1 /* 2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 3 * Written by Alex Tomas <alex@clusterfs.com> 4 * 5 * This program is free software; you can redistribute it and/or modify 6 * it under the terms of the GNU General Public License version 2 as 7 * published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it will be useful, 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 * GNU General Public License for more details. 13 * 14 * You should have received a copy of the GNU General Public License 15 * along with this program; if not, write to the Free Software 16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- 17 */ 18 19 20 /* 21 * mballoc.c contains the multiblocks allocation routines 22 */ 23 24 #include "ext4_jbd2.h" 25 #include "mballoc.h" 26 #include <linux/log2.h> 27 #include <linux/module.h> 28 #include <linux/slab.h> 29 #include <linux/backing-dev.h> 30 #include <trace/events/ext4.h> 31 32 #ifdef CONFIG_EXT4_DEBUG 33 ushort ext4_mballoc_debug __read_mostly; 34 35 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); 36 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); 37 #endif 38 39 /* 40 * MUSTDO: 41 * - test ext4_ext_search_left() and ext4_ext_search_right() 42 * - search for metadata in few groups 43 * 44 * TODO v4: 45 * - normalization should take into account whether file is still open 46 * - discard preallocations if no free space left (policy?) 47 * - don't normalize tails 48 * - quota 49 * - reservation for superuser 50 * 51 * TODO v3: 52 * - bitmap read-ahead (proposed by Oleg Drokin aka green) 53 * - track min/max extents in each group for better group selection 54 * - mb_mark_used() may allocate chunk right after splitting buddy 55 * - tree of groups sorted by number of free blocks 56 * - error handling 57 */ 58 59 /* 60 * The allocation request involve request for multiple number of blocks 61 * near to the goal(block) value specified. 62 * 63 * During initialization phase of the allocator we decide to use the 64 * group preallocation or inode preallocation depending on the size of 65 * the file. The size of the file could be the resulting file size we 66 * would have after allocation, or the current file size, which ever 67 * is larger. If the size is less than sbi->s_mb_stream_request we 68 * select to use the group preallocation. The default value of 69 * s_mb_stream_request is 16 blocks. This can also be tuned via 70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 71 * terms of number of blocks. 72 * 73 * The main motivation for having small file use group preallocation is to 74 * ensure that we have small files closer together on the disk. 75 * 76 * First stage the allocator looks at the inode prealloc list, 77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc 78 * spaces for this particular inode. The inode prealloc space is 79 * represented as: 80 * 81 * pa_lstart -> the logical start block for this prealloc space 82 * pa_pstart -> the physical start block for this prealloc space 83 * pa_len -> length for this prealloc space (in clusters) 84 * pa_free -> free space available in this prealloc space (in clusters) 85 * 86 * The inode preallocation space is used looking at the _logical_ start 87 * block. If only the logical file block falls within the range of prealloc 88 * space we will consume the particular prealloc space. This makes sure that 89 * we have contiguous physical blocks representing the file blocks 90 * 91 * The important thing to be noted in case of inode prealloc space is that 92 * we don't modify the values associated to inode prealloc space except 93 * pa_free. 94 * 95 * If we are not able to find blocks in the inode prealloc space and if we 96 * have the group allocation flag set then we look at the locality group 97 * prealloc space. These are per CPU prealloc list represented as 98 * 99 * ext4_sb_info.s_locality_groups[smp_processor_id()] 100 * 101 * The reason for having a per cpu locality group is to reduce the contention 102 * between CPUs. It is possible to get scheduled at this point. 103 * 104 * The locality group prealloc space is used looking at whether we have 105 * enough free space (pa_free) within the prealloc space. 106 * 107 * If we can't allocate blocks via inode prealloc or/and locality group 108 * prealloc then we look at the buddy cache. The buddy cache is represented 109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 110 * mapped to the buddy and bitmap information regarding different 111 * groups. The buddy information is attached to buddy cache inode so that 112 * we can access them through the page cache. The information regarding 113 * each group is loaded via ext4_mb_load_buddy. The information involve 114 * block bitmap and buddy information. The information are stored in the 115 * inode as: 116 * 117 * { page } 118 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 119 * 120 * 121 * one block each for bitmap and buddy information. So for each group we 122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / 123 * blocksize) blocks. So it can have information regarding groups_per_page 124 * which is blocks_per_page/2 125 * 126 * The buddy cache inode is not stored on disk. The inode is thrown 127 * away when the filesystem is unmounted. 128 * 129 * We look for count number of blocks in the buddy cache. If we were able 130 * to locate that many free blocks we return with additional information 131 * regarding rest of the contiguous physical block available 132 * 133 * Before allocating blocks via buddy cache we normalize the request 134 * blocks. This ensure we ask for more blocks that we needed. The extra 135 * blocks that we get after allocation is added to the respective prealloc 136 * list. In case of inode preallocation we follow a list of heuristics 137 * based on file size. This can be found in ext4_mb_normalize_request. If 138 * we are doing a group prealloc we try to normalize the request to 139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 140 * dependent on the cluster size; for non-bigalloc file systems, it is 141 * 512 blocks. This can be tuned via 142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 143 * terms of number of blocks. If we have mounted the file system with -O 144 * stripe=<value> option the group prealloc request is normalized to the 145 * the smallest multiple of the stripe value (sbi->s_stripe) which is 146 * greater than the default mb_group_prealloc. 147 * 148 * The regular allocator (using the buddy cache) supports a few tunables. 149 * 150 * /sys/fs/ext4/<partition>/mb_min_to_scan 151 * /sys/fs/ext4/<partition>/mb_max_to_scan 152 * /sys/fs/ext4/<partition>/mb_order2_req 153 * 154 * The regular allocator uses buddy scan only if the request len is power of 155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 156 * value of s_mb_order2_reqs can be tuned via 157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 158 * stripe size (sbi->s_stripe), we try to search for contiguous block in 159 * stripe size. This should result in better allocation on RAID setups. If 160 * not, we search in the specific group using bitmap for best extents. The 161 * tunable min_to_scan and max_to_scan control the behaviour here. 162 * min_to_scan indicate how long the mballoc __must__ look for a best 163 * extent and max_to_scan indicates how long the mballoc __can__ look for a 164 * best extent in the found extents. Searching for the blocks starts with 165 * the group specified as the goal value in allocation context via 166 * ac_g_ex. Each group is first checked based on the criteria whether it 167 * can be used for allocation. ext4_mb_good_group explains how the groups are 168 * checked. 169 * 170 * Both the prealloc space are getting populated as above. So for the first 171 * request we will hit the buddy cache which will result in this prealloc 172 * space getting filled. The prealloc space is then later used for the 173 * subsequent request. 174 */ 175 176 /* 177 * mballoc operates on the following data: 178 * - on-disk bitmap 179 * - in-core buddy (actually includes buddy and bitmap) 180 * - preallocation descriptors (PAs) 181 * 182 * there are two types of preallocations: 183 * - inode 184 * assiged to specific inode and can be used for this inode only. 185 * it describes part of inode's space preallocated to specific 186 * physical blocks. any block from that preallocated can be used 187 * independent. the descriptor just tracks number of blocks left 188 * unused. so, before taking some block from descriptor, one must 189 * make sure corresponded logical block isn't allocated yet. this 190 * also means that freeing any block within descriptor's range 191 * must discard all preallocated blocks. 