct super_block *sb; 815 struct buffer_head *bhs; 816 struct buffer_head **bh = NULL; 817 struct inode *inode; 818 char *data; 819 char *bitmap; 820 struct ext4_group_info *grinfo; 821 822 mb_debug(1, "init page %lu\n", page->index); 823 824 inode = page->mapping->host; 825 sb = inode->i_sb; 826 ngroups = ext4_get_groups_count(sb); 827 blocksize = i_blocksize(inode); 828 blocks_per_page = PAGE_SIZE / blocksize; 829 830 groups_per_page = blocks_per_page >> 1; 831 if (groups_per_page == 0) 832 groups_per_page = 1; 833 834 /* allocate buffer_heads to read bitmaps */ 835 if (groups_per_page > 1) { 836 i = sizeof(struct buffer_head *) * groups_per_page; 837 bh = kzalloc(i, gfp); 838 if (bh == NULL) { 839 err = -ENOMEM; 840 goto out; 841 } 842 } else 843 bh = &bhs; 844 845 first_group = page->index * blocks_per_page / 2; 846 847 /* read all groups the page covers into the cache */ 848 for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 849 if (group >= ngroups) 850 break; 851 852 grinfo = ext4_get_group_info(sb, group); 853 /* 854 * If page is uptodate then we came here after online resize 855 * which added some new uninitialized group info structs, so 856 * we must skip all initialized uptodate buddies on the page, 857 * which may be currently in use by an allocating task. 858 */ 859 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { 860 bh[i] = NULL; 861 continue; 862 } 863 bh[i] = ext4_read_block_bitmap_nowait(sb, group); 864 if (IS_ERR(bh[i])) { 865 err = PTR_ERR(bh[i]); 866 bh[i] = NULL; 867 goto out; 868 } 869 mb_debug(1, "read bitmap for group %u\n", group); 870 } 871 872 /* wait for I/O completion */ 873 for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 874 int err2; 875 876 if (!bh[i]) 877 continue; 878 err2 = ext4_wait_block_bitmap(sb, group, bh[i]); 879 if (!err) 880 err = err2; 881 } 882 883 first_block = page->index * blocks_per_page; 884 for (i = 0; i < blocks_per_page; i++) { 885 group = (first_block + i) >> 1; 886 if (group >= ngroups) 887 break; 888 889 if (!bh[group - first_group]) 890 /* skip initialized uptodate buddy */ 891 continue; 892 893 if (!buffer_verified(bh[group - first_group])) 894 /* Skip faulty bitmaps */ 895 continue; 896 err = 0; 897 898 /* 899 * data carry information regarding this 900 * particular group in the format specified 901 * above 902 * 903 */ 904 data = page_address(page) + (i * blocksize); 905 bitmap = bh[group - first_group]->b_data; 906 907 /* 908 * We place the buddy block and bitmap block 909 * close together 910 */ 911 if ((first_block + i) & 1) { 912 /* this is block of buddy */ 913 BUG_ON(incore == NULL); 914 mb_debug(1, "put buddy for group %u in page %lu/%x\n", 915 group, page->index, i * blocksize); 916 trace_ext4_mb_buddy_bitmap_load(sb, group); 917 grinfo = ext4_get_group_info(sb, group); 918 grinfo->bb_fragments = 0; 919 memset(grinfo->bb_counters, 0, 920 sizeof(*grinfo->bb_counters) * 921 (sb->s_blocksize_bits+2)); 922 /* 923 * incore got set to the group block bitmap below 924 */ 925 ext4_lock_group(sb, group); 926 /* init the buddy */ 927 memset(data, 0xff, blocksize); 928 ext4_mb_generate_buddy(sb, data, incore, group); 929 ext4_unlock_group(sb, group); 930 incore = NULL; 931 } else { 932 /* this is block of bitmap */ 933 BUG_ON(incore != NULL); 934 mb_debug(1, "put bitmap for group %u in page %lu/%x\n", 935 group, page->index, i * blocksize); 936 trace_ext4_mb_bitmap_load(sb, group); 937 938 /* see comments in ext4_mb_put_pa() */ 939 ext4_lock_group(sb, group); 940 memcpy(data, bitmap, blocksize); 941 942 /* mark all preallocated blks used in in-core bitmap */ 943 ext4_mb_generate_from_pa(sb, data, group); 944 ext4_mb_generate_from_freelist(sb, data, group); 945 ext4_unlock_group(sb, group); 946 947 /* set incore so that the buddy information can be 948 * generated using this 949 */ 950 incore = data; 951 } 952 } 953 SetPageUptodate(page); 954 955 out: 956 if (bh) { 957 for (i = 0; i < groups_per_page; i++) 958 brelse(bh[i]); 959 if (bh != &bhs) 960 kfree(bh); 961 } 962 return err; 963 } 964 965 /* 966 * Lock the buddy and bitmap pages. This make sure other parallel init_group 967 * on the same buddy page doesn't happen whild holding the buddy page lock. 968 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 969 * are on the same page e4b->bd_buddy_page is NULL and return value is 0. 970 */ 971 static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 972 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) 973 { 974 struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 975 int block, pnum, poff; 976 int blocks_per_page; 977 struct page *page; 978 979 e4b->bd_buddy_page = NULL; 980 e4b->bd_bitmap_page = NULL; 981 982 blocks_per_page = PAGE_SIZE / sb->s_blocksize; 983 /* 984 * the buddy cache inode stores the block bitmap 985 * and buddy information in consecutive blocks. 986 * So for each group we need two blocks. 987 */ 988 block = group * 2; 989 pnum = block / blocks_per_page; 990 poff = block % blocks_per_page; 991 page = find_or_create_page(inode->i_mapping, pnum, gfp); 992 if (!page) 993 return -ENOMEM; 994 BUG_ON(page->mapping != inode->i_mapping); 995 e4b->bd_bitmap_page = page; 996 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 997 998 if (blocks_per_page >= 2) { 999 /* buddy and bitmap are on the same page */ 1000 return 0; 1001 } 1002 1003 block++; 1004 pnum = block / blocks_per_page; 1005 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1006 if (!page) 1007 return -ENOMEM; 1008 BUG_ON(page->mapping != inode->i_mapping); 1009 e4b->bd_buddy_page = page; 1010 return 0; 1011 } 1012 1013 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 1014 { 1015 if (e4b->bd_bitmap_page) { 1016 unlock_page(e4b->bd_bitmap_page); 1017 put_page(e4b->bd_bitmap_page); 1018 } 1019 if (e4b->bd_buddy_page) { 1020 unlock_page(e4b->bd_buddy_page); 1021 put_page(e4b->bd_buddy_page); 1022 } 1023 } 1024 1025 /* 1026 * Locking note: This routine calls ext4_mb_init_cache(), which takes the 1027 * block group lock of all groups for this page; do not hold the BG lock when 1028 * calling this routine! 1029 */ 1030 static noinline_for_stack 1031 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) 1032 { 1033 1034 struct ext4_group_info *this_grp; 1035 struct ext4_buddy e4b; 1036 struct page *page; 1037 int ret = 0; 1038 1039 might_sleep(); 1040 mb_debug(1, "init group %u\n", group); 1041 this_grp = ext4_get_group_info(sb, group); 1042 /* 1043 * This ensures that we don't reinit the buddy cache 1044 * page which map to the group from which we are already 1045 * allocating. If we are looking at the buddy cache we would 1046 * have taken a reference using ext4_mb_load_buddy and that 1047 * would have pinned buddy page to page cache. 1048 * The call to ext4_mb_get_buddy_page_lock will mark the 1049 * page accessed. 1050 */ 1051 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); 1052 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1053 /* 1054 * somebody initialized the group 1055 * return without doing anything 1056 */ 1057 goto err; 1058 } 1059 1060 page = e4b.bd_bitmap_page; 1061 ret = ext4_mb_init_cache(page, NULL, gfp); 1062 if (ret) 1063 goto err; 1064 if (!PageUptodate(page)) { 1065 ret = -EIO; 1066 goto err; 1067 } 1068 1069 if (e4b.bd_buddy_page == NULL) { 1070 /* 1071 * If both the bitmap and buddy are in 1072 * the same page we don't need to force 1073 * init the buddy 1074 */ 1075 ret = 0; 1076 goto err; 1077 } 1078 /* init buddy cache */ 1079 page = e4b.bd_buddy_page; 1080 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); 1081 if (ret) 1082 goto err; 1083 if (!PageUptodate(page)) { 1084 ret = -EIO; 1085 goto err; 1086 } 1087 err: 1088 ext4_mb_put_buddy_page_lock(&e4b); 1089 return ret; 1090 } 1091 1092 /* 1093 * Locking note: This routine calls ext4_mb_init_cache(), which takes the 1094 * block group lock of all groups for this page; do not hold the BG lock when 1095 * calling this routine! 1096 */ 1097 static noinline_for_stack int 1098 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, 1099 struct ext4_buddy *e4b, gfp_t gfp) 1100 { 1101 int blocks_per_page; 1102 int block; 1103 int pnum; 1104 int poff; 1105 struct page *page; 1106 int ret; 1107 struct ext4_group_info *grp; 1108 struct ext4_sb_info *sbi = EXT4_SB(sb); 1109 struct inode *inode = sbi->s_buddy_cache; 1110 1111 might_sleep(); 1112 mb_debug(1, "load group %u\n", group); 1113 1114 blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1115 grp = ext4_get_group_info(sb, group); 1116 1117 e4b->bd_blkbits = sb->s_blocksize_bits; 1118 e4b->bd_info = grp; 1119 e4b->bd_sb = sb; 1120 e4b->bd_group = group; 1121 e4b->bd_buddy_page = NULL; 1122 e4b->bd_bitmap_page = NULL; 1123 1124 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1125 /* 1126 * we need full data about the group 1127 * to make a good selection 1128 */ 1129 ret = ext4_mb_init_group(sb, group, gfp); 1130 if (ret) 1131 return ret; 1132 } 1133 1134 /* 1135 * the buddy cache inode stores the block bitmap 1136 * and buddy information in consecutive blocks. 1137 * So for each group we need two blocks. 1138 */ 1139 block = group * 2; 1140 pnum = block / blocks_per_page; 1141 poff = block % blocks_per_page; 1142 1143 /* we could use find_or_create_page(), but it locks page 1144 * what we'd like to avoid in fast path ... */ 1145 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1146 if (page == NULL || !PageUptodate(page)) { 1147 if (page) 1148 /* 1149 * drop the page reference and try 1150 * to get the page with lock. If we 1151 * are not uptodate that implies 1152 * somebody just created the page but 1153 * is yet to initialize the same. So 1154 * wait for it to initialize. 1155 */ 1156 put_page(page); 1157 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1158 if (page) { 1159 BUG_ON(page->mapping != inode->i_mapping); 1160 if (!PageUptodate(page)) { 1161 ret = ext4_mb_init_cache(page, NULL, gfp); 1162 if (ret) { 1163 unlock_page(page); 1164 goto err; 1165 } 1166 mb_cmp_bitmaps(e4b, page_address(page) + 1167 (poff * sb->s_blocksize)); 1168 } 1169 unlock_page(page); 1170 } 1171 } 1172 if (page == NULL) { 1173 ret = -ENOMEM; 1174 goto err; 1175 } 1176 if (!PageUptodate(page)) { 1177 ret = -EIO; 1178 goto err; 1179 } 1180 1181 /* Pages marked accessed already */ 1182 e4b->bd_bitmap_page = page; 1183 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1184 1185 block++; 1186 pnum = block / blocks_per_page; 1187 poff = block % blocks_per_page; 1188 1189 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1190 if (page == NULL || !PageUptodate(page)) { 1191 if (page) 1192 put_page(page); 1193 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1194 if (page) { 1195 BUG_ON(page->mapping != inode->i_mapping); 1196 if (!PageUptodate(page)) { 1197 ret = ext4_mb_init_cache(page, e4b->bd_bitmap, 1198 gfp); 1199 if (ret) { 1200 unlock_page(page); 1201 goto err; 1202 } 1203 } 1204 unlock_page(page); 1205 } 1206 } 1207 if (page == NULL) { 1208 ret = -ENOMEM; 1209 goto err; 1210 } 1211 if (!PageUptodate(page)) { 1212 ret = -EIO; 1213 goto err; 1214 } 1215 1216 /* Pages marked accessed already */ 1217 e4b->bd_buddy_page = page; 1218 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1219 1220 BUG_ON(e4b->bd_bitmap_page == NULL); 1221 BUG_ON(e4b->bd_buddy_page == NULL); 1222 1223 return 0; 1224 1225 err: 1226 if (page) 1227 put_page(page); 1228 if (e4b->bd_bitmap_page) 1229 put_page(e4b->bd_bitmap_page); 1230 if (e4b->bd_buddy_page) 1231 put_page(e4b->bd_buddy_page); 1232 e4b->bd_buddy = NULL; 1233 e4b->bd_bitmap = NULL; 1234 return ret; 1235 } 1236 1237 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1238 struct ext4_buddy *e4b) 1239 { 1240 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); 1241 } 1242 1243 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1244 { 1245 if (e4b->bd_bitmap_page) 1246 put_page(e4b->bd_bitmap_page); 1247 if (e4b->bd_buddy_page) 1248 put_page(e4b->bd_buddy_page); 1249 } 1250 1251 1252 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1253 { 1254 int order = 1; 1255 int bb_incr = 1 << (e4b->bd_blkbits - 1); 1256 void *bb; 1257 1258 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1259 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1260 1261 bb = e4b->bd_buddy; 1262 while (order <= e4b->bd_blkbits + 1) { 1263 block = block >> 1; 1264 if (!mb_test_bit(block, bb)) { 1265 /* this block is part of buddy of order 'order' */ 1266 return order; 1267 } 1268 bb += bb_incr; 1269 bb_incr >>= 1; 1270 order++; 1271 } 1272 return 0; 1273 } 1274 1275 static void mb_clear_bits(void *bm, int cur, int len) 1276 { 1277 __u32 *addr; 1278 1279 len = cur + len; 1280 while (cur < len) { 1281 if ((cur & 31) == 0 && (len - cur) >= 32) { 1282 /* fast path: clear whole word at once */ 1283 addr = bm + (cur >> 3); 1284 *addr = 0; 1285 cur += 32; 1286 continue; 1287 } 1288 mb_clear_bit(cur, bm); 1289 cur++; 1290 } 1291 } 1292 1293 /* clear bits in given range 1294 * will return first found zero bit if any, -1 otherwise 1295 */ 1296 static int mb_test_and_clear_bits(void *bm, int cur, int len) 1297 { 1298 __u32 *addr; 1299 int zero_bit = -1; 1300 1301 len = cur + len; 1302 while (cur < len) { 1303 if ((cur & 31) == 0 && (len - cur) >= 32) { 1304 /* fast path: clear whole word at once */ 1305 addr = bm + (cur >> 3); 1306 if (*addr != (__u32)(-1) && zero_bit == -1) 1307 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); 1308 *addr = 0; 1309 cur += 32; 1310 continue; 1311 } 1312 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) 1313 zero_bit = cur; 1314 cur++; 1315 } 1316 1317 return zero_bit; 1318 } 1319 1320 void ext4_set_bits(void *bm, int cur, int len) 1321 { 1322 __u32 *addr; 1323 1324 len = cur + len; 1325 while (cur < len) { 1326 if ((cur & 31) == 0 && (len - cur) >= 32) { 1327 /* fast path: set whole word at once */ 1328 addr = bm + (cur >> 3); 1329 *addr = 0xffffffff; 1330 cur += 32; 1331 continue; 1332 } 1333 mb_set_bit(cur, bm); 1334 cur++; 1335 } 1336 } 1337 1338 /* 1339 * _________________________________________________________________ */ 1340 1341 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) 1342 { 1343 if (mb_test_bit(*bit + side, bitmap)) { 1344 mb_clear_bit(*bit, bitmap); 1345 (*bit) -= side; 1346 return 1; 1347 } 1348 else { 1349 (*bit) += side; 1350 mb_set_bit(*bit, bitmap); 1351 return -1; 1352 } 1353 } 1354 1355 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) 1356 { 1357 int max; 1358 int order = 1; 1359 void *buddy = mb_find_buddy(e4b, order, &max); 1360 1361 while (buddy) { 1362 void *buddy2; 1363 1364 /* Bits in range [first; last] are known to be set since 1365 * corresponding blocks were allocated. Bits in range 1366 * (first; last) will stay set because they form buddies on 1367 * upper layer. We just deal with borders if they don't 1368 * align with upper layer and then go up. 1369 * Releasing entire group is all about clearing 1370 * single bit of highest order buddy. 1371 */ 1372 1373 /* Example: 1374 * --------------------------------- 1375 * | 1 | 1 | 1 | 1 | 1376 * --------------------------------- 1377 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1378 * --------------------------------- 1379 * 0 1 2 3 4 5 6 7 1380 * \_____________________/ 1381 * 1382 * Neither [1] nor [6] is aligned to above layer. 1383 * Left neighbour [0] is free, so mark it busy, 1384 * decrease bb_counters and extend range to 1385 * [0; 6] 1386 * Right neighbour [7] is busy. It can't be coaleasced with [6], so 1387 * mark [6] free, increase bb_counters and shrink range to 1388 * [0; 5]. 1389 * Then shift range to [0; 2], go up and do the same. 1390 */ 1391 1392 1393 if (first & 1) 1394 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); 1395 if (!(last & 1)) 1396 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); 1397 if (first > last) 1398 break; 1399 order++; 1400 1401 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { 1402 mb_clear_bits(buddy, first, last - first + 1); 1403 e4b->bd_info->bb_counters[order - 1] += last - first + 1; 1404 break; 1405 } 1406 first >>= 1; 1407 last >>= 1; 1408 buddy = buddy2; 1409 } 1410 } 1411 1412 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1413 int first, int count) 1414 { 1415 int left_is_free = 0; 1416 int right_is_free = 0; 1417 int block; 1418 int last = first + count - 1; 1419 struct super_block *sb = e4b->bd_sb; 1420 1421 if (WARN_ON(count == 0)) 1422 return; 1423 BUG_ON(last >= (sb->s_blocksize << 3)); 1424 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1425 /* Don't bother if the block group is corrupt. */ 1426 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 1427 return; 1428 1429 mb_check_buddy(e4b); 1430 mb_free_blocks_double(inode, e4b, first, count); 1431 1432 e4b->bd_info->bb_free += count; 1433 if (first < e4b->bd_info->bb_first_free) 1434 e4b->bd_info->bb_first_free = first; 1435 1436 /* access memory sequentially: check left neighbour, 1437 * clear range and then check right neighbour 1438 */ 1439 if (first != 0) 1440 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); 1441 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); 1442 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) 1443 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); 1444 1445 if (unlikely(block != -1)) { 1446 struct ext4_sb_info *sbi = EXT4_SB(sb); 1447 ext4_fsblk_t blocknr; 1448 1449 blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 1450 blocknr += EXT4_C2B(sbi, block); 1451 ext4_grp_locked_error(sb, e4b->bd_group, 1452 inode ? inode->i_ino : 0, 1453 blocknr, 1454 "freeing already freed block " 1455 "(bit %u); block bitmap corrupt.", 1456 block); 1457 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 1458 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1459 mb_regenerate_buddy(e4b); 1460 goto done; 1461 } 1462 1463 /* let's maintain fragments counter */ 1464 if (left_is_free && right_is_free) 1465 e4b->bd_info->bb_fragments--; 1466 else if (!left_is_free && !right_is_free) 1467 e4b->bd_info->bb_fragments++; 1468 1469 /* buddy[0] == bd_bitmap is a special case, so handle 1470 * it right away and let mb_buddy_mark_free stay free of 1471 * zero order checks. 