192 * - locality group 193 * assigned to specific locality group which does not translate to 194 * permanent set of inodes: inode can join and leave group. space 195 * from this type of preallocation can be used for any inode. thus 196 * it's consumed from the beginning to the end. 197 * 198 * relation between them can be expressed as: 199 * in-core buddy = on-disk bitmap + preallocation descriptors 200 * 201 * this mean blocks mballoc considers used are: 202 * - allocated blocks (persistent) 203 * - preallocated blocks (non-persistent) 204 * 205 * consistency in mballoc world means that at any time a block is either 206 * free or used in ALL structures. notice: "any time" should not be read 207 * literally -- time is discrete and delimited by locks. 208 * 209 * to keep it simple, we don't use block numbers, instead we count number of 210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 211 * 212 * all operations can be expressed as: 213 * - init buddy: buddy = on-disk + PAs 214 * - new PA: buddy += N; PA = N 215 * - use inode PA: on-disk += N; PA -= N 216 * - discard inode PA buddy -= on-disk - PA; PA = 0 217 * - use locality group PA on-disk += N; PA -= N 218 * - discard locality group PA buddy -= PA; PA = 0 219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 220 * is used in real operation because we can't know actual used 221 * bits from PA, only from on-disk bitmap 222 * 223 * if we follow this strict logic, then all operations above should be atomic. 224 * given some of them can block, we'd have to use something like semaphores 225 * killing performance on high-end SMP hardware. let's try to relax it using 226 * the following knowledge: 227 * 1) if buddy is referenced, it's already initialized 228 * 2) while block is used in buddy and the buddy is referenced, 229 * nobody can re-allocate that block 230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 231 * bit set and PA claims same block, it's OK. IOW, one can set bit in 232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 233 * block 234 * 235 * so, now we're building a concurrency table: 236 * - init buddy vs. 237 * - new PA 238 * blocks for PA are allocated in the buddy, buddy must be referenced 239 * until PA is linked to allocation group to avoid concurrent buddy init 240 * - use inode PA 241 * we need to make sure that either on-disk bitmap or PA has uptodate data 242 * given (3) we care that PA-=N operation doesn't interfere with init 243 * - discard inode PA 244 * the simplest way would be to have buddy initialized by the discard 245 * - use locality group PA 246 * again PA-=N must be serialized with init 247 * - discard locality group PA 248 * the simplest way would be to have buddy initialized by the discard 249 * - new PA vs. 250 * - use inode PA 251 * i_data_sem serializes them 252 * - discard inode PA 253 * discard process must wait until PA isn't used by another process 254 * - use locality group PA 255 * some mutex should serialize them 256 * - discard locality group PA 257 * discard process must wait until PA isn't used by another process 258 * - use inode PA 259 * - use inode PA 260 * i_data_sem or another mutex should serializes them 261 * - discard inode PA 262 * discard process must wait until PA isn't used by another process 263 * - use locality group PA 264 * nothing wrong here -- they're different PAs covering different blocks 265 * - discard locality group PA 266 * discard process must wait until PA isn't used by another process 267 * 268 * now we're ready to make few consequences: 269 * - PA is referenced and while it is no discard is possible 270 * - PA is referenced until block isn't marked in on-disk bitmap 271 * - PA changes only after on-disk bitmap 272 * - discard must not compete with init. either init is done before 273 * any discard or they're serialized somehow 274 * - buddy init as sum of on-disk bitmap and PAs is done atomically 275 * 276 * a special case when we've used PA to emptiness. no need to modify buddy 277 * in this case, but we should care about concurrent init 278 * 279 */ 280 281 /* 282 * Logic in few words: 283 * 284 * - allocation: 285 * load group 286 * find blocks 287 * mark bits in on-disk bitmap 288 * release group 289 * 290 * - use preallocation: 291 * find proper PA (per-inode or group) 292 * load group 293 * mark bits in on-disk bitmap 294 * release group 295 * release PA 296 * 297 * - free: 298 * load group 299 * mark bits in on-disk bitmap 300 * release group 301 * 302 * - discard preallocations in group: 303 * mark PAs deleted 304 * move them onto local list 305 * load on-disk bitmap 306 * load group 307 * remove PA from object (inode or locality group) 308 * mark free blocks in-core 309 * 310 * - discard inode's preallocations: 311 */ 312 313 /* 314 * Locking rules 315 * 316 * Locks: 317 * - bitlock on a group (group) 318 * - object (inode/locality) (object) 319 * - per-pa lock (pa) 320 * 321 * Paths: 322 * - new pa 323 * object 324 * group 325 * 326 * - find and use pa: 327 * pa 328 * 329 * - release consumed pa: 330 * pa 331 * group 332 * object 333 * 334 * - generate in-core bitmap: 335 * group 336 * pa 337 * 338 * - discard all for given object (inode, locality group): 339 * object 340 * pa 341 * group 342 * 343 * - discard all for given group: 344 * group 345 * pa 346 * group 347 * object 348 * 349 */ 350 static struct kmem_cache *ext4_pspace_cachep; 351 static struct kmem_cache *ext4_ac_cachep; 352 static struct kmem_cache *ext4_free_data_cachep; 353 354 /* We create slab caches for groupinfo data structures based on the 355 * superblock block size. There will be one per mounted filesystem for 356 * each unique s_blocksize_bits */ 357 #define NR_GRPINFO_CACHES 8 358 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 359 360 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 363 "ext4_groupinfo_64k", "ext4_groupinfo_128k" 364 }; 365 366 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 367 ext4_group_t group); 368 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 369 ext4_group_t group); 370 371 static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 372 { 373 #if BITS_PER_LONG == 64 374 *bit += ((unsigned long) addr & 7UL) << 3; 375 addr = (void *) ((unsigned long) addr & ~7UL); 376 #elif BITS_PER_LONG == 32 377 *bit += ((unsigned long) addr & 3UL) << 3; 378 addr = (void *) ((unsigned long) addr & ~3UL); 379 #else 380 #error "how many bits you are?!" 381 #endif 382 return addr; 383 } 384 385 static inline int mb_test_bit(int bit, void *addr) 386 { 387 /* 388 * ext4_test_bit on architecture like powerpc 389 * needs unsigned long aligned address 390 */ 391 addr = mb_correct_addr_and_bit(&bit, addr); 392 return ext4_test_bit(bit, addr); 393 } 394 395 static inline void mb_set_bit(int bit, void *addr) 396 { 397 addr = mb_correct_addr_and_bit(&bit, addr); 398 ext4_set_bit(bit, addr); 399 } 400 401 static inline void mb_clear_bit(int bit, void *addr) 402 { 403 addr = mb_correct_addr_and_bit(&bit, addr); 404 ext4_clear_bit(bit, addr); 405 } 406 407 static inline int mb_test_and_clear_bit(int bit, void *addr) 408 { 409 addr = mb_correct_addr_and_bit(&bit, addr); 410 return ext4_test_and_clear_bit(bit, addr); 411 } 412 413 static inline int mb_find_next_zero_bit(void *addr, int max, int start) 414 { 415 int fix = 0, ret, tmpmax; 416 addr = mb_correct_addr_and_bit(&fix, addr); 417 tmpmax = max + fix; 418 start += fix; 419 420 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 421 if (ret > max) 422 return max; 423 return ret; 424 } 425 426 static inline int mb_find_next_bit(void *addr, int max, int start) 427 { 428 int fix = 0, ret, tmpmax; 429 addr = mb_correct_addr_and_bit(&fix, addr); 430 tmpmax = max + fix; 431 start += fix; 432 433 ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 434 if (ret > max) 435 return max; 436 return ret; 437 } 438 439 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 440 { 441 char *bb; 442 443 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 444 BUG_ON(max == NULL); 445 446 if (order > e4b->bd_blkbits + 1) { 447 *max = 0; 448 return NULL; 449 } 450 451 /* at order 0 we see each particular block */ 452 if (order == 0) { 453 *max = 1 << (e4b->bd_blkbits + 3); 454 return e4b->bd_bitmap; 455 } 456 457 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 458 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 459 460 return bb; 461 } 462 463 #ifdef DOUBLE_CHECK 464 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 465 int first, int count) 466 { 467 int i; 468 struct super_block *sb = e4b->bd_sb; 469 470 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 471 return; 472 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 473 for (i = 0; i < count; i++) { 474 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 475 ext4_fsblk_t blocknr; 476 477 blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 478 blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 479 ext4_grp_locked_error(sb, e4b->bd_group, 480 inode ? inode->i_ino : 0, 481 blocknr, 482 "freeing block already freed " 483 "(bit %u)", 484 first + i); 485 } 486 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 487 } 488 } 489 490 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 491 { 492 int i; 493 494 if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 495 return; 496 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 497 for (i = 0; i < count; i++) { 498 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 499 mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 500 } 501 } 502 503 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 504 { 505 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 506 unsigned char *b1, *b2; 507 int i; 508 b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 509 b2 = (unsigned char *) bitmap; 510 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 511 if (b1[i] != b2[i]) { 512 ext4_msg(e4b->bd_sb, KERN_ERR, 513 "corruption in group %u " 514 "at byte %u(%u): %x in copy != %x " 515 "on disk/prealloc", 516 e4b->bd_group, i, i * 8, b1[i], b2[i]); 517 BUG(); 518 } 519 } 520 } 521 } 522 523 #else 524 static inline void mb_free_blocks_double(struct inode *inode, 525 struct ext4_buddy *e4b, int first, int count) 526 { 527 return; 528 } 529 static inline void mb_mark_used_double(struct ext4_buddy *e4b, 530 int first, int count) 531 { 532 return; 533 } 534 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 535 { 536 return; 537 } 538 #endif 539 540 #ifdef AGGRESSIVE_CHECK 541 542 #define MB_CHECK_ASSERT(assert) \ 543 do { \ 544 if (!(assert)) { \ 545 printk(KERN_EMERG \ 546 "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 547 function, file, line, # assert); \ 548 BUG(); \ 549 } \ 550 } while (0) 551 552 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 553 const char *function, int line) 554 { 555 struct super_block *sb = e4b->bd_sb; 556 int order = e4b->bd_blkbits + 1; 557 int max; 558 int max2; 559 int i; 560 int j; 561 int k; 562 int count; 563 struct ext4_group_info *grp; 564 int fragments = 0; 565 int fstart; 566 struct list_head *cur; 567 void *buddy; 568 void *buddy2; 569 570 { 571 static int mb_check_counter; 572 if (mb_check_counter++ % 100 != 0) 573 return 0; 574 } 575 576 while (order > 1) { 577 buddy = mb_find_buddy(e4b, order, &max); 578 MB_CHECK_ASSERT(buddy); 579 buddy2 = mb_find_buddy(e4b, order - 1, &max2); 580 MB_CHECK_ASSERT(buddy2); 581 MB_CHECK_ASSERT(buddy != buddy2); 582 MB_CHECK_ASSERT(max * 2 == max2); 583 584 count = 0; 585 for (i = 0; i < max; i++) { 586 587 if (mb_test_bit(i, buddy)) { 588 /* only single bit in buddy2 may be 1 */ 589 if (!mb_test_bit(i << 1, buddy2)) { 590 MB_CHECK_ASSERT( 591 mb_test_bit((i<<1)+1, buddy2)); 592 } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 593 MB_CHECK_ASSERT( 594 mb_test_bit(i << 1, buddy2)); 595 } 596 continue; 597 } 598 599 /* both bits in buddy2 must be 1 */ 600 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 601 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 602 603 for (j = 0; j < (1 << order); j++) { 604 k = (i * (1 << order)) + j; 605 MB_CHECK_ASSERT( 606 !mb_test_bit(k, e4b->bd_bitmap)); 607 } 608 count++; 609 } 610 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 611 order--; 612 } 613 614 fstart = -1; 615 buddy = mb_find_buddy(e4b, 0, &max); 616 for (i = 0; i < max; i++) { 617 if (!mb_test_bit(i, buddy)) { 618 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 619 if (fstart == -1) { 620 fragments++; 621 fstart = i; 622 } 623 continue; 624 } 625 fstart = -1; 626 /* check used bits only */ 627 for (j = 0; j < e4b->bd_blkbits + 1; j++) { 628 buddy2 = mb_find_buddy(e4b, j, &max2); 629 k = i >> j; 630 MB_CHECK_ASSERT(k < max2); 631 MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 632 } 633 } 634 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 635 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 636 637 grp = ext4_get_group_info(sb, e4b->bd_group); 638 list_for_each(cur, &grp->bb_prealloc_list) { 639 ext4_group_t groupnr; 640 struct ext4_prealloc_space *pa; 641 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 642 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 643 MB_CHECK_ASSERT(groupnr == e4b->bd_group); 644 for (i = 0; i < pa->pa_len; i++) 645 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 646 } 647 return 0; 648 } 649 #undef MB_CHECK_ASSERT 650 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 651 __FILE__, __func__, __LINE__) 652 #else 653 #define mb_check_buddy(e4b) 654 #endif 655 656 /* 657 * Divide blocks started from @first with length @len into 658 * smaller chunks with power of 2 blocks. 659 * Clear the bits in bitmap which the blocks of the chunk(s) covered, 660 * then increase bb_counters[] for corresponded chunk size. 661 */ 662 static void ext4_mb_mark_free_simple(struct super_block *sb, 663 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 664 struct ext4_group_info *grp) 665 { 666 struct ext4_sb_info *sbi = EXT4_SB(sb); 667 ext4_grpblk_t min; 668 ext4_grpblk_t max; 669 ext4_grpblk_t chunk; 670 unsigned int border; 671 672 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 673 674 border = 2 << sb->s_blocksize_bits; 675 676 while (len > 0) { 677 /* find how many blocks can be covered since this position */ 678 max = ffs(first | border) - 1; 679 680 /* find how many blocks of power 2 we need to mark */ 681 min = fls(len) - 1; 682 683 if (max < min) 684 min = max; 685 chunk = 1 << min; 686 687 /* mark multiblock chunks only */ 688 grp->bb_counters[min]++; 689 if (min > 0) 690 mb_clear_bit(first >> min, 691 buddy + sbi->s_mb_offsets[min]); 692 693 len -= chunk; 694 first += chunk; 695 } 696 } 697 698 /* 699 * Cache the order of the largest free extent we have available in this block 700 * group. 701 */ 702 static void 703 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 704 { 705 int i; 706 int bits; 707 708 grp->bb_largest_free_order = -1; /* uninit */ 709 710 bits = sb->s_blocksize_bits + 1; 711 for (i = bits; i >= 0; i--) { 712 if (grp->bb_counters[i] > 0) { 713 grp->bb_largest_free_order = i; 714 break; 715 } 716 } 717 } 718 719 static noinline_for_stack 720 void ext4_mb_generate_buddy(struct super_block *sb, 721 void *buddy, void *bitmap, ext4_group_t group) 722 { 723 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 724 struct ext4_sb_info *sbi = EXT4_SB(sb); 725 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 726 ext4_grpblk_t i = 0; 727 ext4_grpblk_t first; 728 ext4_grpblk_t len; 729 unsigned free = 0; 730 unsigned fragments = 0; 731 unsigned long long period = get_cycles(); 732 733 /* initialize buddy from bitmap which is aggregation 734 * of on-disk bitmap and preallocations */ 735 i = mb_find_next_zero_bit(bitmap, max, 0); 736 grp->bb_first_free = i; 737 while (i < max) { 738 fragments++; 739 first = i; 740 i = mb_find_next_bit(bitmap, max, i); 741 len = i - first; 742 free += len; 743 if (len > 1) 744 ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 745 else 746 grp->bb_counters[0]++; 747 if (i < max) 748 i = mb_find_next_zero_bit(bitmap, max, i); 749 } 750 grp->bb_fragments = fragments; 751 752 if (free != grp->bb_free) { 753 ext4_grp_locked_error(sb, group, 0, 0, 754 "block bitmap and bg descriptor " 755 "inconsistent: %u vs %u free clusters", 756 free, grp->bb_free); 757 /* 758 * If we intend to continue, we consider group descriptor 759 * corrupt and update bb_free using bitmap value 760 */ 761 grp->bb_free = free; 762 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp)) 763 percpu_counter_sub(&sbi->s_freeclusters_counter, 764 grp->bb_free); 765 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state); 766 } 767 mb_set_largest_free_order(sb, grp); 768 769 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 770 771 period = get_cycles() - period; 772 spin_lock(&EXT4_SB(sb)->s_bal_lock); 773 EXT4_SB(sb)->s_mb_buddies_generated++; 774 EXT4_SB(sb)->s_mb_generation_time += period; 775 spin_unlock(&EXT4_SB(sb)->s_bal_lock); 776 } 777 778 static void mb_regenerate_buddy(struct ext4_buddy *e4b) 779 { 780 int count; 781 int order = 1; 782 void *buddy; 783 784 while ((buddy = mb_find_buddy(e4b, order++, &count))) { 785 ext4_set_bits(buddy, 0, count); 786 } 787 e4b->bd_info->bb_fragments = 0; 788 memset(e4b->bd_info->bb_counters, 0, 789 sizeof(*e4b->bd_info->bb_counters) * 790 (e4b->bd_sb->s_blocksize_bits + 2)); 791 792 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, 793 e4b->bd_bitmap, e4b->bd_group); 794 } 795 796 /* The buddy information is attached the buddy cache inode 797 * for convenience. The information regarding each group 798 * is loaded via ext4_mb_load_buddy. The information involve 799 * block bitmap and buddy information. The information are 800 * stored in the inode as 801 * 802 * { page } 803 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 804 * 805 * 806 * one block each for bitmap and buddy information. 