1472 * Check if neighbours are to be coaleasced, 1473 * adjust bitmap bb_counters and borders appropriately. 1474 */ 1475 if (first & 1) { 1476 first += !left_is_free; 1477 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; 1478 } 1479 if (!(last & 1)) { 1480 last -= !right_is_free; 1481 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; 1482 } 1483 1484 if (first <= last) 1485 mb_buddy_mark_free(e4b, first >> 1, last >> 1); 1486 1487 done: 1488 mb_set_largest_free_order(sb, e4b->bd_info); 1489 mb_check_buddy(e4b); 1490 } 1491 1492 static int mb_find_extent(struct ext4_buddy *e4b, int block, 1493 int needed, struct ext4_free_extent *ex) 1494 { 1495 int next = block; 1496 int max, order; 1497 void *buddy; 1498 1499 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1500 BUG_ON(ex == NULL); 1501 1502 buddy = mb_find_buddy(e4b, 0, &max); 1503 BUG_ON(buddy == NULL); 1504 BUG_ON(block >= max); 1505 if (mb_test_bit(block, buddy)) { 1506 ex->fe_len = 0; 1507 ex->fe_start = 0; 1508 ex->fe_group = 0; 1509 return 0; 1510 } 1511 1512 /* find actual order */ 1513 order = mb_find_order_for_block(e4b, block); 1514 block = block >> order; 1515 1516 ex->fe_len = 1 << order; 1517 ex->fe_start = block << order; 1518 ex->fe_group = e4b->bd_group; 1519 1520 /* calc difference from given start */ 1521 next = next - ex->fe_start; 1522 ex->fe_len -= next; 1523 ex->fe_start += next; 1524 1525 while (needed > ex->fe_len && 1526 mb_find_buddy(e4b, order, &max)) { 1527 1528 if (block + 1 >= max) 1529 break; 1530 1531 next = (block + 1) * (1 << order); 1532 if (mb_test_bit(next, e4b->bd_bitmap)) 1533 break; 1534 1535 order = mb_find_order_for_block(e4b, next); 1536 1537 block = next >> order; 1538 ex->fe_len += 1 << order; 1539 } 1540 1541 if (ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))) { 1542 /* Should never happen! (but apparently sometimes does?!?) */ 1543 WARN_ON(1); 1544 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent " 1545 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", 1546 block, order, needed, ex->fe_group, ex->fe_start, 1547 ex->fe_len, ex->fe_logical); 1548 ex->fe_len = 0; 1549 ex->fe_start = 0; 1550 ex->fe_group = 0; 1551 } 1552 return ex->fe_len; 1553 } 1554 1555 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1556 { 1557 int ord; 1558 int mlen = 0; 1559 int max = 0; 1560 int cur; 1561 int start = ex->fe_start; 1562 int len = ex->fe_len; 1563 unsigned ret = 0; 1564 int len0 = len; 1565 void *buddy; 1566 1567 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1568 BUG_ON(e4b->bd_group != ex->fe_group); 1569 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1570 mb_check_buddy(e4b); 1571 mb_mark_used_double(e4b, start, len); 1572 1573 e4b->bd_info->bb_free -= len; 1574 if (e4b->bd_info->bb_first_free == start) 1575 e4b->bd_info->bb_first_free += len; 1576 1577 /* let's maintain fragments counter */ 1578 if (start != 0) 1579 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 1580 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1581 max = !mb_test_bit(start + len, e4b->bd_bitmap); 1582 if (mlen && max) 1583 e4b->bd_info->bb_fragments++; 1584 else if (!mlen && !max) 1585 e4b->bd_info->bb_fragments--; 1586 1587 /* let's maintain buddy itself */ 1588 while (len) { 1589 ord = mb_find_order_for_block(e4b, start); 1590 1591 if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1592 /* the whole chunk may be allocated at once! */ 1593 mlen = 1 << ord; 1594 buddy = mb_find_buddy(e4b, ord, &max); 1595 BUG_ON((start >> ord) >= max); 1596 mb_set_bit(start >> ord, buddy); 1597 e4b->bd_info->bb_counters[ord]--; 1598 start += mlen; 1599 len -= mlen; 1600 BUG_ON(len < 0); 1601 continue; 1602 } 1603 1604 /* store for history */ 1605 if (ret == 0) 1606 ret = len | (ord << 16); 1607 1608 /* we have to split large buddy */ 1609 BUG_ON(ord <= 0); 1610 buddy = mb_find_buddy(e4b, ord, &max); 1611 mb_set_bit(start >> ord, buddy); 1612 e4b->bd_info->bb_counters[ord]--; 1613 1614 ord--; 1615 cur = (start >> ord) & ~1U; 1616 buddy = mb_find_buddy(e4b, ord, &max); 1617 mb_clear_bit(cur, buddy); 1618 mb_clear_bit(cur + 1, buddy); 1619 e4b->bd_info->bb_counters[ord]++; 1620 e4b->bd_info->bb_counters[ord]++; 1621 } 1622 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 1623 1624 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 1625 mb_check_buddy(e4b); 1626 1627 return ret; 1628 } 1629 1630 /* 1631 * Must be called under group lock! 1632 */ 1633 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1634 struct ext4_buddy *e4b) 1635 { 1636 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1637 int ret; 1638 1639 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1640 BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1641 1642 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1643 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1644 ret = mb_mark_used(e4b, &ac->ac_b_ex); 1645 1646 /* preallocation can change ac_b_ex, thus we store actually 1647 * allocated blocks for history */ 1648 ac->ac_f_ex = ac->ac_b_ex; 1649 1650 ac->ac_status = AC_STATUS_FOUND; 1651 ac->ac_tail = ret & 0xffff; 1652 ac->ac_buddy = ret >> 16; 1653 1654 /* 1655 * take the page reference. We want the page to be pinned 1656 * so that we don't get a ext4_mb_init_cache_call for this 1657 * group until we update the bitmap. That would mean we 1658 * double allocate blocks. The reference is dropped 1659 * in ext4_mb_release_context 1660 */ 1661 ac->ac_bitmap_page = e4b->bd_bitmap_page; 1662 get_page(ac->ac_bitmap_page); 1663 ac->ac_buddy_page = e4b->bd_buddy_page; 1664 get_page(ac->ac_buddy_page); 1665 /* store last allocated for subsequent stream allocation */ 1666 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1667 spin_lock(&sbi->s_md_lock); 1668 sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1669 sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1670 spin_unlock(&sbi->s_md_lock); 1671 } 1672 } 1673 1674 /* 1675 * regular allocator, for general purposes allocation 1676 */ 1677 1678 static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1679 struct ext4_buddy *e4b, 1680 int finish_group) 1681 { 1682 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1683 struct ext4_free_extent *bex = &ac->ac_b_ex; 1684 struct ext4_free_extent *gex = &ac->ac_g_ex; 1685 struct ext4_free_extent ex; 1686 int max; 1687 1688 if (ac->ac_status == AC_STATUS_FOUND) 1689 return; 1690 /* 1691 * We don't want to scan for a whole year 1692 */ 1693 if (ac->ac_found > sbi->s_mb_max_to_scan && 1694 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1695 ac->ac_status = AC_STATUS_BREAK; 1696 return; 1697 } 1698 1699 /* 1700 * Haven't found good chunk so far, let's continue 1701 */ 1702 if (bex->fe_len < gex->fe_len) 1703 return; 1704 1705 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1706 && bex->fe_group == e4b->bd_group) { 1707 /* recheck chunk's availability - we don't know 1708 * when it was found (within this lock-unlock 1709 * period or not) */ 1710 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); 1711 if (max >= gex->fe_len) { 1712 ext4_mb_use_best_found(ac, e4b); 1713 return; 1714 } 1715 } 1716 } 1717 1718 /* 1719 * The routine checks whether found extent is good enough. If it is, 1720 * then the extent gets marked used and flag is set to the context 1721 * to stop scanning. Otherwise, the extent is compared with the 1722 * previous found extent and if new one is better, then it's stored 1723 * in the context. Later, the best found extent will be used, if 1724 * mballoc can't find good enough extent. 1725 * 1726 * FIXME: real allocation policy is to be designed yet! 1727 */ 1728 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1729 struct ext4_free_extent *ex, 1730 struct ext4_buddy *e4b) 1731 { 1732 struct ext4_free_extent *bex = &ac->ac_b_ex; 1733 struct ext4_free_extent *gex = &ac->ac_g_ex; 1734 1735 BUG_ON(ex->fe_len <= 0); 1736 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 1737 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 1738 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1739 1740 ac->ac_found++; 1741 1742 /* 1743 * The special case - take what you catch first 1744 */ 1745 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1746 *bex = *ex; 1747 ext4_mb_use_best_found(ac, e4b); 1748 return; 1749 } 1750 1751 /* 1752 * Let's check whether the chuck is good enough 1753 */ 1754 if (ex->fe_len == gex->fe_len) { 1755 *bex = *ex; 1756 ext4_mb_use_best_found(ac, e4b); 1757 return; 1758 } 1759 1760 /* 1761 * If this is first found extent, just store it in the context 1762 */ 1763 if (bex->fe_len == 0) { 1764 *bex = *ex; 1765 return; 1766 } 1767 1768 /* 1769 * If new found extent is better, store it in the context 1770 */ 1771 if (bex->fe_len < gex->fe_len) { 1772 /* if the request isn't satisfied, any found extent 1773 * larger than previous best one is better */ 1774 if (ex->fe_len > bex->fe_len) 1775 *bex = *ex; 1776 } else if (ex->fe_len > gex->fe_len) { 1777 /* if the request is satisfied, then we try to find 1778 * an extent that still satisfy the request, but is 1779 * smaller than previous one */ 1780 if (ex->fe_len < bex->fe_len) 1781 *bex = *ex; 1782 } 1783 1784 ext4_mb_check_limits(ac, e4b, 0); 1785 } 1786 1787 static noinline_for_stack 1788 int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1789 struct ext4_buddy *e4b) 1790 { 1791 struct ext4_free_extent ex = ac->ac_b_ex; 1792 ext4_group_t group = ex.fe_group; 1793 int max; 1794 int err; 1795 1796 BUG_ON(ex.fe_len <= 0); 1797 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1798 if (err) 1799 return err; 1800 1801 ext4_lock_group(ac->ac_sb, group); 1802 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); 1803 1804 if (max > 0) { 1805 ac->ac_b_ex = ex; 1806 ext4_mb_use_best_found(ac, e4b); 1807 } 1808 1809 ext4_unlock_group(ac->ac_sb, group); 1810 ext4_mb_unload_buddy(e4b); 1811 1812 return 0; 1813 } 1814 1815 static noinline_for_stack 1816 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1817 struct ext4_buddy *e4b) 1818 { 1819 ext4_group_t group = ac->ac_g_ex.fe_group; 1820 int max; 1821 int err; 1822 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1823 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1824 struct ext4_free_extent ex; 1825 1826 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1827 return 0; 1828 if (grp->bb_free == 0) 1829 return 0; 1830 1831 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1832 if (err) 1833 return err; 1834 1835 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { 1836 ext4_mb_unload_buddy(e4b); 1837 return 0; 1838 } 1839 1840 ext4_lock_group(ac->ac_sb, group); 1841 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, 1842 ac->ac_g_ex.fe_len, &ex); 1843 ex.fe_logical = 0xDEADFA11; /* debug value */ 1844 1845 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1846 ext4_fsblk_t start; 1847 1848 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + 1849 ex.fe_start; 1850 /* use do_div to get remainder (would be 64-bit modulo) */ 1851 if (do_div(start, sbi->s_stripe) == 0) { 1852 ac->ac_found++; 1853 ac->ac_b_ex = ex; 1854 ext4_mb_use_best_found(ac, e4b); 1855 } 1856 } else if (max >= ac->ac_g_ex.fe_len) { 1857 BUG_ON(ex.fe_len <= 0); 1858 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1859 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1860 ac->ac_found++; 1861 ac->ac_b_ex = ex; 1862 ext4_mb_use_best_found(ac, e4b); 1863 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1864 /* Sometimes, caller may want to merge even small 1865 * number of blocks to an existing extent */ 1866 BUG_ON(ex.fe_len <= 0); 1867 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1868 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1869 ac->ac_found++; 1870 ac->ac_b_ex = ex; 1871 ext4_mb_use_best_found(ac, e4b); 1872 } 1873 ext4_unlock_group(ac->ac_sb, group); 1874 ext4_mb_unload_buddy(e4b); 1875 1876 return 0; 1877 } 1878 1879 /* 1880 * The routine scans buddy structures (not bitmap!) from given order 1881 * to max order and tries to find big enough chunk to satisfy the req 1882 */ 1883 static noinline_for_stack 1884 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1885 struct ext4_buddy *e4b) 1886 { 1887 struct super_block *sb = ac->ac_sb; 1888 struct ext4_group_info *grp = e4b->bd_info; 1889 void *buddy; 1890 int i; 1891 int k; 1892 int max; 1893 1894 BUG_ON(ac->ac_2order <= 0); 1895 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1896 if (grp->bb_counters[i] == 0) 1897 continue; 1898 1899 buddy = mb_find_buddy(e4b, i, &max); 1900 BUG_ON(buddy == NULL); 1901 1902 k = mb_find_next_zero_bit(buddy, max, 0); 1903 BUG_ON(k >= max); 1904 1905 ac->ac_found++; 1906 1907 ac->ac_b_ex.fe_len = 1 << i; 1908 ac->ac_b_ex.fe_start = k << i; 1909 ac->ac_b_ex.fe_group = e4b->bd_group; 1910 1911 ext4_mb_use_best_found(ac, e4b); 1912 1913 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1914 1915 if (EXT4_SB(sb)->s_mb_stats) 1916 atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1917 1918 break; 1919 } 1920 } 1921 1922 /* 1923 * The routine scans the group and measures all found extents. 1924 * In order to optimize scanning, caller must pass number of 1925 * free blocks in the group, so the routine can know upper limit. 1926 */ 1927 static noinline_for_stack 1928 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1929 struct ext4_buddy *e4b) 1930 { 1931 struct super_block *sb = ac->ac_sb; 1932 void *bitmap = e4b->bd_bitmap; 1933 struct ext4_free_extent ex; 1934 int i; 1935 int free; 1936 1937 free = e4b->bd_info->bb_free; 1938 BUG_ON(free <= 0); 1939 1940 i = e4b->bd_info->bb_first_free; 1941 1942 while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1943 i = mb_find_next_zero_bit(bitmap, 1944 EXT4_CLUSTERS_PER_GROUP(sb), i); 1945 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 1946 /* 1947 * IF we have corrupt bitmap, we won't find any 1948 * free blocks even though group info says we 1949 * we have free blocks 1950 */ 1951 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 1952 "%d free clusters as per " 1953 "group info. But bitmap says 0", 1954 free); 1955 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 1956 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1957 break; 1958 } 1959 1960 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); 1961 BUG_ON(ex.fe_len <= 0); 1962 if (free < ex.fe_len) { 1963 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 1964 "%d free clusters as per " 1965 "group info. But got %d blocks", 1966 free, ex.fe_len); 1967 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 1968 EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1969 /* 1970 * The number of free blocks differs. This mostly 1971 * indicate that the bitmap is corrupt. So exit 1972 * without claiming the space. 1973 */ 1974 break; 1975 } 1976 ex.fe_logical = 0xDEADC0DE; /* debug value */ 1977 ext4_mb_measure_extent(ac, &ex, e4b); 1978 1979 i += ex.fe_len; 1980 free -= ex.fe_len; 1981 } 1982 1983 ext4_mb_check_limits(ac, e4b, 1); 1984 } 1985 1986 /* 1987 * This is a special case for storages like raid5 1988 * we try to find stripe-aligned chunks for stripe-size-multiple requests 1989 */ 1990 static noinline_for_stack 1991 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 1992 struct ext4_buddy *e4b) 1993 { 1994 struct super_block *sb = ac->ac_sb; 1995 struct ext4_sb_info *sbi = EXT4_SB(sb); 1996 void *bitmap = e4b->bd_bitmap; 1997 struct ext4_free_extent ex; 1998 ext4_fsblk_t first_group_block; 1999 ext4_fsblk_t a; 2000 ext4_grpblk_t i; 2001 int max; 2002 2003 BUG_ON(sbi->s_stripe == 0); 2004 2005 /* find first stripe-aligned block in group */ 2006 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 2007 2008 a = first_group_block + sbi->s_stripe - 1; 2009 do_div(a, sbi->s_stripe); 2010 i = (a * sbi->s_stripe) - first_group_block; 2011 2012 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 2013 if (!mb_test_bit(i, bitmap)) { 2014 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); 2015 if (max >= sbi->s_stripe) { 2016 ac->ac_found++; 2017 ex.fe_logical = 0xDEADF00D; /* debug value */ 2018 ac->ac_b_ex = ex; 2019 ext4_mb_use_best_found(ac, e4b); 2020 break; 2021 } 2022 } 2023 i += sbi->s_stripe; 2024 } 2025 } 2026 2027 /* 2028 * This is now called BEFORE we load the buddy bitmap. 2029 * Returns either 1 or 0 indicating that the group is either suitable 2030 * for the allocation or not. In addition it can also return negative 2031 * error code when something goes wrong. 2032 */ 2033 static int ext4_mb_good_group(struct ext4_allocation_context *ac, 2034 ext4_group_t group, int cr) 2035 { 2036 unsigned free, fragments; 2037 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 2038 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2039 2040 BUG_ON(cr < 0 || cr >= 4); 2041 2042 free = grp->bb_free; 2043 if (free == 0) 2044 return 0; 2045 if (cr <= 2 && free < ac->ac_g_ex.fe_len) 2046 return 0; 2047 2048 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 2049 return 0; 2050 2051 /* We only do this if the grp has never been initialized */ 2052 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 2053 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS); 2054 if (ret) 2055 return ret; 2056 } 2057 2058 fragments = grp->bb_fragments; 2059 if (fragments == 0) 2060 return 0; 2061 2062 switch (cr) { 2063 case 0: 2064 BUG_ON(ac->ac_2order == 0); 2065 2066 /* Avoid using the first bg of a flexgroup for data files */ 2067 if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 2068 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 2069 ((group % flex_size) == 0)) 2070 return 0; 2071 2072 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || 2073 (free / fragments) >= ac->ac_g_ex.fe_len) 2074 return 1; 2075 2076 if (grp->bb_largest_free_order < ac->ac_2order) 2077 return 0; 2078 2079 return 1; 2080 case 1: 2081 if ((free / fragments) >= ac->ac_g_ex.fe_len) 2082 return 1; 2083 break; 2084 case 2: 2085 if (free >= ac->ac_g_ex.fe_len) 2086 return 1; 2087 break; 2088 case 3: 2089 return 1; 2090 default: 2091 BUG(); 2092 } 2093 2094 return 0; 2095 } 2096 2097 static noinline_for_stack int 2098 ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2099 { 2100 ext4_group_t ngroups, group, i; 2101 int cr; 2102 int err = 0, first_err = 0; 2103 struct ext4_sb_info *sbi; 2104 struct super_block *sb; 2105 struct ext4_buddy e4b; 2106 2107 sb = ac->ac_sb; 2108 sbi = EXT4_SB(sb); 2109 ngroups = ext4_get_groups_count(sb); 2110 /* non-extent files are limited to low blocks/groups */ 2111 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2112 ngroups = sbi->s_blockfile_groups; 2113 2114 BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2115 2116 /* first, try the goal */ 2117 err = ext4_mb_find_by_goal(ac, &e4b); 2118 if (err || ac->ac_status == AC_STATUS_FOUND) 2119 goto out; 2120 2121 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2122 goto out; 2123 2124 /* 2125 * ac->ac2_order is set only if the fe_len is a power of 2 2126 * if ac2_order is set we also set criteria to 0 so that we 2127 * try exact allocation using buddy. 2128 */ 2129 i = fls(ac->ac_g_ex.fe_len); 2130 ac->ac_2order = 0; 2131 /* 2132 * We search using buddy data only if the order of the request 2133 * is greater than equal to the sbi_s_mb_order2_reqs 2134 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2135 * We also support searching for power-of-two requests only for 2136 * requests upto maximum buddy size we have constructed. 2137 */ 2138 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) { 2139 /* 2140 * This should tell if fe_len is exactly power of 2 2141 */ 2142 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 2143 ac->ac_2order = i - 1; 2144 } 2145 2146 /* if stream allocation is enabled, use global goal */ 2147 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2148 /* TBD: may be hot point */ 2149 spin_lock(&sbi->s_md_lock); 2150 ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2151 ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2152 spin_unlock(&sbi->s_md_lock); 2153 } 2154 2155 /* Let's just scan groups to find more-less suitable blocks */ 2156 cr = ac->ac_2order ? 0 : 1; 2157 /* 2158 * cr == 0 try to get exact allocation, 2159 * cr == 3 try to get anything 2160 */ 2161 repeat: 2162 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2163 ac->ac_criteria = cr; 2164 /* 2165 * searching for the right group start 2166 * from the goal value specified 2167 */ 2168 group = ac->ac_g_ex.fe_group; 2169 2170 for (i = 0; i < ngroups; group++, i++) { 2171 int ret = 0; 2172 cond_resched(); 2173 /* 2174 * Artificially restricted ngroups for non-extent 2175 * files makes group > ngroups possible on first loop. 2176 */ 2177 if (group >= ngroups) 2178 group = 0; 2179 2180 /* This now checks without needing the buddy page */ 2181 ret = ext4_mb_good_group(ac, group, cr); 2182 if (ret <= 0) { 2183 if (!first_err) 2184 first_err = ret; 2185 continue; 2186 } 2187 2188 err = ext4_mb_load_buddy(sb, group, &e4b); 2189 if (err) 2190 goto out; 2191 2192 ext4_lock_group(sb, group); 2193 2194 /* 2195 * We need to check again after locking the 2196 * block group 2197 */ 2198 ret = ext4_mb_good_group(ac, group, cr); 2199 if (ret <= 0) { 2200 ext4_unlock_group(sb, group); 2201 ext4_mb_unload_buddy(&e4b); 2202 if (!first_err) 2203 first_err = ret; 2204 continue; 2205 } 2206 2207 ac->ac_groups_scanned++; 2208 if (cr == 0) 2209 ext4_mb_simple_scan_group(ac, &e4b); 2210 else if (cr == 1 && sbi->s_stripe && 2211 !(ac->ac_g_ex.fe_len % sbi->s_stripe)) 2212 ext4_mb_scan_aligned(ac, &e4b); 2213 else 2214 ext4_mb_complex_scan_group(ac, &e4b); 2215 2216 ext4_unlock_group(sb, group); 2217 ext4_mb_unload_buddy(&e4b); 2218 2219 if (ac->ac_status != AC_STATUS_CONTINUE) 2220 break; 2221 } 2222 } 2223 2224 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2225 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2226 /* 2227 * We've been searching too long. Let's try to allocate 2228 * the best chunk we've found so far 2229 */ 2230 2231 ext4_mb_try_best_found(ac, &e4b); 2232 if (ac->ac_status != AC_STATUS_FOUND) { 2233 /* 2234 * Someone more lucky has already allocated it. 2235 * The only thing we can do is just take first 2236 * found block(s) 2237 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2238 */ 2239 ac->ac_b_ex.fe_group = 0; 2240 ac->ac_b_ex.fe_start = 0; 2241 ac->ac_b_ex.fe_len = 0; 2242 ac->ac_status = AC_STATUS_CONTINUE; 2243 ac->ac_flags |= EXT4_MB_HINT_FIRST; 2244 cr = 3; 2245 atomic_inc(&sbi->s_mb_lost_chunks); 2246 goto repeat; 2247 } 2248 } 2249 out: 2250 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) 2251 err = first_err; 2252 return err; 2253 } 2254 2255 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2256 { 2257 struct super_block *sb = PDE_DATA(file_inode(seq->file)); 2258 ext4_group_t group; 2259 2260 if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2261 return NULL; 2262 group = *pos + 1; 2263 return (void *) ((unsigned long) group); 2264 } 2265 2266 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2267 { 2268 struct super_block *sb = PDE_DATA(file_inode(seq->file)); 2269 ext4_group_t group; 2270 2271 ++*pos; 2272 if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2273 return NULL; 2274 group = *pos + 1; 2275 return (void *) ((unsigned long) group); 2276 } 2277 2278 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2279 { 2280 struct super_block *sb = PDE_DATA(file_inode(seq->file)); 2281 ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2282 int i; 2283 int err, buddy_loaded = 0; 2284 struct ext4_buddy e4b; 2285 struct ext4_group_info *grinfo; 2286 unsigned char blocksize_bits = min_t(unsigned char, 2287 sb->s_blocksize_bits, 2288 EXT4_MAX_BLOCK_LOG_SIZE); 2289 struct sg { 2290 struct ext4_group_info info; 2291 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; 2292 } sg; 2293 2294 group--; 2295 if (group == 0) 2296 seq_puts(seq, "#group: free frags first [" 2297 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " 2298 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); 2299 2300 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 2301 sizeof(struct ext4_group_info); 2302 2303 grinfo = ext4_get_group_info(sb, group); 2304 /* Load the group info in memory only if not already loaded. */ 2305 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { 2306 err = ext4_mb_load_buddy(sb, group, &e4b); 2307 if (err) { 2308 seq_printf(seq, "#%-5u: I/O error\n", group); 2309 return 0; 2310 } 2311 buddy_loaded = 1; 2312 } 2313 2314 memcpy(&sg, ext4_get_group_info(sb, group), i); 2315 2316 if (buddy_loaded) 2317 ext4_mb_unload_buddy(&e4b); 2318 2319 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2320 sg.info.bb_fragments, sg.info.bb_first_free); 2321 for (i = 0; i <= 13; i++) 2322 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? 2323 sg.info.bb_counters[i] : 0); 2324 seq_printf(seq, " ]\n"); 2325 2326 return 0; 2327 } 2328 2329 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2330 { 2331 } 2332 2333 const struct seq_operations ext4_mb_seq_groups_ops = { 2334 .start = ext4_mb_seq_groups_start, 2335 .next = ext4_mb_seq_groups_next, 2336 .stop = ext4_mb_seq_groups_stop, 2337 .show = ext4_mb_seq_groups_show, 2338 }; 2339 2340 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 2341 { 2342 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2343 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 2344 2345 BUG_ON(!cachep); 2346 return cachep; 2347 } 2348 2349 /* 2350 * Allocate the top-level s_group_info array for the specified number 2351 * of groups 2352 */ 2353 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) 2354 { 2355 struct ext4_sb_info *sbi = EXT4_SB(sb); 2356 unsigned size; 2357 struct ext4_group_info ***new_groupinfo; 2358 2359 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> 2360 EXT4_DESC_PER_BLOCK_BITS(sb); 2361 if (size <= sbi->s_group_info_size) 2362 return 0; 2363 2364 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); 2365 new_groupinfo = kvzalloc(size, GFP_KERNEL); 2366 if (!new_groupinfo) { 2367 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 2368 return -ENOMEM; 2369 } 2370 if (sbi->s_group_info) { 2371 memcpy(new_groupinfo, sbi->s_group_info, 2372 sbi->s_group_info_size * sizeof(*sbi->s_group_info)); 2373 kvfree(sbi->s_group_info); 2374 } 2375 sbi->s_group_info = new_groupinfo; 2376 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); 2377 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 2378 sbi->s_group_info_size); 2379 return 0; 2380 } 2381 2382 /* Create and initialize ext4_group_info data for the given group. */ 2383 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 2384 struct ext4_group_desc *desc) 2385 { 2386 int i; 2387 int metalen = 0; 2388 struct ext4_sb_info *sbi = EXT4_SB(sb); 2389 struct ext4_group_info **meta_group_info; 2390 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2391 2392 /* 2393 * First check if this group is the first of a reserved block. 2394 * If it's true, we have to allocate a new table of pointers 2395 * to ext4_group_info structures 2396 */ 2397 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 2398 metalen = sizeof(*meta_group_info) << 2399 EXT4_DESC_PER_BLOCK_BITS(sb); 2400 meta_group_info = kmalloc(metalen, GFP_NOFS); 2401 if (meta_group_info == NULL) { 2402 ext4_msg(sb, KERN_ERR, "can't allocate mem " 2403 "for a buddy group"); 2404 goto exit_meta_group_info; 2405 } 2406 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 2407 meta_group_info; 2408 } 2409 2410 meta_group_info = 2411 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 2412 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 2413 2414 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); 2415 if (meta_group_info[i] == NULL) { 2416 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 2417 goto exit_group_info; 2418 } 2419 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 2420 &(meta_group_info[i]->bb_state)); 2421 2422 /* 2423 * initialize bb_free to be able to skip 2424 * empty groups without initialization 2425 */ 2426 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2427 meta_group_info[i]->bb_free = 2428 ext4_free_clusters_after_init(sb, group, desc); 2429 } else { 2430 meta_group_info[i]->bb_free = 2431 ext4_free_group_clusters(sb, desc); 2432 } 2433 2434 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2435 init_rwsem(&meta_group_info[i]->alloc_sem); 2436 meta_group_info[i]->bb_free_root = RB_ROOT; 2437 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 2438 2439 #ifdef DOUBLE_CHECK 2440 { 2441 struct buffer_head *bh; 2442 meta_group_info[i]->bb_bitmap = 2443 kmalloc(sb->s_blocksize, GFP_NOFS); 2444 BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 2445 bh = ext4_read_block_bitmap(sb, group); 2446 BUG_ON(IS_ERR_OR_NULL(bh)); 2447 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 2448 sb->s_blocksize); 2449 put_bh(bh); 2450 } 2451 #endif 2452 2453 return 0; 2454 2455 exit_group_info: 2456 /* If a meta_group_info table has been allocated, release it now */ 2457 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 2458 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 2459 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; 2460 } 2461 exit_meta_group_info: 2462 return -ENOMEM; 2463 } /* ext4_mb_add_groupinfo */ 2464 2465 static int ext4_mb_init_backend(struct super_block *sb) 2466 { 2467 ext4_group_t ngroups = ext4_get_groups_count(sb); 2468 ext4_group_t i; 2469 struct ext4_sb_info *sbi = EXT4_SB(sb); 2470 int err; 2471 struct ext4_group_desc *desc; 2472 struct kmem_cache *cachep; 2473 2474 err = ext4_mb_alloc_groupinfo(sb, ngroups); 2475 if (err) 2476 return err; 2477 2478 sbi->s_buddy_cache = new_inode(sb); 2479 if (sbi->s_buddy_cache == NULL) { 2480 ext4_msg(sb, KERN_ERR, "can't get new inode"); 2481 goto err_freesgi; 2482 } 2483 /* To avoid potentially colliding with an valid on-disk inode number, 2484 * use EXT4_BAD_INO for the buddy cache inode number. This inode is 2485 * not in the inode hash, so it should never be found by iget(), but 2486 * this will avoid confusion if it ever shows up during debugging. */ 2487 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 2488 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 2489 for (i = 0; i < ngroups; i++) { 2490 desc = ext4_get_group_desc(sb, i, NULL); 2491 if (desc == NULL) { 2492 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 2493 goto err_freebuddy; 2494 } 2495 if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 2496 goto err_freebuddy; 2497 } 2498 2499 return 0; 2500 2501 err_freebuddy: 2502 cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2503 while (i-- > 0) 2504 kmem_cache_free(cachep, ext4_get_group_info(sb, i)); 2505 i = sbi->s_group_info_size; 2506 while (i-- > 0) 2507 kfree(sbi->s_group_info[i]); 2508 iput(sbi->s_buddy_cache); 2509 err_freesgi: 2510 kvfree(sbi->s_group_info); 2511 return -ENOMEM; 2512 } 2513 2514 static void ext4_groupinfo_destroy_slabs(void) 2515 { 2516 int i; 2517 2518 for (i = 0; i < NR_GRPINFO_CACHES; i++) { 2519 kmem_cache_destroy(ext4_groupinfo_caches[i]); 2520 ext4_groupinfo_caches[i] = NULL; 2521 } 2522 } 2523 2524 static int ext4_groupinfo_create_slab(size_t size) 2525 { 2526 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 2527 int slab_size; 2528 int blocksize_bits = order_base_2(size); 2529 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2530 struct kmem_cache *cachep; 2531 2532 if (cache_index >= NR_GRPINFO_CACHES) 2533 return -EINVAL; 2534 2535 if (unlikely(cache_index < 0)) 2536 cache_index = 0; 2537 2538 mutex_lock(&ext4_grpinfo_slab_create_mutex); 2539 if (ext4_groupinfo_caches[cache_index]) { 2540 mutex_unlock(&ext4_grpinfo_slab_create_mutex); 2541 return 0; /* Already created */ 2542 } 2543 2544 slab_size = offsetof(struct ext4_group_info, 2545 bb_counters[blocksize_bits + 2]); 2546 2547 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 2548 slab_size, 0, SLAB_RECLAIM_ACCOUNT, 2549 NULL); 2550 2551 ext4_groupinfo_caches[cache_index] = cachep; 2552 2553 mutex_unlock(&ext4_grpinfo_slab_create_mutex); 2554 if (!cachep) { 2555 printk(KERN_EMERG 2556 "EXT4-fs: no memory for groupinfo slab cache\n"); 2557 return -ENOMEM; 2558 } 2559 2560 return 0; 2561 } 2562 2563 int ext4_mb_init(struct super_block *sb) 2564 { 2565 struct ext4_sb_info *sbi = EXT4_SB(sb); 2566 unsigned i, j; 2567 unsigned offset, offset_incr; 2568 unsigned max; 2569 int ret; 2570 2571 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); 2572 2573 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2574 if (sbi->s_mb_offsets == NULL) { 2575 ret = -ENOMEM; 2576 goto out; 2577 } 2578 2579 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); 2580 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2581 if (sbi->s_mb_maxs == NULL) { 2582 ret = -ENOMEM; 2583 goto out; 2584 } 2585 2586 ret = ext4_groupinfo_create_slab(sb->s_blocksize); 2587 if (ret < 0) 2588 goto out; 2589 2590 /* order 0 is regular bitmap */ 2591 sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2592 sbi->s_mb_offsets[0] = 0; 2593 2594 i = 1; 2595 offset = 0; 2596 offset_incr = 1 << (sb->s_blocksize_bits - 1); 2597 max = sb->s_blocksize << 2; 2598 do { 2599 sbi->s_mb_offsets[i] = offset; 2600 sbi->s_mb_maxs[i] = max; 2601 offset += offset_incr; 2602 offset_incr = offset_incr >> 1; 2603 max = max >> 1; 2604 i++; 2605 } while (i <= sb->s_blocksize_bits + 1); 2606 2607 spin_lock_init(&sbi->s_md_lock); 2608 spin_lock_init(&sbi->s_bal_lock); 2609 sbi->s_mb_free_pending = 0; 2610 INIT_LIST_HEAD(&sbi->s_freed_data_list); 2611 2612 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2613 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2614 sbi->s_mb_stats = MB_DEFAULT_STATS; 2615 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2616 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 2617 /* 2618 * The default group preallocation is 512, which for 4k block 2619 * sizes translates to 2 megabytes. However for bigalloc file 2620 * systems, this is probably too big (i.e, if the cluster size 2621 * is 1 megabyte, then group preallocation size becomes half a 2622 * gigabyte!). As a default, we will keep a two megabyte 2623 * group pralloc size for cluster sizes up to 64k, and after 2624 * that, we will force a minimum group preallocation size of 2625 * 32 clusters. This translates to 8 megs when the cluster 2626 * size is 256k, and 32 megs when the cluster size is 1 meg, 2627 * which seems reasonable as a default. 