807 * So for each group we take up 2 blocks. A page can 808 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. 809 * So it can have information regarding groups_per_page which 810 * is blocks_per_page/2 811 * 812 * Locking note: This routine takes the block group lock of all groups 813 * for this page; do not hold this lock when calling this routine! 814 */ 815 816 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) 817 { 818 ext4_group_t ngroups; 819 int blocksize; 820 int blocks_per_page; 821 int groups_per_page; 822 int err = 0; 823 int i; 824 ext4_group_t first_group, group; 825 int first_block; 826 struct super_block *sb; 827 struct buffer_head *bhs; 828 struct buffer_head **bh = NULL; 829 struct inode *inode; 830 char *data; 831 char *bitmap; 832 struct ext4_group_info *grinfo; 833 834 mb_debug(1, "init page %lu\n", page->index); 835 836 inode = page->mapping->host; 837 sb = inode->i_sb; 838 ngroups = ext4_get_groups_count(sb); 839 blocksize = i_blocksize(inode); 840 blocks_per_page = PAGE_SIZE / blocksize; 841 842 groups_per_page = blocks_per_page >> 1; 843 if (groups_per_page == 0) 844 groups_per_page = 1; 845 846 /* allocate buffer_heads to read bitmaps */ 847 if (groups_per_page > 1) { 848 i = sizeof(struct buffer_head *) * groups_per_page; 849 bh = kzalloc(i, gfp); 850 if (bh == NULL) { 851 err = -ENOMEM; 852 goto out; 853 } 854 } else 855 bh = &bhs; 856 857 first_group = page->index * blocks_per_page / 2; 858 859 /* read all groups the page covers into the cache */ 860 for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 861 if (group >= ngroups) 862 break; 863 864 grinfo = ext4_get_group_info(sb, group); 865 /* 866 * If page is uptodate then we came here after online resize 867 * which added some new uninitialized group info structs, so 868 * we must skip all initialized uptodate buddies on the page, 869 * which may be currently in use by an allocating task. 870 */ 871 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { 872 bh[i] = NULL; 873 continue; 874 } 875 bh[i] = ext4_read_block_bitmap_nowait(sb, group); 876 if (IS_ERR(bh[i])) { 877 err = PTR_ERR(bh[i]); 878 bh[i] = NULL; 879 goto out; 880 } 881 mb_debug(1, "read bitmap for group %u\n", group); 882 } 883 884 /* wait for I/O completion */ 885 for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 886 int err2; 887 888 if (!bh[i]) 889 continue; 890 err2 = ext4_wait_block_bitmap(sb, group, bh[i]); 891 if (!err) 892 err = err2; 893 } 894 895 first_block = page->index * blocks_per_page; 896 for (i = 0; i < blocks_per_page; i++) { 897 group = (first_block + i) >> 1; 898 if (group >= ngroups) 899 break; 900 901 if (!bh[group - first_group]) 902 /* skip initialized uptodate buddy */ 903 continue; 904 905 if (!buffer_verified(bh[group - first_group])) 906 /* Skip faulty bitmaps */ 907 continue; 908 err = 0; 909 910 /* 911 * data carry information regarding this 912 * particular group in the format specified 913 * above 914 * 915 */ 916 data = page_address(page) + (i * blocksize); 917 bitmap = bh[group - first_group]->b_data; 918 919 /* 920 * We place the buddy block and bitmap block 921 * close together 922 */ 923 if ((first_block + i) & 1) { 924 /* this is block of buddy */ 925 BUG_ON(incore == NULL); 926 mb_debug(1, "put buddy for group %u in page %lu/%x\n", 927 group, page->index, i * blocksize); 928 trace_ext4_mb_buddy_bitmap_load(sb, group); 929 grinfo = ext4_get_group_info(sb, group); 930 grinfo->bb_fragments = 0; 931 memset(grinfo->bb_counters, 0, 932 sizeof(*grinfo->bb_counters) * 933 (sb->s_blocksize_bits+2)); 934 /* 935 * incore got set to the group block bitmap below 936 */ 937 ext4_lock_group(sb, group); 938 /* init the buddy */ 939 memset(data, 0xff, blocksize); 940 ext4_mb_generate_buddy(sb, data, incore, group); 941 ext4_unlock_group(sb, group); 942 incore = NULL; 943 } else { 944 /* this is block of bitmap */ 945 BUG_ON(incore != NULL); 946 mb_debug(1, "put bitmap for group %u in page %lu/%x\n", 947 group, page->index, i * blocksize); 948 trace_ext4_mb_bitmap_load(sb, group); 949 950 /* see comments in ext4_mb_put_pa() */ 951 ext4_lock_group(sb, group); 952 memcpy(data, bitmap, blocksize); 953 954 /* mark all preallocated blks used in in-core bitmap */ 955 ext4_mb_generate_from_pa(sb, data, group); 956 ext4_mb_generate_from_freelist(sb, data, group); 957 ext4_unlock_group(sb, group); 958 959 /* set incore so that the buddy information can be 960 * generated using this 961 */ 962 incore = data; 963 } 964 } 965 SetPageUptodate(page); 966 967 out: 968 if (bh) { 969 for (i = 0; i < groups_per_page; i++) 970 brelse(bh[i]); 971 if (bh != &bhs) 972 kfree(bh); 973 } 974 return err; 975 } 976 977 /* 978 * Lock the buddy and bitmap pages. This make sure other parallel init_group 979 * on the same buddy page doesn't happen whild holding the buddy page lock. 980 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 981 * are on the same page e4b->bd_buddy_page is NULL and return value is 0. 982 */ 983 static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 984 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) 985 { 986 struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 987 int block, pnum, poff; 988 int blocks_per_page; 989 struct page *page; 990 991 e4b->bd_buddy_page = NULL; 992 e4b->bd_bitmap_page = NULL; 993 994 blocks_per_page = PAGE_SIZE / sb->s_blocksize; 995 /* 996 * the buddy cache inode stores the block bitmap 997 * and buddy information in consecutive blocks. 998 * So for each group we need two blocks. 999 */ 1000 block = group * 2; 1001 pnum = block / blocks_per_page; 1002 poff = block % blocks_per_page; 1003 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1004 if (!page) 1005 return -ENOMEM; 1006 BUG_ON(page->mapping != inode->i_mapping); 1007 e4b->bd_bitmap_page = page; 1008 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1009 1010 if (blocks_per_page >= 2) { 1011 /* buddy and bitmap are on the same page */ 1012 return 0; 1013 } 1014 1015 block++; 1016 pnum = block / blocks_per_page; 1017 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1018 if (!page) 1019 return -ENOMEM; 1020 BUG_ON(page->mapping != inode->i_mapping); 1021 e4b->bd_buddy_page = page; 1022 return 0; 1023 } 1024 1025 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 1026 { 1027 if (e4b->bd_bitmap_page) { 1028 unlock_page(e4b->bd_bitmap_page); 1029 put_page(e4b->bd_bitmap_page); 1030 } 1031 if (e4b->bd_buddy_page) { 1032 unlock_page(e4b->bd_buddy_page); 1033 put_page(e4b->bd_buddy_page); 1034 } 1035 } 1036 1037 /* 1038 * Locking note: This routine calls ext4_mb_init_cache(), which takes the 1039 * block group lock of all groups for this page; do not hold the BG lock when 1040 * calling this routine! 1041 */ 1042 static noinline_for_stack 1043 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) 1044 { 1045 1046 struct ext4_group_info *this_grp; 1047 struct ext4_buddy e4b; 1048 struct page *page; 1049 int ret = 0; 1050 1051 might_sleep(); 1052 mb_debug(1, "init group %u\n", group); 1053 this_grp = ext4_get_group_info(sb, group); 1054 /* 1055 * This ensures that we don't reinit the buddy cache 1056 * page which map to the group from which we are already 1057 * allocating. If we are looking at the buddy cache we would 1058 * have taken a reference using ext4_mb_load_buddy and that 1059 * would have pinned buddy page to page cache. 