2628 */ 2629 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 2630 sbi->s_cluster_bits, 32); 2631 /* 2632 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 2633 * to the lowest multiple of s_stripe which is bigger than 2634 * the s_mb_group_prealloc as determined above. We want 2635 * the preallocation size to be an exact multiple of the 2636 * RAID stripe size so that preallocations don't fragment 2637 * the stripes. 2638 */ 2639 if (sbi->s_stripe > 1) { 2640 sbi->s_mb_group_prealloc = roundup( 2641 sbi->s_mb_group_prealloc, sbi->s_stripe); 2642 } 2643 2644 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2645 if (sbi->s_locality_groups == NULL) { 2646 ret = -ENOMEM; 2647 goto out; 2648 } 2649 for_each_possible_cpu(i) { 2650 struct ext4_locality_group *lg; 2651 lg = per_cpu_ptr(sbi->s_locality_groups, i); 2652 mutex_init(&lg->lg_mutex); 2653 for (j = 0; j < PREALLOC_TB_SIZE; j++) 2654 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2655 spin_lock_init(&lg->lg_prealloc_lock); 2656 } 2657 2658 /* init file for buddy data */ 2659 ret = ext4_mb_init_backend(sb); 2660 if (ret != 0) 2661 goto out_free_locality_groups; 2662 2663 return 0; 2664 2665 out_free_locality_groups: 2666 free_percpu(sbi->s_locality_groups); 2667 sbi->s_locality_groups = NULL; 2668 out: 2669 kfree(sbi->s_mb_offsets); 2670 sbi->s_mb_offsets = NULL; 2671 kfree(sbi->s_mb_maxs); 2672 sbi->s_mb_maxs = NULL; 2673 return ret; 2674 } 2675 2676 /* need to called with the ext4 group lock held */ 2677 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2678 { 2679 struct ext4_prealloc_space *pa; 2680 struct list_head *cur, *tmp; 2681 int count = 0; 2682 2683 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2684 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2685 list_del(&pa->pa_group_list); 2686 count++; 2687 kmem_cache_free(ext4_pspace_cachep, pa); 2688 } 2689 if (count) 2690 mb_debug(1, "mballoc: %u PAs left\n", count); 2691 2692 } 2693 2694 int ext4_mb_release(struct super_block *sb) 2695 { 2696 ext4_group_t ngroups = ext4_get_groups_count(sb); 2697 ext4_group_t i; 2698 int num_meta_group_infos; 2699 struct ext4_group_info *grinfo; 2700 struct ext4_sb_info *sbi = EXT4_SB(sb); 2701 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2702 2703 if (sbi->s_group_info) { 2704 for (i = 0; i < ngroups; i++) { 2705 grinfo = ext4_get_group_info(sb, i); 2706 #ifdef DOUBLE_CHECK 2707 kfree(grinfo->bb_bitmap); 2708 #endif 2709 ext4_lock_group(sb, i); 2710 ext4_mb_cleanup_pa(grinfo); 2711 ext4_unlock_group(sb, i); 2712 kmem_cache_free(cachep, grinfo); 2713 } 2714 num_meta_group_infos = (ngroups + 2715 EXT4_DESC_PER_BLOCK(sb) - 1) >> 2716 EXT4_DESC_PER_BLOCK_BITS(sb); 2717 for (i = 0; i < num_meta_group_infos; i++) 2718 kfree(sbi->s_group_info[i]); 2719 kvfree(sbi->s_group_info); 2720 } 2721 kfree(sbi->s_mb_offsets); 2722 kfree(sbi->s_mb_maxs); 2723 iput(sbi->s_buddy_cache); 2724 if (sbi->s_mb_stats) { 2725 ext4_msg(sb, KERN_INFO, 2726 "mballoc: %u blocks %u reqs (%u success)", 2727 atomic_read(&sbi->s_bal_allocated), 2728 atomic_read(&sbi->s_bal_reqs), 2729 atomic_read(&sbi->s_bal_success)); 2730 ext4_msg(sb, KERN_INFO, 2731 "mballoc: %u extents scanned, %u goal hits, " 2732 "%u 2^N hits, %u breaks, %u lost", 2733 atomic_read(&sbi->s_bal_ex_scanned), 2734 atomic_read(&sbi->s_bal_goals), 2735 atomic_read(&sbi->s_bal_2orders), 2736 atomic_read(&sbi->s_bal_breaks), 2737 atomic_read(&sbi->s_mb_lost_chunks)); 2738 ext4_msg(sb, KERN_INFO, 2739 "mballoc: %lu generated and it took %Lu", 2740 sbi->s_mb_buddies_generated, 2741 sbi->s_mb_generation_time); 2742 ext4_msg(sb, KERN_INFO, 2743 "mballoc: %u preallocated, %u discarded", 2744 atomic_read(&sbi->s_mb_preallocated), 2745 atomic_read(&sbi->s_mb_discarded)); 2746 } 2747 2748 free_percpu(sbi->s_locality_groups); 2749 2750 return 0; 2751 } 2752 2753 static inline int ext4_issue_discard(struct super_block *sb, 2754 ext4_group_t block_group, ext4_grpblk_t cluster, int count, 2755 struct bio **biop) 2756 { 2757 ext4_fsblk_t discard_block; 2758 2759 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 2760 ext4_group_first_block_no(sb, block_group)); 2761 count = EXT4_C2B(EXT4_SB(sb), count); 2762 trace_ext4_discard_blocks(sb, 2763 (unsigned long long) discard_block, count); 2764 if (biop) { 2765 return __blkdev_issue_discard(sb->s_bdev, 2766 (sector_t)discard_block << (sb->s_blocksize_bits - 9), 2767 (sector_t)count << (sb->s_blocksize_bits - 9), 2768 GFP_NOFS, 0, biop); 2769 } else 2770 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 2771 } 2772 2773 static void ext4_free_data_in_buddy(struct super_block *sb, 2774 struct ext4_free_data *entry) 2775 { 2776 struct ext4_buddy e4b; 2777 struct ext4_group_info *db; 2778 int err, count = 0, count2 = 0; 2779 2780 mb_debug(1, "gonna free %u blocks in group %u (0x%p):", 2781 entry->efd_count, entry->efd_group, entry); 2782 2783 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 2784 /* we expect to find existing buddy because it's pinned */ 2785 BUG_ON(err != 0); 2786 2787 spin_lock(&EXT4_SB(sb)->s_md_lock); 2788 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; 2789 spin_unlock(&EXT4_SB(sb)->s_md_lock); 2790 2791 db = e4b.bd_info; 2792 /* there are blocks to put in buddy to make them really free */ 2793 count += entry->efd_count; 2794 count2++; 2795 ext4_lock_group(sb, entry->efd_group); 2796 /* Take it out of per group rb tree */ 2797 rb_erase(&entry->efd_node, &(db->bb_free_root)); 2798 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 2799 2800 /* 2801 * Clear the trimmed flag for the group so that the next 2802 * ext4_trim_fs can trim it. 2803 * If the volume is mounted with -o discard, online discard 2804 * is supported and the free blocks will be trimmed online. 2805 */ 2806 if (!test_opt(sb, DISCARD)) 2807 EXT4_MB_GRP_CLEAR_TRIMMED(db); 2808 2809 if (!db->bb_free_root.rb_node) { 2810 /* No more items in the per group rb tree 2811 * balance refcounts from ext4_mb_free_metadata() 2812 */ 2813 put_page(e4b.bd_buddy_page); 2814 put_page(e4b.bd_bitmap_page); 2815 } 2816 ext4_unlock_group(sb, entry->efd_group); 2817 kmem_cache_free(ext4_free_data_cachep, entry); 2818 ext4_mb_unload_buddy(&e4b); 2819 2820 mb_debug(1, "freed %u blocks in %u structures\n", count, count2); 2821 } 2822 2823 /* 2824 * This function is called by the jbd2 layer once the commit has finished, 2825 * so we know we can free the blocks that were released with that commit. 2826 */ 2827 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) 2828 { 2829 struct ext4_sb_info *sbi = EXT4_SB(sb); 2830 struct ext4_free_data *entry, *tmp; 2831 struct bio *discard_bio = NULL; 2832 struct list_head freed_data_list; 2833 struct list_head *cut_pos = NULL; 2834 int err; 2835 2836 INIT_LIST_HEAD(&freed_data_list); 2837 2838 spin_lock(&sbi->s_md_lock); 2839 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { 2840 if (entry->efd_tid != commit_tid) 2841 break; 2842 cut_pos = &entry->efd_list; 2843 } 2844 if (cut_pos) 2845 list_cut_position(&freed_data_list, &sbi->s_freed_data_list, 2846 cut_pos); 2847 spin_unlock(&sbi->s_md_lock); 2848 2849 if (test_opt(sb, DISCARD)) { 2850 list_for_each_entry(entry, &freed_data_list, efd_list) { 2851 err = ext4_issue_discard(sb, entry->efd_group, 2852 entry->efd_start_cluster, 2853 entry->efd_count, 2854 &discard_bio); 2855 if (err && err != -EOPNOTSUPP) { 2856 ext4_msg(sb, KERN_WARNING, "discard request in" 2857 " group:%d block:%d count:%d failed" 2858 " with %d", entry->efd_group, 2859 entry->efd_start_cluster, 2860 entry->efd_count, err); 2861 } else if (err == -EOPNOTSUPP) 2862 break; 2863 } 2864 2865 if (discard_bio) { 2866 submit_bio_wait(discard_bio); 2867 bio_put(discard_bio); 2868 } 2869 } 2870 2871 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) 2872 ext4_free_data_in_buddy(sb, entry); 2873 } 2874 2875 int __init ext4_init_mballoc(void) 2876 { 2877 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 2878 SLAB_RECLAIM_ACCOUNT); 2879 if (ext4_pspace_cachep == NULL) 2880 return -ENOMEM; 2881 2882 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 2883 SLAB_RECLAIM_ACCOUNT); 2884 if (ext4_ac_cachep == NULL) { 2885 kmem_cache_destroy(ext4_pspace_cachep); 2886 return -ENOMEM; 2887 } 2888 2889 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 2890 SLAB_RECLAIM_ACCOUNT); 2891 if (ext4_free_data_cachep == NULL) { 2892 kmem_cache_destroy(ext4_pspace_cachep); 2893 kmem_cache_destroy(ext4_ac_cachep); 2894 return -ENOMEM; 2895 } 2896 return 0; 2897 } 2898 2899 void ext4_exit_mballoc(void) 2900 { 2901 /* 2902 * Wait for completion of call_rcu()'s on ext4_pspace_cachep 2903 * before destroying the slab cache. 2904 */ 2905 rcu_barrier(); 2906 kmem_cache_destroy(ext4_pspace_cachep); 2907 kmem_cache_destroy(ext4_ac_cachep); 2908 kmem_cache_destroy(ext4_free_data_cachep); 2909 ext4_groupinfo_destroy_slabs(); 2910 } 2911 2912 2913 /* 2914 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 2915 * Returns 0 if success or error code 2916 */ 2917 static noinline_for_stack int 2918 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 2919 handle_t *handle, unsigned int reserv_clstrs) 2920 { 2921 struct buffer_head *bitmap_bh = NULL; 2922 struct ext4_group_desc *gdp; 2923 struct buffer_head *gdp_bh; 2924 struct ext4_sb_info *sbi; 2925 struct super_block *sb; 2926 ext4_fsblk_t block; 2927 int err, len; 2928 2929 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2930 BUG_ON(ac->ac_b_ex.fe_len <= 0); 2931 2932 sb = ac->ac_sb; 2933 sbi = EXT4_SB(sb); 2934 2935 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 2936 if (IS_ERR(bitmap_bh)) { 2937 err = PTR_ERR(bitmap_bh); 2938 bitmap_bh = NULL; 2939 goto out_err; 2940 } 2941 2942 BUFFER_TRACE(bitmap_bh, "getting write access"); 2943 err = ext4_journal_get_write_access(handle, bitmap_bh); 2944 if (err) 2945 goto out_err; 2946 2947 err = -EIO; 2948 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2949 if (!gdp) 2950 goto out_err; 2951 2952 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 2953 ext4_free_group_clusters(sb, gdp)); 2954 2955 BUFFER_TRACE(gdp_bh, "get_write_access"); 2956 err = ext4_journal_get_write_access(handle, gdp_bh); 2957 if (err) 2958 goto out_err; 2959 2960 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 2961 2962 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 2963 if (!ext4_data_block_valid(sbi, block, len)) { 2964 ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 2965 "fs metadata", block, block+len); 2966 /* File system mounted not to panic on error 2967 * Fix the bitmap and return EFSCORRUPTED 2968 * We leak some of the blocks here. 2969 */ 2970 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2971 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2972 ac->ac_b_ex.fe_len); 2973 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 2974 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2975 if (!err) 2976 err = -EFSCORRUPTED; 2977 goto out_err; 2978 } 2979 2980 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2981 #ifdef AGGRESSIVE_CHECK 2982 { 2983 int i; 2984 for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 2985 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 2986 bitmap_bh->b_data)); 2987 } 2988 } 2989 #endif 2990 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2991 ac->ac_b_ex.fe_len); 2992 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2993 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 2994 ext4_free_group_clusters_set(sb, gdp, 2995 ext4_free_clusters_after_init(sb, 2996 ac->ac_b_ex.fe_group, gdp)); 2997 } 2998 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; 2999 ext4_free_group_clusters_set(sb, gdp, len); 3000 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); 3001 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); 3002 3003 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3004 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 3005 /* 3006 * Now reduce the dirty block count also. Should not go negative 3007 */ 3008 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 3009 /* release all the reserved blocks if non delalloc */ 3010 percpu_counter_sub(&sbi->s_dirtyclusters_counter, 3011 reserv_clstrs); 3012 3013 if (sbi->s_log_groups_per_flex) { 3014 ext4_group_t flex_group = ext4_flex_group(sbi, 3015 ac->ac_b_ex.fe_group); 3016 atomic64_sub(ac->ac_b_ex.fe_len, 3017 &sbi->s_flex_groups[flex_group].free_clusters); 3018 } 3019 3020 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 3021 if (err) 3022 goto out_err; 3023 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 3024 3025 out_err: 3026 brelse(bitmap_bh); 3027 return err; 3028 } 3029 3030 /* 3031 * here we normalize request for locality group 3032 * Group request are normalized to s_mb_group_prealloc, which goes to 3033 * s_strip if we set the same via mount option. 3034 * s_mb_group_prealloc can be configured via 3035 * /sys/fs/ext4/<partition>/mb_group_prealloc 3036 * 3037 * XXX: should we try to preallocate more than the group has now? 3038 */ 3039 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 3040 { 3041 struct super_block *sb = ac->ac_sb; 3042 struct ext4_locality_group *lg = ac->ac_lg; 3043 3044 BUG_ON(lg == NULL); 3045 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 3046 mb_debug(1, "#%u: goal %u blocks for locality group\n", 3047 current->pid, ac->ac_g_ex.fe_len); 3048 } 3049 3050 /* 3051 * Normalization means making request better in terms of 3052 * size and alignment 3053 */ 3054 static noinline_for_stack void 3055 ext4_mb_normalize_request(struct ext4_allocation_context *ac, 3056 struct ext4_allocation_request *ar) 3057 { 3058 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3059 int bsbits, max; 3060 ext4_lblk_t end; 3061 loff_t size, start_off; 3062 loff_t orig_size __maybe_unused; 3063 ext4_lblk_t start; 3064 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3065 struct ext4_prealloc_space *pa; 3066 3067 /* do normalize only data requests, metadata requests 3068 do not need preallocation */ 3069 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3070 return; 3071 3072 /* sometime caller may want exact blocks */ 3073 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 3074 return; 3075 3076 /* caller may indicate that preallocation isn't 3077 * required (it's a tail, for example) */ 3078 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 3079 return; 3080 3081 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 3082 ext4_mb_normalize_group_request(ac); 3083 return ; 3084 } 3085 3086 bsbits = ac->ac_sb->s_blocksize_bits; 3087 3088 /* first, let's learn actual file size 3089 * given current request is allocated */ 3090 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 3091 size = size << bsbits; 3092 if (size < i_size_read(ac->ac_inode)) 3093 size = i_size_read(ac->ac_inode); 3094 orig_size = size; 3095 3096 /* max size of free chunks */ 3097 max = 2 << bsbits; 3098 3099 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 3100 (req <= (size) || max <= (chunk_size)) 3101 3102 /* first, try to predict filesize */ 3103 /* XXX: should this table be tunable? */ 3104 start_off = 0; 3105 if (size <= 16 * 1024) { 3106 size = 16 * 1024; 3107 } else if (size <= 32 * 1024) { 3108 size = 32 * 1024; 3109 } else if (size <= 64 * 1024) { 3110 size = 64 * 1024; 3111 } else if (size <= 128 * 1024) { 3112 size = 128 * 1024; 3113 } else if (size <= 256 * 1024) { 3114 size = 256 * 1024; 3115 } else if (size <= 512 * 1024) { 3116 size = 512 * 1024; 3117 } else if (size <= 1024 * 1024) { 3118 size = 1024 * 1024; 3119 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 3120 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3121 (21 - bsbits)) << 21; 3122 size = 2 * 1024 * 1024; 3123 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 3124 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3125 (22 - bsbits)) << 22; 3126 size = 4 * 1024 * 1024; 3127 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 3128 (8<<20)>>bsbits, max, 8 * 1024)) { 3129 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3130 (23 - bsbits)) << 23; 3131 size = 8 * 1024 * 1024; 3132 } else { 3133 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; 3134 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), 3135 ac->ac_o_ex.fe_len) << bsbits; 3136 } 3137 size = size >> bsbits; 3138 start = start_off >> bsbits; 3139 3140 /* don't cover already allocated blocks in selected range */ 3141 if (ar->pleft && start <= ar->lleft) { 3142 size -= ar->lleft + 1 - start; 3143 start = ar->lleft + 1; 3144 } 3145 if (ar->pright && start + size - 1 >= ar->lright) 3146 size -= start + size - ar->lright; 3147 3148 /* 3149 * Trim allocation request for filesystems with artificially small 3150 * groups. 3151 */ 3152 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) 3153 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); 3154 3155 end = start + size; 3156 3157 /* check we don't cross already preallocated blocks */ 3158 rcu_read_lock(); 3159 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3160 ext4_lblk_t pa_end; 3161 3162 if (pa->pa_deleted) 3163 continue; 3164 spin_lock(&pa->pa_lock); 3165 if (pa->pa_deleted) { 3166 spin_unlock(&pa->pa_lock); 3167 continue; 3168 } 3169 3170 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 3171 pa->pa_len); 3172 3173 /* PA must not overlap original request */ 3174 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 3175 ac->ac_o_ex.fe_logical < pa->pa_lstart)); 3176 3177 /* skip PAs this normalized request doesn't overlap with */ 3178 if (pa->pa_lstart >= end || pa_end <= start) { 3179 spin_unlock(&pa->pa_lock); 3180 continue; 3181 } 3182 BUG_ON(pa->pa_lstart <= start && pa_end >= end); 3183 3184 /* adjust start or end to be adjacent to this pa */ 3185 if (pa_end <= ac->ac_o_ex.fe_logical) { 3186 BUG_ON(pa_end < start); 3187 start = pa_end; 3188 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 3189 BUG_ON(pa->pa_lstart > end); 3190 end = pa->pa_lstart; 3191 } 3192 spin_unlock(&pa->pa_lock); 3193 } 3194 rcu_read_unlock(); 3195 size = end - start; 3196 3197 /* XXX: extra loop to check we really don't overlap preallocations */ 3198 rcu_read_lock(); 3199 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3200 ext4_lblk_t pa_end; 3201 3202 spin_lock(&pa->pa_lock); 3203 if (pa->pa_deleted == 0) { 3204 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 3205 pa->pa_len); 3206 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 3207 } 3208 spin_unlock(&pa->pa_lock); 3209 } 3210 rcu_read_unlock(); 3211 3212 if (start + size <= ac->ac_o_ex.fe_logical && 3213 start > ac->ac_o_ex.fe_logical) { 3214 ext4_msg(ac->ac_sb, KERN_ERR, 3215 "start %lu, size %lu, fe_logical %lu", 3216 (unsigned long) start, (unsigned long) size, 3217 (unsigned long) ac->ac_o_ex.fe_logical); 3218 BUG(); 3219 } 3220 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 3221 3222 /* now prepare goal request */ 3223 3224 /* XXX: is it better to align blocks WRT to logical 3225 * placement or satisfy big request as is */ 3226 ac->ac_g_ex.fe_logical = start; 3227 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 3228 3229 /* define goal start in order to merge */ 3230 if (ar->pright && (ar->lright == (start + size))) { 3231 /* merge to the right */ 3232 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 3233 &ac->ac_f_ex.fe_group, 3234 &ac->ac_f_ex.fe_start); 3235 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3236 } 3237 if (ar->pleft && (ar->lleft + 1 == start)) { 3238 /* merge to the left */ 3239 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 3240 &ac->ac_f_ex.fe_group, 3241 &ac->ac_f_ex.fe_start); 3242 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3243 } 3244 3245 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, 3246 (unsigned) orig_size, (unsigned) start); 3247 } 3248 3249 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 3250 { 3251 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3252 3253 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 3254 atomic_inc(&sbi->s_bal_reqs); 3255 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 3256 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 3257 atomic_inc(&sbi->s_bal_success); 3258 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 3259 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 3260 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 3261 atomic_inc(&sbi->s_bal_goals); 3262 if (ac->ac_found > sbi->s_mb_max_to_scan) 3263 atomic_inc(&sbi->s_bal_breaks); 3264 } 3265 3266 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 3267 trace_ext4_mballoc_alloc(ac); 3268 else 3269 trace_ext4_mballoc_prealloc(ac); 3270 } 3271 3272 /* 3273 * Called on failure; free up any blocks from the inode PA for this 3274 * context. We don't need this for MB_GROUP_PA because we only change 3275 * pa_free in ext4_mb_release_context(), but on failure, we've already 3276 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 3277 */ 3278 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 3279 { 3280 struct ext4_prealloc_space *pa = ac->ac_pa; 3281 struct ext4_buddy e4b; 3282 int err; 3283 3284 if (pa == NULL) { 3285 if (ac->ac_f_ex.fe_len == 0) 3286 return; 3287 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); 3288 if (err) { 3289 /* 3290 * This should never happen since we pin the 3291 * pages in the ext4_allocation_context so 3292 * ext4_mb_load_buddy() should never fail. 3293 */ 3294 WARN(1, "mb_load_buddy failed (%d)", err); 3295 return; 3296 } 3297 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 3298 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, 3299 ac->ac_f_ex.fe_len); 3300 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 3301 ext4_mb_unload_buddy(&e4b); 3302 return; 3303 } 3304 if (pa->pa_type == MB_INODE_PA) 3305 pa->pa_free += ac->ac_b_ex.fe_len; 3306 } 3307 3308 /* 3309 * use blocks preallocated to inode 3310 */ 3311 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3312 struct ext4_prealloc_space *pa) 3313 { 3314 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3315 ext4_fsblk_t start; 3316 ext4_fsblk_t end; 3317 int len; 3318 3319 /* found preallocated blocks, use them */ 3320 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 3321 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 3322 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 3323 len = EXT4_NUM_B2C(sbi, end - start); 3324 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3325 &ac->ac_b_ex.fe_start); 3326 ac->ac_b_ex.fe_len = len; 3327 ac->ac_status = AC_STATUS_FOUND; 3328 ac->ac_pa = pa; 3329 3330 BUG_ON(start < pa->pa_pstart); 3331 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 3332 BUG_ON(pa->pa_free < len); 3333 pa->pa_free -= len; 3334 3335 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); 3336 } 3337 3338 /* 3339 * use blocks preallocated to locality group 3340 */ 3341 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3342 struct ext4_prealloc_space *pa) 3343 { 3344 unsigned int len = ac->ac_o_ex.fe_len; 3345 3346 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3347 &ac->ac_b_ex.fe_group, 3348 &ac->ac_b_ex.fe_start); 3349 ac->ac_b_ex.fe_len = len; 3350 ac->ac_status = AC_STATUS_FOUND; 3351 ac->ac_pa = pa; 3352 3353 /* we don't correct pa_pstart or pa_plen here to avoid 3354 * possible race when the group is being loaded concurrently 3355 * instead we correct pa later, after blocks are marked 3356 * in on-disk bitmap -- see ext4_mb_release_context() 3357 * Other CPUs are prevented from allocating from this pa by lg_mutex 3358 */ 3359 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3360 } 3361 3362 /* 3363 * Return the prealloc space that have minimal distance 3364 * from the goal block. @cpa is the prealloc 3365 * space that is having currently known minimal distance 3366 * from the goal block. 3367 */ 3368 static struct ext4_prealloc_space * 3369 ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 3370 struct ext4_prealloc_space *pa, 3371 struct ext4_prealloc_space *cpa) 3372 { 3373 ext4_fsblk_t cur_distance, new_distance; 3374 3375 if (cpa == NULL) { 3376 atomic_inc(&pa->pa_count); 3377 return pa; 3378 } 3379 cur_distance = abs(goal_block - cpa->pa_pstart); 3380 new_distance = abs(goal_block - pa->pa_pstart); 3381 3382 if (cur_distance <= new_distance) 3383 return cpa; 3384 3385 /* drop the previous reference */ 3386 atomic_dec(&cpa->pa_count); 3387 atomic_inc(&pa->pa_count); 3388 return pa; 3389 } 3390 3391 /* 3392 * search goal blocks in preallocated space 3393 */ 3394 static noinline_for_stack int 3395 ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3396 { 3397 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3398 int order, i; 3399 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3400 struct ext4_locality_group *lg; 3401 struct ext4_prealloc_space *pa, *cpa = NULL; 3402 ext4_fsblk_t goal_block; 3403 3404 /* only data can be preallocated */ 3405 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3406 return 0; 3407 3408 /* first, try per-file preallocation */ 3409 rcu_read_lock(); 3410 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3411 3412 /* all fields in this condition don't change, 3413 * so we can skip locking for them */ 3414 if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 3415 ac->ac_o_ex.fe_logical >= (pa->pa_lstart + 3416 EXT4_C2B(sbi, pa->pa_len))) 3417 continue; 3418 3419 /* non-extent files can't have physical blocks past 2^32 */ 3420 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 3421 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > 3422 EXT4_MAX_BLOCK_FILE_PHYS)) 3423 continue; 3424 3425 /* found preallocated blocks, use them */ 3426 spin_lock(&pa->pa_lock); 3427 if (pa->pa_deleted == 0 && pa->pa_free) { 3428 atomic_inc(&pa->pa_count); 3429 ext4_mb_use_inode_pa(ac, pa); 3430 spin_unlock(&pa->pa_lock); 3431 ac->ac_criteria = 10; 3432 rcu_read_unlock(); 3433 return 1; 3434 } 3435 spin_unlock(&pa->pa_lock); 3436 } 3437 rcu_read_unlock(); 3438 3439 /* can we use group allocation? */ 3440 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3441 return 0; 3442 3443 /* inode may have no locality group for some reason */ 3444 lg = ac->ac_lg; 3445 if (lg == NULL) 3446 return 0; 3447 order = fls(ac->ac_o_ex.fe_len) - 1; 3448 if (order > PREALLOC_TB_SIZE - 1) 3449 /* The max size of hash table is PREALLOC_TB_SIZE */ 3450 order = PREALLOC_TB_SIZE - 1; 3451 3452 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 3453 /* 3454 * search for the prealloc space that is having 3455 * minimal distance from the goal block. 3456 */ 3457 for (i = order; i < PREALLOC_TB_SIZE; i++) { 3458 rcu_read_lock(); 3459 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 3460 pa_inode_list) { 3461 spin_lock(&pa->pa_lock); 3462 if (pa->pa_deleted == 0 && 3463 pa->pa_free >= ac->ac_o_ex.fe_len) { 3464 3465 cpa = ext4_mb_check_group_pa(goal_block, 3466 pa, cpa); 3467 } 3468 spin_unlock(&pa->pa_lock); 3469 } 3470 rcu_read_unlock(); 3471 } 3472 if (cpa) { 3473 ext4_mb_use_group_pa(ac, cpa); 3474 ac->ac_criteria = 20; 3475 return 1; 3476 } 3477 return 0; 3478 } 3479 3480 /* 3481 * the function goes through all block freed in the group 3482 * but not yet committed and marks them used in in-core bitmap. 3483 * buddy must be generated from this bitmap 3484 * Need to be called with the ext4 group lock held 3485 */ 3486 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 3487 ext4_group_t group) 3488 { 3489 struct rb_node *n; 3490 struct ext4_group_info *grp; 3491 struct ext4_free_data *entry; 3492 3493 grp = ext4_get_group_info(sb, group); 3494 n = rb_first(&(grp->bb_free_root)); 3495 3496 while (n) { 3497 entry = rb_entry(n, struct ext4_free_data, efd_node); 3498 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); 3499 n = rb_next(n); 3500 } 3501 return; 3502 } 3503 3504 /* 3505 * the function goes through all preallocation in this group and marks them 3506 * used in in-core bitmap. buddy must be generated from this bitmap 3507 * Need to be called with ext4 group lock held 3508 */ 3509 static noinline_for_stack 3510 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3511 ext4_group_t group) 3512 { 3513 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3514 struct ext4_prealloc_space *pa; 3515 struct list_head *cur; 3516 ext4_group_t groupnr; 3517 ext4_grpblk_t start; 3518 int preallocated = 0; 3519 int len; 3520 3521 /* all form of preallocation discards first load group, 3522 * so the only competing code is preallocation use. 3523 * we don't need any locking here 3524 * notice we do NOT ignore preallocations with pa_deleted 3525 * otherwise we could leave used blocks available for 3526 * allocation in buddy when concurrent ext4_mb_put_pa() 3527 * is dropping preallocation 3528 */ 3529 list_for_each(cur, &grp->bb_prealloc_list) { 3530 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3531 spin_lock(&pa->pa_lock); 3532 ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3533 &groupnr, &start); 3534 len = pa->pa_len; 3535 spin_unlock(&pa->pa_lock); 3536 if (unlikely(len == 0)) 3537 continue; 3538 BUG_ON(groupnr != group); 3539 ext4_set_bits(bitmap, start, len); 3540 preallocated += len; 3541 } 3542 mb_debug(1, "preallocated %u for group %u\n", preallocated, group); 3543 } 3544 3545 static void ext4_mb_pa_callback(struct rcu_head *head) 3546 { 3547 struct ext4_prealloc_space *pa; 3548 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 3549 3550 BUG_ON(atomic_read(&pa->pa_count)); 3551 BUG_ON(pa->pa_deleted == 0); 3552 kmem_cache_free(ext4_pspace_cachep, pa); 3553 } 3554 3555 /* 3556 * drops a reference to preallocated space descriptor 3557 * if this was the last reference and the space is consumed 3558 */ 3559 static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3560 struct super_block *sb, struct ext4_prealloc_space *pa) 3561 { 3562 ext4_group_t grp; 3563 ext4_fsblk_t grp_blk; 3564 3565 /* in this short window concurrent discard can set pa_deleted */ 3566 spin_lock(&pa->pa_lock); 3567 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { 3568 spin_unlock(&pa->pa_lock); 3569 return; 3570 } 3571 3572 if (pa->pa_deleted == 1) { 3573 spin_unlock(&pa->pa_lock); 3574 return; 3575 } 3576 3577 pa->pa_deleted = 1; 3578 spin_unlock(&pa->pa_lock); 3579 3580 grp_blk = pa->pa_pstart; 3581 /* 3582 * If doing group-based preallocation, pa_pstart may be in the 3583 * next group when pa is used up 3584 */ 3585 if (pa->pa_type == MB_GROUP_PA) 3586 grp_blk--; 3587 3588 grp = ext4_get_group_number(sb, grp_blk); 3589 3590 /* 3591 * possible race: 3592 * 3593 * P1 (buddy init) P2 (regular allocation) 3594 * find block B in PA 3595 * copy on-disk bitmap to buddy 3596 * mark B in on-disk bitmap 3597 * drop PA from group 3598 * mark all PAs in buddy 3599 * 3600 * thus, P1 initializes buddy with B available. to prevent this 3601 * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3602 * against that pair 3603 */ 3604 ext4_lock_group(sb, grp); 3605 list_del(&pa->pa_group_list); 3606 ext4_unlock_group(sb, grp); 3607 3608 spin_lock(pa->pa_obj_lock); 3609 list_del_rcu(&pa->pa_inode_list); 3610 spin_unlock(pa->pa_obj_lock); 3611 3612 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3613 } 3614 3615 /* 3616 * creates new preallocated space for given inode 3617 */ 3618 static noinline_for_stack int 3619 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3620 { 3621 struct super_block *sb = ac->ac_sb; 3622 struct ext4_sb_info *sbi = EXT4_SB(sb); 3623 struct ext4_prealloc_space *pa; 3624 struct ext4_group_info *grp; 3625 struct ext4_inode_info *ei; 3626 3627 /* preallocate only when found space is larger then requested */ 3628 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3629 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3630 BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3631 3632 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3633 if (pa == NULL) 3634 return -ENOMEM; 3635 3636 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3637 int winl; 3638 int wins; 3639 int win; 3640 int offs; 3641 3642 /* we can't allocate as much as normalizer wants. 