1060 * The call to ext4_mb_get_buddy_page_lock will mark the 1061 * page accessed. 1062 */ 1063 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); 1064 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1065 /* 1066 * somebody initialized the group 1067 * return without doing anything 1068 */ 1069 goto err; 1070 } 1071 1072 page = e4b.bd_bitmap_page; 1073 ret = ext4_mb_init_cache(page, NULL, gfp); 1074 if (ret) 1075 goto err; 1076 if (!PageUptodate(page)) { 1077 ret = -EIO; 1078 goto err; 1079 } 1080 1081 if (e4b.bd_buddy_page == NULL) { 1082 /* 1083 * If both the bitmap and buddy are in 1084 * the same page we don't need to force 1085 * init the buddy 1086 */ 1087 ret = 0; 1088 goto err; 1089 } 1090 /* init buddy cache */ 1091 page = e4b.bd_buddy_page; 1092 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); 1093 if (ret) 1094 goto err; 1095 if (!PageUptodate(page)) { 1096 ret = -EIO; 1097 goto err; 1098 } 1099 err: 1100 ext4_mb_put_buddy_page_lock(&e4b); 1101 return ret; 1102 } 1103 1104 /* 1105 * Locking note: This routine calls ext4_mb_init_cache(), which takes the 1106 * block group lock of all groups for this page; do not hold the BG lock when 1107 * calling this routine! 1108 */ 1109 static noinline_for_stack int 1110 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, 1111 struct ext4_buddy *e4b, gfp_t gfp) 1112 { 1113 int blocks_per_page; 1114 int block; 1115 int pnum; 1116 int poff; 1117 struct page *page; 1118 int ret; 1119 struct ext4_group_info *grp; 1120 struct ext4_sb_info *sbi = EXT4_SB(sb); 1121 struct inode *inode = sbi->s_buddy_cache; 1122 1123 might_sleep(); 1124 mb_debug(1, "load group %u\n", group); 1125 1126 blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1127 grp = ext4_get_group_info(sb, group); 1128 1129 e4b->bd_blkbits = sb->s_blocksize_bits; 1130 e4b->bd_info = grp; 1131 e4b->bd_sb = sb; 1132 e4b->bd_group = group; 1133 e4b->bd_buddy_page = NULL; 1134 e4b->bd_bitmap_page = NULL; 1135 1136 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1137 /* 1138 * we need full data about the group 1139 * to make a good selection 1140 */ 1141 ret = ext4_mb_init_group(sb, group, gfp); 1142 if (ret) 1143 return ret; 1144 } 1145 1146 /* 1147 * the buddy cache inode stores the block bitmap 1148 * and buddy information in consecutive blocks. 1149 * So for each group we need two blocks. 1150 */ 1151 block = group * 2; 1152 pnum = block / blocks_per_page; 1153 poff = block % blocks_per_page; 1154 1155 /* we could use find_or_create_page(), but it locks page 1156 * what we'd like to avoid in fast path ... */ 1157 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1158 if (page == NULL || !PageUptodate(page)) { 1159 if (page) 1160 /* 1161 * drop the page reference and try 1162 * to get the page with lock. If we 1163 * are not uptodate that implies 1164 * somebody just created the page but 1165 * is yet to initialize the same. So 1166 * wait for it to initialize. 1167 */ 1168 put_page(page); 1169 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1170 if (page) { 1171 BUG_ON(page->mapping != inode->i_mapping); 1172 if (!PageUptodate(page)) { 1173 ret = ext4_mb_init_cache(page, NULL, gfp); 1174 if (ret) { 1175 unlock_page(page); 1176 goto err; 1177 } 1178 mb_cmp_bitmaps(e4b, page_address(page) + 1179 (poff * sb->s_blocksize)); 1180 } 1181 unlock_page(page); 1182 } 1183 } 1184 if (page == NULL) { 1185 ret = -ENOMEM; 1186 goto err; 1187 } 1188 if (!PageUptodate(page)) { 1189 ret = -EIO; 1190 goto err; 1191 } 1192 1193 /* Pages marked accessed already */ 1194 e4b->bd_bitmap_page = page; 1195 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1196 1197 block++; 1198 pnum = block / blocks_per_page; 1199 poff = block % blocks_per_page; 1200 1201 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1202 if (page == NULL || !PageUptodate(page)) { 1203 if (page) 1204 put_page(page); 1205 page = find_or_create_page(inode->i_mapping, pnum, gfp); 1206 if (page) { 1207 BUG_ON(page->mapping != inode->i_mapping); 1208 if (!PageUptodate(page)) { 1209 ret = ext4_mb_init_cache(page, e4b->bd_bitmap, 1210 gfp); 1211 if (ret) { 1212 unlock_page(page); 1213 goto err; 1214 } 1215 } 1216 unlock_page(page); 1217 } 1218 } 1219 if (page == NULL) { 1220 ret = -ENOMEM; 1221 goto err; 1222 } 1223 if (!PageUptodate(page)) { 1224 ret = -EIO; 1225 goto err; 1226 } 1227 1228 /* Pages marked accessed already */ 1229 e4b->bd_buddy_page = page; 1230 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1231 1232 BUG_ON(e4b->bd_bitmap_page == NULL); 1233 BUG_ON(e4b->bd_buddy_page == NULL); 1234 1235 return 0; 1236 1237 err: 1238 if (page) 1239 put_page(page); 1240 if (e4b->bd_bitmap_page) 1241 put_page(e4b->bd_bitmap_page); 1242 if (e4b->bd_buddy_page) 1243 put_page(e4b->bd_buddy_page); 1244 e4b->bd_buddy = NULL; 1245 e4b->bd_bitmap = NULL; 1246 return ret; 1247 } 1248 1249 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1250 struct ext4_buddy *e4b) 1251 { 1252 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); 1253 } 1254 1255 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1256 { 1257 if (e4b->bd_bitmap_page) 1258 put_page(e4b->bd_bitmap_page); 1259 if (e4b->bd_buddy_page) 1260 put_page(e4b->bd_buddy_page); 1261 } 1262 1263 1264 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1265 { 1266 int order = 1; 1267 int bb_incr = 1 << (e4b->bd_blkbits - 1); 1268 void *bb; 1269 1270 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1271 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1272 1273 bb = e4b->bd_buddy; 1274 while (order <= e4b->bd_blkbits + 1) { 1275 block = block >> 1; 1276 if (!mb_test_bit(block, bb)) { 1277 /* this block is part of buddy of order 'order' */ 1278 return order; 1279 } 1280 bb += bb_incr; 1281 bb_incr >>= 1; 1282 order++; 1283 } 1284 return 0; 1285 } 1286 1287 static void mb_clear_bits(void *bm, int cur, int len) 1288 { 1289 __u32 *addr; 1290 1291 len = cur + len; 1292 while (cur < len) { 1293 if ((cur & 31) == 0 && (len - cur) >= 32) { 1294 /* fast path: clear whole word at once */ 1295 addr = bm + (cur >> 3); 1296 *addr = 0; 1297 cur += 32; 1298 continue; 1299 } 1300 mb_clear_bit(cur, bm); 1301 cur++; 1302 } 1303 } 1304 1305 /* clear bits in given range 1306 * will return first found zero bit if any, -1 otherwise 1307 */ 1308 static int mb_test_and_clear_bits(void *bm, int cur, int len) 1309 { 1310 __u32 *addr; 1311 int zero_bit = -1; 1312 1313 len = cur + len; 1314 while (cur < len) { 1315 if ((cur & 31) == 0 && (len - cur) >= 32) { 1316 /* fast path: clear whole word at once */ 1317 addr = bm + (cur >> 3); 1318 if (*addr != (__u32)(-1) && zero_bit == -1) 1319 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); 1320 *addr = 0; 1321 cur += 32; 1322 continue; 1323 } 1324 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) 1325 zero_bit = cur; 1326 cur++; 1327 } 1328 1329 return zero_bit; 1330 } 1331 1332 void ext4_set_bits(void *bm, int cur, int len) 1333 { 1334 __u32 *addr; 1335 1336 len = cur + len; 1337 while (cur < len) { 1338 if ((cur & 31) == 0 && (len - cur) >= 32) { 1339 /* fast path: set whole word at once */ 1340 addr = bm + (cur >> 3); 1341 *addr = 0xffffffff; 1342 cur += 32; 1343 continue; 1344 } 1345 mb_set_bit(cur, bm); 1346 cur++; 1347 } 1348 } 1349 1350 /* 1351 * _________________________________________________________________ */ 1352 1353 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) 1354 { 1355 if (mb_test_bit(*bit + side, bitmap)) { 1356 mb_clear_bit(*bit, bitmap); 1357 (*bit) -= side; 1358 return 1; 1359 } 1360 else { 1361 (*bit) += side; 1362 mb_set_bit(*bit, bitmap); 1363 return -1; 1364 } 1365 } 1366 1367 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) 1368 { 1369 int max; 1370 int order = 1; 1371 void *buddy = mb_find_buddy(e4b, order, &max); 1372 1373 while (buddy) { 1374 void *buddy2; 1375 1376 /* Bits in range [first; last] are known to be set since 1377 * corresponding blocks were allocated. Bits in range 1378 * (first; last) will stay set because they form buddies on 1379 * upper layer. We just deal with borders if they don't 1380 * align with upper layer and then go up. 1381 * Releasing entire group is all about clearing 1382 * single bit of highest order buddy. 