3643 * so, found space must get proper lstart 3644 * to cover original request */ 3645 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3646 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3647 3648 /* we're limited by original request in that 3649 * logical block must be covered any way 3650 * winl is window we can move our chunk within */ 3651 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3652 3653 /* also, we should cover whole original request */ 3654 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); 3655 3656 /* the smallest one defines real window */ 3657 win = min(winl, wins); 3658 3659 offs = ac->ac_o_ex.fe_logical % 3660 EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 3661 if (offs && offs < win) 3662 win = offs; 3663 3664 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - 3665 EXT4_NUM_B2C(sbi, win); 3666 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3667 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3668 } 3669 3670 /* preallocation can change ac_b_ex, thus we store actually 3671 * allocated blocks for history */ 3672 ac->ac_f_ex = ac->ac_b_ex; 3673 3674 pa->pa_lstart = ac->ac_b_ex.fe_logical; 3675 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3676 pa->pa_len = ac->ac_b_ex.fe_len; 3677 pa->pa_free = pa->pa_len; 3678 atomic_set(&pa->pa_count, 1); 3679 spin_lock_init(&pa->pa_lock); 3680 INIT_LIST_HEAD(&pa->pa_inode_list); 3681 INIT_LIST_HEAD(&pa->pa_group_list); 3682 pa->pa_deleted = 0; 3683 pa->pa_type = MB_INODE_PA; 3684 3685 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, 3686 pa->pa_pstart, pa->pa_len, pa->pa_lstart); 3687 trace_ext4_mb_new_inode_pa(ac, pa); 3688 3689 ext4_mb_use_inode_pa(ac, pa); 3690 atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 3691 3692 ei = EXT4_I(ac->ac_inode); 3693 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3694 3695 pa->pa_obj_lock = &ei->i_prealloc_lock; 3696 pa->pa_inode = ac->ac_inode; 3697 3698 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3699 list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3700 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3701 3702 spin_lock(pa->pa_obj_lock); 3703 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3704 spin_unlock(pa->pa_obj_lock); 3705 3706 return 0; 3707 } 3708 3709 /* 3710 * creates new preallocated space for locality group inodes belongs to 3711 */ 3712 static noinline_for_stack int 3713 ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3714 { 3715 struct super_block *sb = ac->ac_sb; 3716 struct ext4_locality_group *lg; 3717 struct ext4_prealloc_space *pa; 3718 struct ext4_group_info *grp; 3719 3720 /* preallocate only when found space is larger then requested */ 3721 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3722 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3723 BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3724 3725 BUG_ON(ext4_pspace_cachep == NULL); 3726 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3727 if (pa == NULL) 3728 return -ENOMEM; 3729 3730 /* preallocation can change ac_b_ex, thus we store actually 3731 * allocated blocks for history */ 3732 ac->ac_f_ex = ac->ac_b_ex; 3733 3734 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3735 pa->pa_lstart = pa->pa_pstart; 3736 pa->pa_len = ac->ac_b_ex.fe_len; 3737 pa->pa_free = pa->pa_len; 3738 atomic_set(&pa->pa_count, 1); 3739 spin_lock_init(&pa->pa_lock); 3740 INIT_LIST_HEAD(&pa->pa_inode_list); 3741 INIT_LIST_HEAD(&pa->pa_group_list); 3742 pa->pa_deleted = 0; 3743 pa->pa_type = MB_GROUP_PA; 3744 3745 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, 3746 pa->pa_pstart, pa->pa_len, pa->pa_lstart); 3747 trace_ext4_mb_new_group_pa(ac, pa); 3748 3749 ext4_mb_use_group_pa(ac, pa); 3750 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3751 3752 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3753 lg = ac->ac_lg; 3754 BUG_ON(lg == NULL); 3755 3756 pa->pa_obj_lock = &lg->lg_prealloc_lock; 3757 pa->pa_inode = NULL; 3758 3759 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3760 list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3761 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3762 3763 /* 3764 * We will later add the new pa to the right bucket 3765 * after updating the pa_free in ext4_mb_release_context 3766 */ 3767 return 0; 3768 } 3769 3770 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3771 { 3772 int err; 3773 3774 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3775 err = ext4_mb_new_group_pa(ac); 3776 else 3777 err = ext4_mb_new_inode_pa(ac); 3778 return err; 3779 } 3780 3781 /* 3782 * finds all unused blocks in on-disk bitmap, frees them in 3783 * in-core bitmap and buddy. 3784 * @pa must be unlinked from inode and group lists, so that 3785 * nobody else can find/use it. 3786 * the caller MUST hold group/inode locks. 3787 * TODO: optimize the case when there are no in-core structures yet 3788 */ 3789 static noinline_for_stack int 3790 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 3791 struct ext4_prealloc_space *pa) 3792 { 3793 struct super_block *sb = e4b->bd_sb; 3794 struct ext4_sb_info *sbi = EXT4_SB(sb); 3795 unsigned int end; 3796 unsigned int next; 3797 ext4_group_t group; 3798 ext4_grpblk_t bit; 3799 unsigned long long grp_blk_start; 3800 int err = 0; 3801 int free = 0; 3802 3803 BUG_ON(pa->pa_deleted == 0); 3804 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3805 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 3806 BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3807 end = bit + pa->pa_len; 3808 3809 while (bit < end) { 3810 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3811 if (bit >= end) 3812 break; 3813 next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 3814 mb_debug(1, " free preallocated %u/%u in group %u\n", 3815 (unsigned) ext4_group_first_block_no(sb, group) + bit, 3816 (unsigned) next - bit, (unsigned) group); 3817 free += next - bit; 3818 3819 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 3820 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 3821 EXT4_C2B(sbi, bit)), 3822 next - bit); 3823 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3824 bit = next + 1; 3825 } 3826 if (free != pa->pa_free) { 3827 ext4_msg(e4b->bd_sb, KERN_CRIT, 3828 "pa %p: logic %lu, phys. %lu, len %lu", 3829 pa, (unsigned long) pa->pa_lstart, 3830 (unsigned long) pa->pa_pstart, 3831 (unsigned long) pa->pa_len); 3832 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 3833 free, pa->pa_free); 3834 /* 3835 * pa is already deleted so we use the value obtained 3836 * from the bitmap and continue. 3837 */ 3838 } 3839 atomic_add(free, &sbi->s_mb_discarded); 3840 3841 return err; 3842 } 3843 3844 static noinline_for_stack int 3845 ext4_mb_release_group_pa(struct ext4_buddy *e4b, 3846 struct ext4_prealloc_space *pa) 3847 { 3848 struct super_block *sb = e4b->bd_sb; 3849 ext4_group_t group; 3850 ext4_grpblk_t bit; 3851 3852 trace_ext4_mb_release_group_pa(sb, pa); 3853 BUG_ON(pa->pa_deleted == 0); 3854 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3855 BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3856 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3857 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 3858 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 3859 3860 return 0; 3861 } 3862 3863 /* 3864 * releases all preallocations in given group 3865 * 3866 * first, we need to decide discard policy: 3867 * - when do we discard 3868 * 1) ENOSPC 3869 * - how many do we discard 3870 * 1) how many requested 3871 */ 3872 static noinline_for_stack int 3873 ext4_mb_discard_group_preallocations(struct super_block *sb, 3874 ext4_group_t group, int needed) 3875 { 3876 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3877 struct buffer_head *bitmap_bh = NULL; 3878 struct ext4_prealloc_space *pa, *tmp; 3879 struct list_head list; 3880 struct ext4_buddy e4b; 3881 int err; 3882 int busy = 0; 3883 int free = 0; 3884 3885 mb_debug(1, "discard preallocation for group %u\n", group); 3886 3887 if (list_empty(&grp->bb_prealloc_list)) 3888 return 0; 3889 3890 bitmap_bh = ext4_read_block_bitmap(sb, group); 3891 if (IS_ERR(bitmap_bh)) { 3892 err = PTR_ERR(bitmap_bh); 3893 ext4_error(sb, "Error %d reading block bitmap for %u", 3894 err, group); 3895 return 0; 3896 } 3897 3898 err = ext4_mb_load_buddy(sb, group, &e4b); 3899 if (err) { 3900 ext4_warning(sb, "Error %d loading buddy information for %u", 3901 err, group); 3902 put_bh(bitmap_bh); 3903 return 0; 3904 } 3905 3906 if (needed == 0) 3907 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 3908 3909 INIT_LIST_HEAD(&list); 3910 repeat: 3911 ext4_lock_group(sb, group); 3912 list_for_each_entry_safe(pa, tmp, 3913 &grp->bb_prealloc_list, pa_group_list) { 3914 spin_lock(&pa->pa_lock); 3915 if (atomic_read(&pa->pa_count)) { 3916 spin_unlock(&pa->pa_lock); 3917 busy = 1; 3918 continue; 3919 } 3920 if (pa->pa_deleted) { 3921 spin_unlock(&pa->pa_lock); 3922 continue; 3923 } 3924 3925 /* seems this one can be freed ... */ 3926 pa->pa_deleted = 1; 3927 3928 /* we can trust pa_free ... */ 3929 free += pa->pa_free; 3930 3931 spin_unlock(&pa->pa_lock); 3932 3933 list_del(&pa->pa_group_list); 3934 list_add(&pa->u.pa_tmp_list, &list); 3935 } 3936 3937 /* if we still need more blocks and some PAs were used, try again */ 3938 if (free < needed && busy) { 3939 busy = 0; 3940 ext4_unlock_group(sb, group); 3941 cond_resched(); 3942 goto repeat; 3943 } 3944 3945 /* found anything to free? */ 3946 if (list_empty(&list)) { 3947 BUG_ON(free != 0); 3948 goto out; 3949 } 3950 3951 /* now free all selected PAs */ 3952 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3953 3954 /* remove from object (inode or locality group) */ 3955 spin_lock(pa->pa_obj_lock); 3956 list_del_rcu(&pa->pa_inode_list); 3957 spin_unlock(pa->pa_obj_lock); 3958 3959 if (pa->pa_type == MB_GROUP_PA) 3960 ext4_mb_release_group_pa(&e4b, pa); 3961 else 3962 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3963 3964 list_del(&pa->u.pa_tmp_list); 3965 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3966 } 3967 3968 out: 3969 ext4_unlock_group(sb, group); 3970 ext4_mb_unload_buddy(&e4b); 3971 put_bh(bitmap_bh); 3972 return free; 3973 } 3974 3975 /* 3976 * releases all non-used preallocated blocks for given inode 3977 * 3978 * It's important to discard preallocations under i_data_sem 3979 * We don't want another block to be served from the prealloc 3980 * space when we are discarding the inode prealloc space. 3981 * 3982 * FIXME!! Make sure it is valid at all the call sites 3983 */ 3984 void ext4_discard_preallocations(struct inode *inode) 3985 { 3986 struct ext4_inode_info *ei = EXT4_I(inode); 3987 struct super_block *sb = inode->i_sb; 3988 struct buffer_head *bitmap_bh = NULL; 3989 struct ext4_prealloc_space *pa, *tmp; 3990 ext4_group_t group = 0; 3991 struct list_head list; 3992 struct ext4_buddy e4b; 3993 int err; 3994 3995 if (!S_ISREG(inode->i_mode)) { 3996 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 3997 return; 3998 } 3999 4000 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); 4001 trace_ext4_discard_preallocations(inode); 4002 4003 INIT_LIST_HEAD(&list); 4004 4005 repeat: 4006 /* first, collect all pa's in the inode */ 4007 spin_lock(&ei->i_prealloc_lock); 4008 while (!list_empty(&ei->i_prealloc_list)) { 4009 pa = list_entry(ei->i_prealloc_list.next, 4010 struct ext4_prealloc_space, pa_inode_list); 4011 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 4012 spin_lock(&pa->pa_lock); 4013 if (atomic_read(&pa->pa_count)) { 4014 /* this shouldn't happen often - nobody should 4015 * use preallocation while we're discarding it */ 4016 spin_unlock(&pa->pa_lock); 4017 spin_unlock(&ei->i_prealloc_lock); 4018 ext4_msg(sb, KERN_ERR, 4019 "uh-oh! used pa while discarding"); 4020 WARN_ON(1); 4021 schedule_timeout_uninterruptible(HZ); 4022 goto repeat; 4023 4024 } 4025 if (pa->pa_deleted == 0) { 4026 pa->pa_deleted = 1; 4027 spin_unlock(&pa->pa_lock); 4028 list_del_rcu(&pa->pa_inode_list); 4029 list_add(&pa->u.pa_tmp_list, &list); 4030 continue; 4031 } 4032 4033 /* someone is deleting pa right now */ 4034 spin_unlock(&pa->pa_lock); 4035 spin_unlock(&ei->i_prealloc_lock); 4036 4037 /* we have to wait here because pa_deleted 4038 * doesn't mean pa is already unlinked from 4039 * the list. as we might be called from 4040 * ->clear_inode() the inode will get freed 4041 * and concurrent thread which is unlinking 4042 * pa from inode's list may access already 4043 * freed memory, bad-bad-bad */ 4044 4045 /* XXX: if this happens too often, we can 4046 * add a flag to force wait only in case 4047 * of ->clear_inode(), but not in case of 4048 * regular truncate */ 4049 schedule_timeout_uninterruptible(HZ); 4050 goto repeat; 4051 } 4052 spin_unlock(&ei->i_prealloc_lock); 4053 4054 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 4055 BUG_ON(pa->pa_type != MB_INODE_PA); 4056 group = ext4_get_group_number(sb, pa->pa_pstart); 4057 4058 err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 4059 GFP_NOFS|__GFP_NOFAIL); 4060 if (err) { 4061 ext4_error(sb, "Error %d loading buddy information for %u", 4062 err, group); 4063 continue; 4064 } 4065 4066 bitmap_bh = ext4_read_block_bitmap(sb, group); 4067 if (IS_ERR(bitmap_bh)) { 4068 err = PTR_ERR(bitmap_bh); 4069 ext4_error(sb, "Error %d reading block bitmap for %u", 4070 err, group); 4071 ext4_mb_unload_buddy(&e4b); 4072 continue; 4073 } 4074 4075 ext4_lock_group(sb, group); 4076 list_del(&pa->pa_group_list); 4077 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 4078 ext4_unlock_group(sb, group); 4079 4080 ext4_mb_unload_buddy(&e4b); 4081 put_bh(bitmap_bh); 4082 4083 list_del(&pa->u.pa_tmp_list); 4084 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 4085 } 4086 } 4087 4088 #ifdef CONFIG_EXT4_DEBUG 4089 static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4090 { 4091 struct super_block *sb = ac->ac_sb; 4092 ext4_group_t ngroups, i; 4093 4094 if (!ext4_mballoc_debug || 4095 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) 4096 return; 4097 4098 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" 4099 " Allocation context details:"); 4100 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", 4101 ac->ac_status, ac->ac_flags); 4102 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " 4103 "goal %lu/%lu/%lu@%lu, " 4104 "best %lu/%lu/%lu@%lu cr %d", 4105 (unsigned long)ac->ac_o_ex.fe_group, 4106 (unsigned long)ac->ac_o_ex.fe_start, 4107 (unsigned long)ac->ac_o_ex.fe_len, 4108 (unsigned long)ac->ac_o_ex.fe_logical, 4109 (unsigned long)ac->ac_g_ex.fe_group, 4110 (unsigned long)ac->ac_g_ex.fe_start, 4111 (unsigned long)ac->ac_g_ex.fe_len, 4112 (unsigned long)ac->ac_g_ex.fe_logical, 4113 (unsigned long)ac->ac_b_ex.fe_group, 4114 (unsigned long)ac->ac_b_ex.fe_start, 4115 (unsigned long)ac->ac_b_ex.fe_len, 4116 (unsigned long)ac->ac_b_ex.fe_logical, 4117 (int)ac->ac_criteria); 4118 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); 4119 ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); 4120 ngroups = ext4_get_groups_count(sb); 4121 for (i = 0; i < ngroups; i++) { 4122 struct ext4_group_info *grp = ext4_get_group_info(sb, i); 4123 struct ext4_prealloc_space *pa; 4124 ext4_grpblk_t start; 4125 struct list_head *cur; 4126 ext4_lock_group(sb, i); 4127 list_for_each(cur, &grp->bb_prealloc_list) { 4128 pa = list_entry(cur, struct ext4_prealloc_space, 4129 pa_group_list); 4130 spin_lock(&pa->pa_lock); 4131 ext4_get_group_no_and_offset(sb, pa->pa_pstart, 4132 NULL, &start); 4133 spin_unlock(&pa->pa_lock); 4134 printk(KERN_ERR "PA:%u:%d:%u \n", i, 4135 start, pa->pa_len); 4136 } 4137 ext4_unlock_group(sb, i); 4138 4139 if (grp->bb_free == 0) 4140 continue; 4141 printk(KERN_ERR "%u: %d/%d \n", 4142 i, grp->bb_free, grp->bb_fragments); 4143 } 4144 printk(KERN_ERR "\n"); 4145 } 4146 #else 4147 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4148 { 4149 return; 4150 } 4151 #endif 4152 4153 /* 4154 * We use locality group preallocation for small size file. The size of the 4155 * file is determined by the current size or the resulting size after 4156 * allocation which ever is larger 4157 * 4158 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 4159 */ 4160 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 4161 { 4162 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4163 int bsbits = ac->ac_sb->s_blocksize_bits; 4164 loff_t size, isize; 4165 4166 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4167 return; 4168 4169 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 4170 return; 4171 4172 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 4173 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 4174 >> bsbits; 4175 4176 if ((size == isize) && 4177 !ext4_fs_is_busy(sbi) && 4178 (atomic_read(&ac->ac_inode->i_writecount) == 0)) { 4179 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 4180 return; 4181 } 4182 4183 if (sbi->s_mb_group_prealloc <= 0) { 4184 ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4185 return; 4186 } 4187 4188 /* don't use group allocation for large files */ 4189 size = max(size, isize); 4190 if (size > sbi->s_mb_stream_request) { 4191 ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4192 return; 4193 } 4194 4195 BUG_ON(ac->ac_lg != NULL); 4196 /* 4197 * locality group prealloc space are per cpu. The reason for having 4198 * per cpu locality group is to reduce the contention between block 4199 * request from multiple CPUs. 4200 */ 4201 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); 4202 4203 /* we're going to use group allocation */ 4204 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 4205 4206 /* serialize all allocations in the group */ 4207 mutex_lock(&ac->ac_lg->lg_mutex); 4208 } 4209 4210 static noinline_for_stack int 4211 ext4_mb_initialize_context(struct ext4_allocation_context *ac, 4212 struct ext4_allocation_request *ar) 4213 { 4214 struct super_block *sb = ar->inode->i_sb; 4215 struct ext4_sb_info *sbi = EXT4_SB(sb); 4216 struct ext4_super_block *es = sbi->s_es; 4217 ext4_group_t group; 4218 unsigned int len; 4219 ext4_fsblk_t goal; 4220 ext4_grpblk_t block; 4221 4222 /* we can't allocate > group size */ 4223 len = ar->len; 4224 4225 /* just a dirty hack to filter too big requests */ 4226 if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) 4227 len = EXT4_CLUSTERS_PER_GROUP(sb); 4228 4229 /* start searching from the goal */ 4230 goal = ar->goal; 4231 if (goal < le32_to_cpu(es->s_first_data_block) || 4232 goal >= ext4_blocks_count(es)) 4233 goal = le32_to_cpu(es->s_first_data_block); 4234 ext4_get_group_no_and_offset(sb, goal, &group, &block); 4235 4236 /* set up allocation goals */ 4237 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); 4238 ac->ac_status = AC_STATUS_CONTINUE; 4239 ac->ac_sb = sb; 4240 ac->ac_inode = ar->inode; 4241 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 4242 ac->ac_o_ex.fe_group = group; 4243 ac->ac_o_ex.fe_start = block; 4244 ac->ac_o_ex.fe_len = len; 4245 ac->ac_g_ex = ac->ac_o_ex; 4246 ac->ac_flags = ar->flags; 4247 4248 /* we have to define context: we'll we work with a file or 4249 * locality group. this is a policy, actually */ 4250 ext4_mb_group_or_file(ac); 4251 4252 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 4253 "left: %u/%u, right %u/%u to %swritable\n", 4254 (unsigned) ar->len, (unsigned) ar->logical, 4255 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 4256 (unsigned) ar->lleft, (unsigned) ar->pleft, 4257 (unsigned) ar->lright, (unsigned) ar->pright, 4258 atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 4259 return 0; 4260 4261 } 4262 4263 static noinline_for_stack void 4264 ext4_mb_discard_lg_preallocations(struct super_block *sb, 4265 struct ext4_locality_group *lg, 4266 int order, int total_entries) 4267 { 4268 ext4_group_t group = 0; 4269 struct ext4_buddy e4b; 4270 struct list_head discard_list; 4271 struct ext4_prealloc_space *pa, *tmp; 4272 4273 mb_debug(1, "discard locality group preallocation\n"); 4274 4275 INIT_LIST_HEAD(&discard_list); 4276 4277 spin_lock(&lg->lg_prealloc_lock); 4278 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 4279 pa_inode_list) { 4280 spin_lock(&pa->pa_lock); 4281 if (atomic_read(&pa->pa_count)) { 4282 /* 4283 * This is the pa that we just used 4284 * for block allocation. So don't 4285 * free that 4286 */ 4287 spin_unlock(&pa->pa_lock); 4288 continue; 4289 } 4290 if (pa->pa_deleted) { 4291 spin_unlock(&pa->pa_lock); 4292 continue; 4293 } 4294 /* only lg prealloc space */ 4295 BUG_ON(pa->pa_type != MB_GROUP_PA); 4296 4297 /* seems this one can be freed ... */ 4298 pa->pa_deleted = 1; 4299 spin_unlock(&pa->pa_lock); 4300 4301 list_del_rcu(&pa->pa_inode_list); 4302 list_add(&pa->u.pa_tmp_list, &discard_list); 4303 4304 total_entries--; 4305 if (total_entries <= 5) { 4306 /* 4307 * we want to keep only 5 entries 4308 * allowing it to grow to 8. This 4309 * mak sure we don't call discard 4310 * soon for this list. 4311 */ 4312 break; 4313 } 4314 } 4315 spin_unlock(&lg->lg_prealloc_lock); 4316 4317 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 4318 int err; 4319 4320 group = ext4_get_group_number(sb, pa->pa_pstart); 4321 err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 4322 GFP_NOFS|__GFP_NOFAIL); 4323 if (err) { 4324 ext4_error(sb, "Error %d loading buddy information for %u", 4325 err, group); 4326 continue; 4327 } 4328 ext4_lock_group(sb, group); 4329 list_del(&pa->pa_group_list); 4330 ext4_mb_release_group_pa(&e4b, pa); 4331 ext4_unlock_group(sb, group); 4332 4333 ext4_mb_unload_buddy(&e4b); 4334 list_del(&pa->u.pa_tmp_list); 4335 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 4336 } 4337 } 4338 4339 /* 4340 * We have incremented pa_count. So it cannot be freed at this 4341 * point. Also we hold lg_mutex. So no parallel allocation is 4342 * possible from this lg. That means pa_free cannot be updated. 4343 * 4344 * A parallel ext4_mb_discard_group_preallocations is possible. 4345 * which can cause the lg_prealloc_list to be updated. 4346 */ 4347 4348 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 4349 { 4350 int order, added = 0, lg_prealloc_count = 1; 4351 struct super_block *sb = ac->ac_sb; 4352 struct ext4_locality_group *lg = ac->ac_lg; 4353 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 4354 4355 order = fls(pa->pa_free) - 1; 4356 if (order > PREALLOC_TB_SIZE - 1) 4357 /* The max size of hash table is PREALLOC_TB_SIZE */ 4358 order = PREALLOC_TB_SIZE - 1; 4359 /* Add the prealloc space to lg */ 4360 spin_lock(&lg->lg_prealloc_lock); 4361 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 4362 pa_inode_list) { 4363 spin_lock(&tmp_pa->pa_lock); 4364 if (tmp_pa->pa_deleted) { 4365 spin_unlock(&tmp_pa->pa_lock); 4366 continue; 4367 } 4368 if (!added && pa->pa_free < tmp_pa->pa_free) { 4369 /* Add to the tail of the previous entry */ 4370 list_add_tail_rcu(&pa->pa_inode_list, 4371 &tmp_pa->pa_inode_list); 4372 added = 1; 4373 /* 4374 * we want to count the total 4375 * number of entries in the list 4376 */ 4377 } 4378 spin_unlock(&tmp_pa->pa_lock); 4379 lg_prealloc_count++; 4380 } 4381 if (!added) 4382 list_add_tail_rcu(&pa->pa_inode_list, 4383 &lg->lg_prealloc_list[order]); 4384 spin_unlock(&lg->lg_prealloc_lock); 4385 4386 /* Now trim the list to be not more than 8 elements */ 4387 if (lg_prealloc_count > 8) { 4388 ext4_mb_discard_lg_preallocations(sb, lg, 4389 order, lg_prealloc_count); 4390 return; 4391 } 4392 return ; 4393 } 4394 4395 /* 4396 * release all resource we used in allocation 4397 */ 4398 static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4399 { 4400 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4401 struct ext4_prealloc_space *pa = ac->ac_pa; 4402 if (pa) { 4403 if (pa->pa_type == MB_GROUP_PA) { 4404 /* see comment in ext4_mb_use_group_pa() */ 4405 spin_lock(&pa->pa_lock); 4406 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 4407 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 4408 pa->pa_free -= ac->ac_b_ex.fe_len; 4409 pa->pa_len -= ac->ac_b_ex.fe_len; 4410 spin_unlock(&pa->pa_lock); 4411 } 4412 } 4413 if (pa) { 4414 /* 4415 * We want to add the pa to the right bucket. 4416 * Remove it from the list and while adding 4417 * make sure the list to which we are adding 4418 * doesn't grow big. 4419 */ 4420 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { 4421 spin_lock(pa->pa_obj_lock); 4422 list_del_rcu(&pa->pa_inode_list); 4423 spin_unlock(pa->pa_obj_lock); 4424 ext4_mb_add_n_trim(ac); 4425 } 4426 ext4_mb_put_pa(ac, ac->ac_sb, pa); 4427 } 4428 if (ac->ac_bitmap_page) 4429 put_page(ac->ac_bitmap_page); 4430 if (ac->ac_buddy_page) 4431 put_page(ac->ac_buddy_page); 4432 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4433 mutex_unlock(&ac->ac_lg->lg_mutex); 4434 ext4_mb_collect_stats(ac); 4435 return 0; 4436 } 4437 4438 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4439 { 4440 ext4_group_t i, ngroups = ext4_get_groups_count(sb); 4441 int ret; 4442 int freed = 0; 4443 4444 trace_ext4_mb_discard_preallocations(sb, needed); 4445 for (i = 0; i < ngroups && needed > 0; i++) { 4446 ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4447 freed += ret; 4448 needed -= ret; 4449 } 4450 4451 return freed; 4452 } 4453 4454 /* 4455 * Main entry point into mballoc to allocate blocks 4456 * it tries to use preallocation first, then falls back 4457 * to usual allocation 4458 */ 4459 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4460 struct ext4_allocation_request *ar, int *errp) 4461 { 4462 int freed; 4463 struct ext4_allocation_context *ac = NULL; 4464 struct ext4_sb_info *sbi; 4465 struct super_block *sb; 4466 ext4_fsblk_t block = 0; 4467 unsigned int inquota = 0; 4468 unsigned int reserv_clstrs = 0; 4469 4470 might_sleep(); 4471 sb = ar->inode->i_sb; 4472 sbi = EXT4_SB(sb); 4473 4474 trace_ext4_request_blocks(ar); 4475 4476 /* Allow to use superuser reservation for quota file */ 4477 if (ext4_is_quota_file(ar->inode)) 4478 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 4479 4480 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { 4481 /* Without delayed allocation we need to verify 4482 * there is enough free blocks to do block allocation 4483 * and verify allocation doesn't exceed the quota limits. 4484 */ 4485 while (ar->len && 4486 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 4487 4488 /* let others to free the space */ 4489 cond_resched(); 4490 ar->len = ar->len >> 1; 4491 } 4492 if (!ar->len) { 4493 *errp = -ENOSPC; 4494 return 0; 4495 } 4496 reserv_clstrs = ar->len; 4497 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 4498 dquot_alloc_block_nofail(ar->inode, 4499 EXT4_C2B(sbi, ar->len)); 4500 } else { 4501 while (ar->len && 4502 dquot_alloc_block(ar->inode, 4503 EXT4_C2B(sbi, ar->len))) { 4504 4505 ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4506 ar->len--; 4507 } 4508 } 4509 inquota = ar->len; 4510 if (ar->len == 0) { 4511 *errp = -EDQUOT; 4512 goto out; 4513 } 4514 } 4515 4516 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); 4517 if (!ac) { 4518 ar->len = 0; 4519 *errp = -ENOMEM; 4520 goto out; 4521 } 4522 4523 *errp = ext4_mb_initialize_context(ac, ar); 4524 if (*errp) { 4525 ar->len = 0; 4526 goto out; 4527 } 4528 4529 ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4530 if (!ext4_mb_use_preallocated(ac)) { 4531 ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4532 ext4_mb_normalize_request(ac, ar); 4533 repeat: 4534 /* allocate space in core */ 4535 *errp = ext4_mb_regular_allocator(ac); 4536 if (*errp) 4537 goto discard_and_exit; 4538 4539 /* as we've just preallocated more space than 4540 * user requested originally, we store allocated 4541 * space in a special descriptor */ 4542 if (ac->ac_status == AC_STATUS_FOUND && 4543 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 4544 *errp = ext4_mb_new_preallocation(ac); 4545 if (*errp) { 4546 discard_and_exit: 4547 ext4_discard_allocated_blocks(ac); 4548 goto errout; 4549 } 4550 } 4551 if (likely(ac->ac_status == AC_STATUS_FOUND)) { 4552 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 4553 if (*errp) { 4554 ext4_discard_allocated_blocks(ac); 4555 goto errout; 4556 } else { 4557 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4558 ar->len = ac->ac_b_ex.fe_len; 4559 } 4560 } else { 4561 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4562 if (freed) 4563 goto repeat; 4564 *errp = -ENOSPC; 4565 } 4566 4567 errout: 4568 if (*errp) { 4569 ac->ac_b_ex.fe_len = 0; 4570 ar->len = 0; 4571 ext4_mb_show_ac(ac); 4572 } 4573 ext4_mb_release_context(ac); 4574 out: 4575 if (ac) 4576 kmem_cache_free(ext4_ac_cachep, ac); 4577 if (inquota && ar->len < inquota) 4578 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 4579 if (!ar->len) { 4580 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) 4581 /* release all the reserved blocks if non delalloc */ 4582 percpu_counter_sub(&sbi->s_dirtyclusters_counter, 4583 reserv_clstrs); 4584 } 4585 4586 trace_ext4_allocate_blocks(ar, (unsigned long long)block); 4587 4588 return block; 4589 } 4590 4591 /* 4592 * We can merge two free data extents only if the physical blocks 4593 * are contiguous, AND the extents were freed by the same transaction, 4594 * AND the blocks are associated with the same group. 4595 */ 4596 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, 4597 struct ext4_free_data *entry, 4598 struct ext4_free_data *new_entry, 4599 struct rb_root *entry_rb_root) 4600 { 4601 if ((entry->efd_tid != new_entry->efd_tid) || 4602 (entry->efd_group != new_entry->efd_group)) 4603 return; 4604 if (entry->efd_start_cluster + entry->efd_count == 4605 new_entry->efd_start_cluster) { 4606 new_entry->efd_start_cluster = entry->efd_start_cluster; 4607 new_entry->efd_count += entry->efd_count; 4608 } else if (new_entry->efd_start_cluster + new_entry->efd_count == 4609 entry->efd_start_cluster) { 4610 new_entry->efd_count += entry->efd_count; 4611 } else 4612 return; 4613 spin_lock(&sbi->s_md_lock); 4614 list_del(&entry->efd_list); 4615 spin_unlock(&sbi->s_md_lock); 4616 rb_erase(&entry->efd_node, entry_rb_root); 4617 kmem_cache_free(ext4_free_data_cachep, entry); 4618 } 4619 4620 static noinline_for_stack int 4621 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 4622 struct ext4_free_data *new_entry) 4623 { 4624 ext4_group_t group = e4b->bd_group; 4625 ext4_grpblk_t cluster; 4626 ext4_grpblk_t clusters = new_entry->efd_count; 4627 struct ext4_free_data *entry; 4628 struct ext4_group_info *db = e4b->bd_info; 4629 struct super_block *sb = e4b->bd_sb; 4630 struct ext4_sb_info *sbi = EXT4_SB(sb); 4631 struct rb_node **n = &db->bb_free_root.rb_node, *node; 4632 struct rb_node *parent = NULL, *new_node; 4633 4634 BUG_ON(!ext4_handle_valid(handle)); 4635 BUG_ON(e4b->bd_bitmap_page == NULL); 4636 BUG_ON(e4b->bd_buddy_page == NULL); 4637 4638 new_node = &new_entry->efd_node; 4639 cluster = new_entry->efd_start_cluster; 4640 4641 if (!*n) { 4642 /* first free block exent. We need to 4643 protect buddy cache from being freed, 4644 * otherwise we'll refresh it from 4645 * on-disk bitmap and lose not-yet-available 4646 * blocks */ 4647 get_page(e4b->bd_buddy_page); 4648 get_page(e4b->bd_bitmap_page); 4649 } 4650 while (*n) { 4651 parent = *n; 4652 entry = rb_entry(parent, struct ext4_free_data, efd_node); 4653 if (cluster < entry->efd_start_cluster) 4654 n = &(*n)->rb_left; 4655 else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 4656 n = &(*n)->rb_right; 4657 else { 4658 ext4_grp_locked_error(sb, group, 0, 4659 ext4_group_first_block_no(sb, group) + 4660 EXT4_C2B(sbi, cluster), 4661 "Block already on to-be-freed list"); 4662 return 0; 4663 } 4664 } 4665 4666 rb_link_node(new_node, parent, n); 4667 rb_insert_color(new_node, &db->bb_free_root); 4668 4669 /* Now try to see the extent can be merged to left and right */ 4670 node = rb_prev(new_node); 4671 if (node) { 4672 entry = rb_entry(node, struct ext4_free_data, efd_node); 4673 ext4_try_merge_freed_extent(sbi, entry, new_entry, 4674 &(db->bb_free_root)); 4675 } 4676 4677 node = rb_next(new_node); 4678 if (node) { 4679 entry = rb_entry(node, struct ext4_free_data, efd_node); 4680 ext4_try_merge_freed_extent(sbi, entry, new_entry, 4681 &(db->bb_free_root)); 4682 } 4683 4684 spin_lock(&sbi->s_md_lock); 4685 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); 4686 sbi->s_mb_free_pending += clusters; 4687 spin_unlock(&sbi->s_md_lock); 4688 return 0; 4689 } 4690 4691 /** 4692 * ext4_free_blocks() -- Free given blocks and update quota 4693 * @handle: handle for this transaction 4694 * @inode: inode 4695 * @block: start physical block to free 4696 * @count: number of blocks to count 4697 * @flags: flags used by ext4_free_blocks 4698 */ 4699 void ext4_free_blocks(handle_t *handle, struct inode *inode, 4700 struct buffer_head *bh, ext4_fsblk_t block, 4701 unsigned long count, int flags) 4702 { 4703 struct buffer_head *bitmap_bh = NULL; 4704 struct super_block *sb = inode->i_sb; 4705 struct ext4_group_desc *gdp; 4706 unsigned int overflow; 4707 ext4_grpblk_t bit; 4708 struct buffer_head *gd_bh; 4709 ext4_group_t block_group; 4710 struct ext4_sb_info *sbi; 4711 struct ext4_buddy e4b; 4712 unsigned int count_clusters; 4713 int err = 0; 4714 int ret; 4715 4716 might_sleep(); 4717 if (bh) { 4718 if (block) 4719 BUG_ON(block != bh->b_blocknr); 4720 else 4721 block = bh->b_blocknr; 4722 } 4723 4724 sbi = EXT4_SB(sb); 4725 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 4726 !ext4_data_block_valid(sbi, block, count)) { 4727 ext4_error(sb, "Freeing blocks not in datazone - " 4728 "block = %llu, count = %lu", block, count); 4729 goto error_return; 4730 } 4731 4732 ext4_debug("freeing block %llu\n", block); 4733 trace_ext4_free_blocks(inode, block, count, flags); 4734 4735 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 4736 BUG_ON(count > 1); 4737 4738 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 4739 inode, bh, block); 4740 } 4741 4742 /* 4743 * If the extent to be freed does not begin on a cluster 4744 * boundary, we need to deal with partial clusters at the 4745 * beginning and end of the extent. Normally we will free 4746 * blocks at the beginning or the end unless we are explicitly 4747 * requested to avoid doing so. 4748 */ 4749 overflow = EXT4_PBLK_COFF(sbi, block); 4750 if (overflow) { 4751 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 4752 overflow = sbi->s_cluster_ratio - overflow; 4753 block += overflow; 4754 if (count > overflow) 4755 count -= overflow; 4756 else 4757 return; 4758 } else { 4759 block -= overflow; 4760 count += overflow; 4761 } 4762 } 4763 overflow = EXT4_LBLK_COFF(sbi, count); 4764 if (overflow) { 4765 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 4766 if (count > overflow) 4767 count -= overflow; 4768 else 4769 return; 4770 } else 4771 count += sbi->s_cluster_ratio - overflow; 4772 } 4773 4774 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 4775 int i; 4776 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; 4777 4778 for (i = 0; i < count; i++) { 4779 cond_resched(); 4780 if (is_metadata) 4781 bh = sb_find_get_block(inode->i_sb, block + i); 4782 ext4_forget(handle, is_metadata, inode, bh, block + i); 4783 } 4784 } 4785 4786 do_more: 4787 overflow = 0; 4788 ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4789 4790 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( 4791 ext4_get_group_info(sb, block_group)))) 4792 return; 4793 4794 /* 4795 * Check to see if we are freeing blocks across a group 4796 * boundary. 