1383 */ 1384 1385 /* Example: 1386 * --------------------------------- 1387 * | 1 | 1 | 1 | 1 | 1388 * --------------------------------- 1389 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1390 * --------------------------------- 1391 * 0 1 2 3 4 5 6 7 1392 * \_____________________/ 1393 * 1394 * Neither [1] nor [6] is aligned to above layer. 1395 * Left neighbour [0] is free, so mark it busy, 1396 * decrease bb_counters and extend range to 1397 * [0; 6] 1398 * Right neighbour [7] is busy. It can't be coaleasced with [6], so 1399 * mark [6] free, increase bb_counters and shrink range to 1400 * [0; 5]. 1401 * Then shift range to [0; 2], go up and do the same. 1402 */ 1403 1404 1405 if (first & 1) 1406 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); 1407 if (!(last & 1)) 1408 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); 1409 if (first > last) 1410 break; 1411 order++; 1412 1413 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { 1414 mb_clear_bits(buddy, first, last - first + 1); 1415 e4b->bd_info->bb_counters[order - 1] += last - first + 1; 1416 break; 1417 } 1418 first >>= 1; 1419 last >>= 1; 1420 buddy = buddy2; 1421 } 1422 } 1423 1424 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1425 int first, int count) 1426 { 1427 int left_is_free = 0; 1428 int right_is_free = 0; 1429 int block; 1430 int last = first + count - 1; 1431 struct super_block *sb = e4b->bd_sb; 1432 1433 if (WARN_ON(count == 0)) 1434 return; 1435 BUG_ON(last >= (sb->s_blocksize << 3)); 1436 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1437 /* Don't bother if the block group is corrupt. */ 1438 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 1439 return; 1440 1441 mb_check_buddy(e4b); 1442 mb_free_blocks_double(inode, e4b, first, count); 1443 1444 e4b->bd_info->bb_free += count; 1445 if (first < e4b->bd_info->bb_first_free) 1446 e4b->bd_info->bb_first_free = first; 1447 1448 /* access memory sequentially: check left neighbour, 1449 * clear range and then check right neighbour 1450 */ 1451 if (first != 0) 1452 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); 1453 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); 1454 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) 1455 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); 1456 1457 if (unlikely(block != -1)) { 1458 struct ext4_sb_info *sbi = EXT4_SB(sb); 1459 ext4_fsblk_t blocknr; 1460 1461 blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 1462 blocknr += EXT4_C2B(EXT4_SB(sb), block); 1463 ext4_grp_locked_error(sb, e4b->bd_group, 1464 inode ? inode->i_ino : 0, 1465 blocknr, 1466 "freeing already freed block " 1467 "(bit %u); block bitmap corrupt.", 1468 block); 1469 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)) 1470 percpu_counter_sub(&sbi->s_freeclusters_counter, 1471 e4b->bd_info->bb_free); 1472 /* Mark the block group as corrupt. */ 1473 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, 1474 &e4b->bd_info->bb_state); 1475 mb_regenerate_buddy(e4b); 1476 goto done; 1477 } 1478 1479 /* let's maintain fragments counter */ 1480 if (left_is_free && right_is_free) 1481 e4b->bd_info->bb_fragments--; 1482 else if (!left_is_free && !right_is_free) 1483 e4b->bd_info->bb_fragments++; 1484 1485 /* buddy[0] == bd_bitmap is a special case, so handle 1486 * it right away and let mb_buddy_mark_free stay free of 1487 * zero order checks. 1488 * Check if neighbours are to be coaleasced, 1489 * adjust bitmap bb_counters and borders appropriately. 1490 */ 1491 if (first & 1) { 1492 first += !left_is_free; 1493 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; 1494 } 1495 if (!(last & 1)) { 1496 last -= !right_is_free; 1497 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; 1498 } 1499 1500 if (first <= last) 1501 mb_buddy_mark_free(e4b, first >> 1, last >> 1); 1502 1503 done: 1504 mb_set_largest_free_order(sb, e4b->bd_info); 1505 mb_check_buddy(e4b); 1506 } 1507 1508 static int mb_find_extent(struct ext4_buddy *e4b, int block, 1509 int needed, struct ext4_free_extent *ex) 1510 { 1511 int next = block; 1512 int max, order; 1513 void *buddy; 1514 1515 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1516 BUG_ON(ex == NULL); 1517 1518 buddy = mb_find_buddy(e4b, 0, &max); 1519 BUG_ON(buddy == NULL); 1520 BUG_ON(block >= max); 1521 if (mb_test_bit(block, buddy)) { 1522 ex->fe_len = 0; 1523 ex->fe_start = 0; 1524 ex->fe_group = 0; 1525 return 0; 1526 } 1527 1528 /* find actual order */ 1529 order = mb_find_order_for_block(e4b, block); 1530 block = block >> order; 1531 1532 ex->fe_len = 1 << order; 1533 ex->fe_start = block << order; 1534 ex->fe_group = e4b->bd_group; 1535 1536 /* calc difference from given start */ 1537 next = next - ex->fe_start; 1538 ex->fe_len -= next; 1539 ex->fe_start += next; 1540 1541 while (needed > ex->fe_len && 1542 mb_find_buddy(e4b, order, &max)) { 1543 1544 if (block + 1 >= max) 1545 break; 1546 1547 next = (block + 1) * (1 << order); 1548 if (mb_test_bit(next, e4b->bd_bitmap)) 1549 break; 1550 1551 order = mb_find_order_for_block(e4b, next); 1552 1553 block = next >> order; 1554 ex->fe_len += 1 << order; 1555 } 1556 1557 if (ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))) { 1558 /* Should never happen! (but apparently sometimes does?!?) */ 1559 WARN_ON(1); 1560 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent " 1561 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", 1562 block, order, needed, ex->fe_group, ex->fe_start, 1563 ex->fe_len, ex->fe_logical); 1564 ex->fe_len = 0; 1565 ex->fe_start = 0; 1566 ex->fe_group = 0; 1567 } 1568 return ex->fe_len; 1569 } 1570 1571 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1572 { 1573 int ord; 1574 int mlen = 0; 1575 int max = 0; 1576 int cur; 1577 int start = ex->fe_start; 1578 int len = ex->fe_len; 1579 unsigned ret = 0; 1580 int len0 = len; 1581 void *buddy; 1582 1583 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1584 BUG_ON(e4b->bd_group != ex->fe_group); 1585 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1586 mb_check_buddy(e4b); 1587 mb_mark_used_double(e4b, start, len); 1588 1589 e4b->bd_info->bb_free -= len; 1590 if (e4b->bd_info->bb_first_free == start) 1591 e4b->bd_info->bb_first_free += len; 1592 1593 /* let's maintain fragments counter */ 1594 if (start != 0) 1595 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 1596 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1597 max = !mb_test_bit(start + len, e4b->bd_bitmap); 1598 if (mlen && max) 1599 e4b->bd_info->bb_fragments++; 1600 else if (!mlen && !max) 1601 e4b->bd_info->bb_fragments--; 1602 1603 /* let's maintain buddy itself */ 1604 while (len) { 1605 ord = mb_find_order_for_block(e4b, start); 1606 1607 if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1608 /* the whole chunk may be allocated at once! */ 1609 mlen = 1 << ord; 1610 buddy = mb_find_buddy(e4b, ord, &max); 1611 BUG_ON((start >> ord) >= max); 1612 mb_set_bit(start >> ord, buddy); 1613 e4b->bd_info->bb_counters[ord]--; 1614 start += mlen; 1615 len -= mlen; 1616 BUG_ON(len < 0); 1617 continue; 1618 } 1619 1620 /* store for history */ 1621 if (ret == 0) 1622 ret = len | (ord << 16); 1623 1624 /* we have to split large buddy */ 1625 BUG_ON(ord <= 0); 1626 buddy = mb_find_buddy(e4b, ord, &max); 1627 mb_set_bit(start >> ord, buddy); 1628 e4b->bd_info->bb_counters[ord]--; 1629 1630 ord--; 1631 cur = (start >> ord) & ~1U; 1632 buddy = mb_find_buddy(e4b, ord, &max); 1633 mb_clear_bit(cur, buddy); 1634 mb_clear_bit(cur + 1, buddy); 1635 e4b->bd_info->bb_counters[ord]++; 1636 e4b->bd_info->bb_counters[ord]++; 1637 } 1638 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 1639 1640 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 1641 mb_check_buddy(e4b); 1642 1643 return ret; 1644 } 1645 1646 /* 1647 * Must be called under group lock! 1648 */ 1649 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1650 struct ext4_buddy *e4b) 1651 { 1652 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1653 int ret; 1654 1655 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1656 BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1657 1658 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1659 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1660 ret = mb_mark_used(e4b, &ac->ac_b_ex); 1661 1662 /* preallocation can change ac_b_ex, thus we store actually 1663 * allocated blocks for history */ 1664 ac->ac_f_ex = ac->ac_b_ex; 1665 1666 ac->ac_status = AC_STATUS_FOUND; 1667 ac->ac_tail = ret & 0xffff; 1668 ac->ac_buddy = ret >> 16; 1669 1670 /* 1671 * take the page reference. We want the page to