4797 */ 4798 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 4799 overflow = EXT4_C2B(sbi, bit) + count - 4800 EXT4_BLOCKS_PER_GROUP(sb); 4801 count -= overflow; 4802 } 4803 count_clusters = EXT4_NUM_B2C(sbi, count); 4804 bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4805 if (IS_ERR(bitmap_bh)) { 4806 err = PTR_ERR(bitmap_bh); 4807 bitmap_bh = NULL; 4808 goto error_return; 4809 } 4810 gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4811 if (!gdp) { 4812 err = -EIO; 4813 goto error_return; 4814 } 4815 4816 if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4817 in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4818 in_range(block, ext4_inode_table(sb, gdp), 4819 sbi->s_itb_per_group) || 4820 in_range(block + count - 1, ext4_inode_table(sb, gdp), 4821 sbi->s_itb_per_group)) { 4822 4823 ext4_error(sb, "Freeing blocks in system zone - " 4824 "Block = %llu, count = %lu", block, count); 4825 /* err = 0. ext4_std_error should be a no op */ 4826 goto error_return; 4827 } 4828 4829 BUFFER_TRACE(bitmap_bh, "getting write access"); 4830 err = ext4_journal_get_write_access(handle, bitmap_bh); 4831 if (err) 4832 goto error_return; 4833 4834 /* 4835 * We are about to modify some metadata. Call the journal APIs 4836 * to unshare ->b_data if a currently-committing transaction is 4837 * using it 4838 */ 4839 BUFFER_TRACE(gd_bh, "get_write_access"); 4840 err = ext4_journal_get_write_access(handle, gd_bh); 4841 if (err) 4842 goto error_return; 4843 #ifdef AGGRESSIVE_CHECK 4844 { 4845 int i; 4846 for (i = 0; i < count_clusters; i++) 4847 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4848 } 4849 #endif 4850 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 4851 4852 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ 4853 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, 4854 GFP_NOFS|__GFP_NOFAIL); 4855 if (err) 4856 goto error_return; 4857 4858 /* 4859 * We need to make sure we don't reuse the freed block until after the 4860 * transaction is committed. We make an exception if the inode is to be 4861 * written in writeback mode since writeback mode has weak data 4862 * consistency guarantees. 4863 */ 4864 if (ext4_handle_valid(handle) && 4865 ((flags & EXT4_FREE_BLOCKS_METADATA) || 4866 !ext4_should_writeback_data(inode))) { 4867 struct ext4_free_data *new_entry; 4868 /* 4869 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed 4870 * to fail. 4871 */ 4872 new_entry = kmem_cache_alloc(ext4_free_data_cachep, 4873 GFP_NOFS|__GFP_NOFAIL); 4874 new_entry->efd_start_cluster = bit; 4875 new_entry->efd_group = block_group; 4876 new_entry->efd_count = count_clusters; 4877 new_entry->efd_tid = handle->h_transaction->t_tid; 4878 4879 ext4_lock_group(sb, block_group); 4880 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 4881 ext4_mb_free_metadata(handle, &e4b, new_entry); 4882 } else { 4883 /* need to update group_info->bb_free and bitmap 4884 * with group lock held. generate_buddy look at 4885 * them with group lock_held 4886 */ 4887 if (test_opt(sb, DISCARD)) { 4888 err = ext4_issue_discard(sb, block_group, bit, count, 4889 NULL); 4890 if (err && err != -EOPNOTSUPP) 4891 ext4_msg(sb, KERN_WARNING, "discard request in" 4892 " group:%d block:%d count:%lu failed" 4893 " with %d", block_group, bit, count, 4894 err); 4895 } else 4896 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); 4897 4898 ext4_lock_group(sb, block_group); 4899 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 4900 mb_free_blocks(inode, &e4b, bit, count_clusters); 4901 } 4902 4903 ret = ext4_free_group_clusters(sb, gdp) + count_clusters; 4904 ext4_free_group_clusters_set(sb, gdp, ret); 4905 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); 4906 ext4_group_desc_csum_set(sb, block_group, gdp); 4907 ext4_unlock_group(sb, block_group); 4908 4909 if (sbi->s_log_groups_per_flex) { 4910 ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 4911 atomic64_add(count_clusters, 4912 &sbi->s_flex_groups[flex_group].free_clusters); 4913 } 4914 4915 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 4916 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 4917 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); 4918 4919 ext4_mb_unload_buddy(&e4b); 4920 4921 /* We dirtied the bitmap block */ 4922 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 4923 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4924 4925 /* And the group descriptor block */ 4926 BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 4927 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4928 if (!err) 4929 err = ret; 4930 4931 if (overflow && !err) { 4932 block += count; 4933 count = overflow; 4934 put_bh(bitmap_bh); 4935 goto do_more; 4936 } 4937 error_return: 4938 brelse(bitmap_bh); 4939 ext4_std_error(sb, err); 4940 return; 4941 } 4942 4943 /** 4944 * ext4_group_add_blocks() -- Add given blocks to an existing group 4945 * @handle: handle to this transaction 4946 * @sb: super block 4947 * @block: start physical block to add to the block group 4948 * @count: number of blocks to free 4949 * 4950 * This marks the blocks as free in the bitmap and buddy. 4951 */ 4952 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 4953 ext4_fsblk_t block, unsigned long count) 4954 { 4955 struct buffer_head *bitmap_bh = NULL; 4956 struct buffer_head *gd_bh; 4957 ext4_group_t block_group; 4958 ext4_grpblk_t bit; 4959 unsigned int i; 4960 struct ext4_group_desc *desc; 4961 struct ext4_sb_info *sbi = EXT4_SB(sb); 4962 struct ext4_buddy e4b; 4963 int err = 0, ret, free_clusters_count; 4964 ext4_grpblk_t clusters_freed; 4965 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); 4966 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); 4967 unsigned long cluster_count = last_cluster - first_cluster + 1; 4968 4969 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 4970 4971 if (count == 0) 4972 return 0; 4973 4974 ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4975 /* 4976 * Check to see if we are freeing blocks across a group 4977 * boundary. 4978 */ 4979 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { 4980 ext4_warning(sb, "too many blocks added to group %u", 4981 block_group); 4982 err = -EINVAL; 4983 goto error_return; 4984 } 4985 4986 bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4987 if (IS_ERR(bitmap_bh)) { 4988 err = PTR_ERR(bitmap_bh); 4989 bitmap_bh = NULL; 4990 goto error_return; 4991 } 4992 4993 desc = ext4_get_group_desc(sb, block_group, &gd_bh); 4994 if (!desc) { 4995 err = -EIO; 4996 goto error_return; 4997 } 4998 4999 if (in_range(ext4_block_bitmap(sb, desc), block, count) || 5000 in_range(ext4_inode_bitmap(sb, desc), block, count) || 5001 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || 5002 in_range(block + count - 1, ext4_inode_table(sb, desc), 5003 sbi->s_itb_per_group)) { 5004 ext4_error(sb, "Adding blocks in system zones - " 5005 "Block = %llu, count = %lu", 5006 block, count); 5007 err = -EINVAL; 5008 goto error_return; 5009 } 5010 5011 BUFFER_TRACE(bitmap_bh, "getting write access"); 5012 err = ext4_journal_get_write_access(handle, bitmap_bh); 5013 if (err) 5014 goto error_return; 5015 5016 /* 5017 * We are about to modify some metadata. Call the journal APIs 5018 * to unshare ->b_data if a currently-committing transaction is 5019 * using it 5020 */ 5021 BUFFER_TRACE(gd_bh, "get_write_access"); 5022 err = ext4_journal_get_write_access(handle, gd_bh); 5023 if (err) 5024 goto error_return; 5025 5026 for (i = 0, clusters_freed = 0; i < cluster_count; i++) { 5027 BUFFER_TRACE(bitmap_bh, "clear bit"); 5028 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { 5029 ext4_error(sb, "bit already cleared for block %llu", 5030 (ext4_fsblk_t)(block + i)); 5031 BUFFER_TRACE(bitmap_bh, "bit already cleared"); 5032 } else { 5033 clusters_freed++; 5034 } 5035 } 5036 5037 err = ext4_mb_load_buddy(sb, block_group, &e4b); 5038 if (err) 5039 goto error_return; 5040 5041 /* 5042 * need to update group_info->bb_free and bitmap 5043 * with group lock held. generate_buddy look at 5044 * them with group lock_held 5045 */ 5046 ext4_lock_group(sb, block_group); 5047 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); 5048 mb_free_blocks(NULL, &e4b, bit, cluster_count); 5049 free_clusters_count = clusters_freed + 5050 ext4_free_group_clusters(sb, desc); 5051 ext4_free_group_clusters_set(sb, desc, free_clusters_count); 5052 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); 5053 ext4_group_desc_csum_set(sb, block_group, desc); 5054 ext4_unlock_group(sb, block_group); 5055 percpu_counter_add(&sbi->s_freeclusters_counter, 5056 clusters_freed); 5057 5058 if (sbi->s_log_groups_per_flex) { 5059 ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 5060 atomic64_add(clusters_freed, 5061 &sbi->s_flex_groups[flex_group].free_clusters); 5062 } 5063 5064 ext4_mb_unload_buddy(&e4b); 5065 5066 /* We dirtied the bitmap block */ 5067 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 5068 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 5069 5070 /* And the group descriptor block */ 5071 BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 5072 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 5073 if (!err) 5074 err = ret; 5075 5076 error_return: 5077 brelse(bitmap_bh); 5078 ext4_std_error(sb, err); 5079 return err; 5080 } 5081 5082 /** 5083 * ext4_trim_extent -- function to TRIM one single free extent in the group 5084 * @sb: super block for the file system 5085 * @start: starting block of the free extent in the alloc. group 5086 * @count: number of blocks to TRIM 5087 * @group: alloc. group we are working with 5088 * @e4b: ext4 buddy for the group 5089 * 5090 * Trim "count" blocks starting at "start" in the "group". To assure that no 5091 * one will allocate those blocks, mark it as used in buddy bitmap. This must 5092 * be called with under the group lock. 5093 */ 5094 static int ext4_trim_extent(struct super_block *sb, int start, int count, 5095 ext4_group_t group, struct ext4_buddy *e4b) 5096 __releases(bitlock) 5097 __acquires(bitlock) 5098 { 5099 struct ext4_free_extent ex; 5100 int ret = 0; 5101 5102 trace_ext4_trim_extent(sb, group, start, count); 5103 5104 assert_spin_locked(ext4_group_lock_ptr(sb, group)); 5105 5106 ex.fe_start = start; 5107 ex.fe_group = group; 5108 ex.fe_len = count; 5109 5110 /* 5111 * Mark blocks used, so no one can reuse them while 5112 * being trimmed. 5113 */ 5114 mb_mark_used(e4b, &ex); 5115 ext4_unlock_group(sb, group); 5116 ret = ext4_issue_discard(sb, group, start, count, NULL); 5117 ext4_lock_group(sb, group); 5118 mb_free_blocks(NULL, e4b, start, ex.fe_len); 5119 return ret; 5120 } 5121 5122 /** 5123 * ext4_trim_all_free -- function to trim all free space in alloc. group 5124 * @sb: super block for file system 5125 * @group: group to be trimmed 5126 * @start: first group block to examine 5127 * @max: last group block to examine 5128 * @minblocks: minimum extent block count 5129 * 5130 * ext4_trim_all_free walks through group's buddy bitmap searching for free 5131 * extents. When the free block is found, ext4_trim_extent is called to TRIM 5132 * the extent. 5133 * 5134 * 5135 * ext4_trim_all_free walks through group's block bitmap searching for free 5136 * extents. When the free extent is found, mark it as used in group buddy 5137 * bitmap. Then issue a TRIM command on this extent and free the extent in 5138 * the group buddy bitmap. This is done until whole group is scanned. 5139 */ 5140 static ext4_grpblk_t 5141 ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 5142 ext4_grpblk_t start, ext4_grpblk_t max, 5143 ext4_grpblk_t minblocks) 5144 { 5145 void *bitmap; 5146 ext4_grpblk_t next, count = 0, free_count = 0; 5147 struct ext4_buddy e4b; 5148 int ret = 0; 5149 5150 trace_ext4_trim_all_free(sb, group, start, max); 5151 5152 ret = ext4_mb_load_buddy(sb, group, &e4b); 5153 if (ret) { 5154 ext4_warning(sb, "Error %d loading buddy information for %u", 5155 ret, group); 5156 return ret; 5157 } 5158 bitmap = e4b.bd_bitmap; 5159 5160 ext4_lock_group(sb, group); 5161 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && 5162 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) 5163 goto out; 5164 5165 start = (e4b.bd_info->bb_first_free > start) ? 5166 e4b.bd_info->bb_first_free : start; 5167 5168 while (start <= max) { 5169 start = mb_find_next_zero_bit(bitmap, max + 1, start); 5170 if (start > max) 5171 break; 5172 next = mb_find_next_bit(bitmap, max + 1, start); 5173 5174 if ((next - start) >= minblocks) { 5175 ret = ext4_trim_extent(sb, start, 5176 next - start, group, &e4b); 5177 if (ret && ret != -EOPNOTSUPP) 5178 break; 5179 ret = 0; 5180 count += next - start; 5181 } 5182 free_count += next - start; 5183 start = next + 1; 5184 5185 if (fatal_signal_pending(current)) { 5186 count = -ERESTARTSYS; 5187 break; 5188 } 5189 5190 if (need_resched()) { 5191 ext4_unlock_group(sb, group); 5192 cond_resched(); 5193 ext4_lock_group(sb, group); 5194 } 5195 5196 if ((e4b.bd_info->bb_free - free_count) < minblocks) 5197 break; 5198 } 5199 5200 if (!ret) { 5201 ret = count; 5202 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); 5203 } 5204 out: 5205 ext4_unlock_group(sb, group); 5206 ext4_mb_unload_buddy(&e4b); 5207 5208 ext4_debug("trimmed %d blocks in the group %d\n", 5209 count, group); 5210 5211 return ret; 5212 } 5213 5214 /** 5215 * ext4_trim_fs() -- trim ioctl handle function 5216 * @sb: superblock for filesystem 5217 * @range: fstrim_range structure 5218 * 5219 * start: First Byte to trim 5220 * len: number of Bytes to trim from start 5221 * minlen: minimum extent length in Bytes 5222 * ext4_trim_fs goes through all allocation groups containing Bytes from 5223 * start to start+len. For each such a group ext4_trim_all_free function 5224 * is invoked to trim all free space. 5225 */ 5226 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 5227 { 5228 struct ext4_group_info *grp; 5229 ext4_group_t group, first_group, last_group; 5230 ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 5231 uint64_t start, end, minlen, trimmed = 0; 5232 ext4_fsblk_t first_data_blk = 5233 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 5234 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 5235 int ret = 0; 5236 5237 start = range->start >> sb->s_blocksize_bits; 5238 end = start + (range->len >> sb->s_blocksize_bits) - 1; 5239 minlen = EXT4_NUM_B2C(EXT4_SB(sb), 5240 range->minlen >> sb->s_blocksize_bits); 5241 5242 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || 5243 start >= max_blks || 5244 range->len < sb->s_blocksize) 5245 return -EINVAL; 5246 if (end >= max_blks) 5247 end = max_blks - 1; 5248 if (end <= first_data_blk) 5249 goto out; 5250 if (start < first_data_blk) 5251 start = first_data_blk; 5252 5253 /* Determine first and last group to examine based on start and end */ 5254 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 5255 &first_group, &first_cluster); 5256 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 5257 &last_group, &last_cluster); 5258 5259 /* end now represents the last cluster to discard in this group */ 5260 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 5261 5262 for (group = first_group; group <= last_group; group++) { 5263 grp = ext4_get_group_info(sb, group); 5264 /* We only do this if the grp has never been initialized */ 5265 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 5266 ret = ext4_mb_init_group(sb, group, GFP_NOFS); 5267 if (ret) 5268 break; 5269 } 5270 5271 /* 5272 * For all the groups except the last one, last cluster will 5273 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 5274 * change it for the last group, note that last_cluster is 5275 * already computed earlier by ext4_get_group_no_and_offset() 5276 */ 5277 if (group == last_group) 5278 end = last_cluster; 5279 5280 if (grp->bb_free >= minlen) { 5281 cnt = ext4_trim_all_free(sb, group, first_cluster, 5282 end, minlen); 5283 if (cnt < 0) { 5284 ret = cnt; 5285 break; 5286 } 5287 trimmed += cnt; 5288 } 5289 5290 /* 5291 * For every group except the first one, we are sure 5292 * that the first cluster to discard will be cluster #0. 5293 */ 5294 first_cluster = 0; 5295 } 5296 5297 if (!ret) 5298 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); 5299 5300 out: 5301 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; 5302 return ret; 5303 } 5304 5305 /* Iterate all the free extents in the group. */ 5306 int 5307 ext4_mballoc_query_range( 5308 struct super_block *sb, 5309 ext4_group_t group, 5310 ext4_grpblk_t start, 5311 ext4_grpblk_t end, 5312 ext4_mballoc_query_range_fn formatter, 5313 void *priv) 5314 { 5315 void *bitmap; 5316 ext4_grpblk_t next; 5317 struct ext4_buddy e4b; 5318 int error; 5319 5320 error = ext4_mb_load_buddy(sb, group, &e4b); 5321 if (error) 5322 return error; 5323 bitmap = e4b.bd_bitmap; 5324 5325 ext4_lock_group(sb, group); 5326 5327 start = (e4b.bd_info->bb_first_free > start) ? 5328 e4b.bd_info->bb_first_free : start; 5329 if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) 5330 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 5331 5332 while (start <= end) { 5333 start = mb_find_next_zero_bit(bitmap, end + 1, start); 5334 if (start > end) 5335 break; 5336 next = mb_find_next_bit(bitmap, end + 1, start); 5337 5338 ext4_unlock_group(sb, group); 5339 error = formatter(sb, group, start, next - start, priv); 5340 if (error) 5341 goto out_unload; 5342 ext4_lock_group(sb, group); 5343 5344 start = next + 1; 5345 } 5346 5347 ext4_unlock_group(sb, group); 5348 out_unload: 5349 ext4_mb_unload_buddy(&e4b); 5350 5351 return error; 5352 } 5353