be pinned 1672 * so that we don't get a ext4_mb_init_cache_call for this 1673 * group until we update the bitmap. That would mean we 1674 * double allocate blocks. The reference is dropped 1675 * in ext4_mb_release_context 1676 */ 1677 ac->ac_bitmap_page = e4b->bd_bitmap_page; 1678 get_page(ac->ac_bitmap_page); 1679 ac->ac_buddy_page = e4b->bd_buddy_page; 1680 get_page(ac->ac_buddy_page); 1681 /* store last allocated for subsequent stream allocation */ 1682 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1683 spin_lock(&sbi->s_md_lock); 1684 sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1685 sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1686 spin_unlock(&sbi->s_md_lock); 1687 } 1688 } 1689 1690 /* 1691 * regular allocator, for general purposes allocation 1692 */ 1693 1694 static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1695 struct ext4_buddy *e4b, 1696 int finish_group) 1697 { 1698 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1699 struct ext4_free_extent *bex = &ac->ac_b_ex; 1700 struct ext4_free_extent *gex = &ac->ac_g_ex; 1701 struct ext4_free_extent ex; 1702 int max; 1703 1704 if (ac->ac_status == AC_STATUS_FOUND) 1705 return; 1706 /* 1707 * We don't want to scan for a whole year 1708 */ 1709 if (ac->ac_found > sbi->s_mb_max_to_scan && 1710 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1711 ac->ac_status = AC_STATUS_BREAK; 1712 return; 1713 } 1714 1715 /* 1716 * Haven't found good chunk so far, let's continue 1717 */ 1718 if (bex->fe_len < gex->fe_len) 1719 return; 1720 1721 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1722 && bex->fe_group == e4b->bd_group) { 1723 /* recheck chunk's availability - we don't know 1724 * when it was found (within this lock-unlock 1725 * period or not) */ 1726 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); 1727 if (max >= gex->fe_len) { 1728 ext4_mb_use_best_found(ac, e4b); 1729 return; 1730 } 1731 } 1732 } 1733 1734 /* 1735 * The routine checks whether found extent is good enough. If it is, 1736 * then the extent gets marked used and flag is set to the context 1737 * to stop scanning. Otherwise, the extent is compared with the 1738 * previous found extent and if new one is better, then it's stored 1739 * in the context. Later, the best found extent will be used, if 1740 * mballoc can't find good enough extent. 1741 * 1742 * FIXME: real allocation policy is to be designed yet! 1743 */ 1744 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1745 struct ext4_free_extent *ex, 1746 struct ext4_buddy *e4b) 1747 { 1748 struct ext4_free_extent *bex = &ac->ac_b_ex; 1749 struct ext4_free_extent *gex = &ac->ac_g_ex; 1750 1751 BUG_ON(ex->fe_len <= 0); 1752 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 1753 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 1754 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1755 1756 ac->ac_found++; 1757 1758 /* 1759 * The special case - take what you catch first 1760 */ 1761 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1762 *bex = *ex; 1763 ext4_mb_use_best_found(ac, e4b); 1764 return; 1765 } 1766 1767 /* 1768 * Let's check whether the chuck is good enough 1769 */ 1770 if (ex->fe_len == gex->fe_len) { 1771 *bex = *ex; 1772 ext4_mb_use_best_found(ac, e4b); 1773 return; 1774 } 1775 1776 /* 1777 * If this is first found extent, just store it in the context 1778 */ 1779 if (bex->fe_len == 0) { 1780 *bex = *ex; 1781 return; 1782 } 1783 1784 /* 1785 * If new found extent is better, store it in the context 1786 */ 1787 if (bex->fe_len < gex->fe_len) { 1788 /* if the request isn't satisfied, any found extent 1789 * larger than previous best one is better */ 1790 if (ex->fe_len > bex->fe_len) 1791 *bex = *ex; 1792 } else if (ex->fe_len > gex->fe_len) { 1793 /* if the request is satisfied, then we try to find 1794 * an extent that still satisfy the request, but is 1795 * smaller than previous one */ 1796 if (ex->fe_len < bex->fe_len) 1797 *bex = *ex; 1798 } 1799 1800 ext4_mb_check_limits(ac, e4b, 0); 1801 } 1802 1803 static noinline_for_stack 1804 int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1805 struct ext4_buddy *e4b) 1806 { 1807 struct ext4_free_extent ex = ac->ac_b_ex; 1808 ext4_group_t group = ex.fe_group; 1809 int max; 1810 int err; 1811 1812 BUG_ON(ex.fe_len <= 0); 1813 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1814 if (err) 1815 return err; 1816 1817 ext4_lock_group(ac->ac_sb, group); 1818 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); 1819 1820 if (max > 0) { 1821 ac->ac_b_ex = ex; 1822 ext4_mb_use_best_found(ac, e4b); 1823 } 1824 1825 ext4_unlock_group(ac->ac_sb, group); 1826 ext4_mb_unload_buddy(e4b); 1827 1828 return 0; 1829 } 1830 1831 static noinline_for_stack 1832 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1833 struct ext4_buddy *e4b) 1834 { 1835 ext4_group_t group = ac->ac_g_ex.fe_group; 1836 int max; 1837 int err; 1838 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1839 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1840 struct ext4_free_extent ex; 1841 1842 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1843 return 0; 1844 if (grp->bb_free == 0) 1845 return 0; 1846 1847 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1848 if (err) 1849 return err; 1850 1851 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { 1852 ext4_mb_unload_buddy(e4b); 1853 return 0; 1854 } 1855 1856 ext4_lock_group(ac->ac_sb, group); 1857 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, 1858 ac->ac_g_ex.fe_len, &ex); 1859 ex.fe_logical = 0xDEADFA11; /* debug value */ 1860 1861 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1862 ext4_fsblk_t start; 1863 1864 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + 1865 ex.fe_start; 1866 /* use do_div to get remainder (would be 64-bit modulo) */ 1867 if (do_div(start, sbi->s_stripe) == 0) { 1868 ac->ac_found++; 1869 ac->ac_b_ex = ex; 1870 ext4_mb_use_best_found(ac, e4b); 1871 } 1872 } else if (max >= ac->ac_g_ex.fe_len) { 1873 BUG_ON(ex.fe_len <= 0); 1874 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1875 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1876 ac->ac_found++; 1877 ac->ac_b_ex = ex; 1878 ext4_mb_use_best_found(ac, e4b); 1879 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1880 /* Sometimes, caller may want to merge even small 1881 * number of blocks to an existing extent */ 1882 BUG_ON(ex.fe_len <= 0); 1883 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1884 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1885 ac->ac_found++; 1886 ac->ac_b_ex = ex; 1887 ext4_mb_use_best_found(ac, e4b); 1888 } 1889 ext4_unlock_group(ac->ac_sb, group); 1890 ext4_mb_unload_buddy(e4b); 1891 1892 return 0; 1893 } 1894 1895 /* 1896 * The routine scans buddy structures (not bitmap!) from given order 1897 * to max order and tries to find big enough chunk to satisfy the req 1898 */ 1899 static noinline_for_stack 1900 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1901 struct ext4_buddy *e4b) 1902 { 1903 struct super_block *sb = ac->ac_sb; 1904 struct ext4_group_info *grp = e4b->bd_info; 1905 void *buddy; 1906 int i; 1907 int k; 1908 int max; 1909 1910 BUG_ON(ac->ac_2order <= 0); 1911 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1912 if (grp->bb_counters[i] == 0) 1913 continue; 1914 1915 buddy = mb_find_buddy(e4b, i, &max); 1916 BUG_ON(buddy == NULL); 1917 1918 k = mb_find_next_zero_bit(buddy, max, 0); 1919 BUG_ON(k >= max); 1920 1921 ac->ac_found++; 1922 1923 ac->ac_b_ex.fe_len = 1 << i; 1924 ac->ac_b_ex.fe_start = k << i; 1925 ac->ac_b_ex.fe_group = e4b->bd_group; 1926 1927 ext4_mb_use_best_found(ac, e4b); 1928 1929 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1930 1931 if (EXT4_SB(sb)->s_mb_stats) 1932 atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1933 1934 break; 1935 } 1936 } 1937 1938 /* 1939 * The routine scans the group and measures all found extents. 1940 * In order to optimize scanning, caller must pass number of 1941 * free blocks in the group, so the routine can know upper limit. 1942 */ 1943 static noinline_for_stack 1944 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1945 struct ext4_buddy *e4b) 1946 { 1947 struct super_block *sb = ac->ac_sb; 1948 void *bitmap = e4b->bd_bitmap; 1949 struct ext4_free_extent ex; 1950 int i; 1951 int free; 1952 1953 free = e4b->bd_info->bb_free; 1954 BUG_ON(free <= 0); 1955 1956 i = e4b->bd_info->bb_first_free; 1957 1958 while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1959 i = mb_find_next_zero_bit(bitmap, 1960 EXT4_CLUSTERS_PER_GROUP(sb), i); 1961 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 1962 /* 1963 * IF we have corrupt bitmap, we won't find any 1964 * free blocks even though group info says we 1965 * we have free blocks 1966 */ 1967 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 1968 "%d free clusters as per " 1969 "group info. But bitmap says 0", 1970 free); 1971 break; 1972 } 1973 1974 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); 1975 BUG_ON(ex.fe_len <= 0); 1976 if (free < ex.fe_len) { 1977 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 1978 "%d free clusters as per " 1979 "group info. But got %d blocks", 1980 free, ex.fe_len); 1981 /* 1982 * The number of free blocks differs. This mostly 1983 * indicate that the bitmap is corrupt. So exit 1984 * without claiming the space. 1985 */ 1986 break; 1987 } 1988 ex.fe_logical = 0xDEADC0DE; /* debug value */ 1989 ext4_mb_measure_extent(ac, &ex, e4b); 1990 1991 i += ex.fe_len; 1992 free -= ex.fe_len; 1993 } 1994 1995 ext4_mb_check_limits(ac, e4b, 1); 1996 } 1997 1998 /* 1999 * This is a special case for storages like raid5 2000 * we try to find stripe-aligned chunks for stripe-size-multiple requests 2001 */ 2002 static noinline_for_stack 2003 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 2004 struct ext4_buddy *e4b) 2005 { 2006 struct super_block *sb = ac->ac_sb; 2007 struct ext4_sb_info *sbi = EXT4_SB(sb); 2008 void *bitmap = e4b->bd_bitmap; 2009 struct ext4_free_extent ex; 2010 ext4_fsblk_t first_group_block; 2011 ext4_fsblk_t a; 2012 ext4_grpblk_t i; 2013 int max; 2014 2015 BUG_ON(sbi->s_stripe == 0); 2016 2017 /* find first stripe-aligned block in group */ 2018 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 2019 2020 a = first_group_block + sbi->s_stripe - 1; 2021 do_div(a, sbi->s_stripe); 2022 i = (a * sbi->s_stripe) - first_group_block; 2023 2024 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 2025 if (!mb_test_bit(i, bitmap)) { 2026 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); 2027 if (max >= sbi->s_stripe) { 2028 ac->ac_found++; 2029 ex.fe_logical = 0xDEADF00D; /* debug value */ 2030 ac->ac_b_ex = ex; 2031 ext4_mb_use_best_found(ac, e4b); 2032 break; 2033 } 2034 } 2035 i += sbi->s_stripe; 2036 } 2037 } 2038 2039 /* 2040 * This is now called BEFORE we load the buddy bitmap. 2041 * Returns either 1 or 0 indicating that the group is either suitable 2042 * for the allocation or not. In addition it can also return negative 2043 * error code when something goes wrong. 2044 */ 2045 static int ext4_mb_good_group(struct ext4_allocation_context *ac, 2046 ext4_group_t group, int cr) 2047 { 2048 unsigned free, fragments; 2049 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 2050 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2051 2052 BUG_ON(cr < 0 || cr >= 4); 2053 2054 free = grp->bb_free; 2055 if (free == 0) 2056 return 0; 2057 if (cr <= 2 && free < ac->ac_g_ex.fe_len) 2058 return 0; 2059 2060 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 2061 return 0; 2062 2063 /* We only do this if the grp has never been initialized */ 2064 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 2065 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS); 2066 if (ret) 2067 return ret; 2068 } 2069 2070 fragments = grp->bb_fragments; 2071 if (fragments == 0) 2072 return 0; 2073 2074 switch (cr) { 2075 case 0: 2076 BUG_ON(ac->ac_2order == 0); 2077 2078 /* Avoid using the first bg of a flexgroup for data files */ 2079 if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 2080 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 2081 ((group % flex_size) == 0)) 2082 return 0; 2083 2084 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || 2085 (free / fragments) >= ac->ac_g_ex.fe_len) 2086 return 1; 2087 2088 if (grp->bb_largest_free_order < ac->ac_2order) 2089 return 0; 2090 2091 return 1; 2092 case 1: 2093 if ((free / fragments) >= ac->ac_g_ex.fe_len) 2094 return 1; 2095 break; 2096 case 2: 2097 if (free >= ac->ac_g_ex.fe_len) 2098 return 1; 2099 break; 2100 case 3: 2101 return 1; 2102 default: 2103 BUG(); 2104 } 2105 2106 return 0; 2107 } 2108 2109 static noinline_for_stack int 2110 ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2111 { 2112 ext4_group_t ngroups, group, i; 2113 int cr; 2114 int err = 0, first_err = 0; 2115 struct ext4_sb_info *sbi; 2116 struct super_block *sb; 2117 struct ext4_buddy e4b; 2118 2119 sb = ac->ac_sb; 2120 sbi = EXT4_SB(sb); 2121 ngroups = ext4_get_groups_count(sb); 2122 /* non-extent files are limited to low blocks/groups */ 2123 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2124 ngroups = sbi->s_blockfile_groups; 2125 2126 BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2127 2128 /* first, try the goal */ 2129 err = ext4_mb_find_by_goal(ac, &e4b); 2130 if (err || ac->ac_status == AC_STATUS_FOUND) 2131 goto out; 2132 2133 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2134 goto out; 2135 2136 /* 2137 * ac->ac2_order is set only if the fe_len is a power of 2 2138 * if ac2_order is set we also set criteria to 0 so that we 2139 * try exact allocation using buddy. 2140 */ 2141 i = fls(ac->ac_g_ex.fe_len); 2142 ac->ac_2order = 0; 2143 /* 2144 * We search using buddy data only if the order of the request 2145 * is greater than equal to the sbi_s_mb_order2_reqs 2146 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2147 * We also support searching for power-of-two requests only for 2148 * requests upto maximum buddy size we have constructed. 2149 */ 2150 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) { 2151 /* 2152 * This should tell if fe_len is exactly power of 2 2153 */ 2154 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 2155 ac->ac_2order = i - 1; 2156 } 2157 2158 /* if stream allocation is enabled, use global goal */ 2159 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2160 /* TBD: may be hot point */ 2161 spin_lock(&sbi->s_md_lock); 2162 ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2163 ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2164 spin_unlock(&sbi->s_md_lock); 2165 } 2166 2167 /* Let's just scan groups to find more-less suitable blocks */ 2168 cr = ac->ac_2order ? 0 : 1; 2169 /* 2170 * cr == 0 try to get exact allocation, 2171 * cr == 3 try to get anything 2172 */ 2173 repeat: 2174 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2175 ac->ac_criteria = cr; 2176 /* 2177 * searching for the right group start 2178 * from the goal value specified 2179 */ 2180 group = ac->ac_g_ex.fe_group; 2181 2182 for (i = 0; i < ngroups; group++, i++) { 2183 int ret = 0; 2184 cond_resched(); 2185 /* 2186 * Artificially restricted ngroups for non-extent 2187 * files makes group > ngroups possible on first loop. 2188 */ 2189 if (group >= ngroups) 2190 group = 0; 2191 2192 /* This now checks without needing the buddy page */ 2193 ret = ext4_mb_good_group(ac, group, cr); 2194 if (ret <= 0) { 2195 if (!first_err) 2196 first_err = ret; 2197 continue; 2198 } 2199 2200 err = ext4_mb_load_buddy(sb, group, &e4b); 2201 if (err) 2202 goto out; 2203 2204 ext4_lock_group(sb, group); 2205 2206 /* 2207 * We need to check again after locking the 2208 * block group 2209 */ 2210 ret = ext4_mb_good_group(ac, group, cr); 2211 if (ret <= 0) { 2212 ext4_unlock_group(sb, group); 2213 ext4_mb_unload_buddy(&e4b); 2214 if (!first_err) 2215 first_err = ret; 2216 continue; 2217 } 2218 2219 ac->ac_groups_scanned++; 2220 if (cr == 0) 2221 ext4_mb_simple_scan_group(ac, &e4b); 2222 else if (cr == 1 && sbi->s_stripe && 2223 !(ac->ac_g_ex.fe_len % sbi->s_stripe)) 2224 ext4_mb_scan_aligned(ac, &e4b); 2225 else 2226 ext4_mb_complex_scan_group(ac, &e4b); 2227 2228 ext4_unlock_group(sb, group); 2229 ext4_mb_unload_buddy(&e4b); 2230 2231 if (ac->ac_status != AC_STATUS_CONTINUE) 2232 break; 2233 } 2234 } 2235 2236 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2237 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2238 /* 2239 * We've been searching too long. Let's try to allocate 2240 * the best chunk we've found so far 2241 */ 2242 2243 ext4_mb_try_best_found(ac, &e4b); 2244 if (ac->ac_status != AC_STATUS_FOUND) { 2245 /* 2246 * Someone more lucky has already allocated it. 2247 * The only thing we can do is just take first 2248 * found block(s) 2249 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2250 */ 2251 ac->ac_b_ex.fe_group = 0; 2252 ac->ac_b_ex.fe_start = 0; 2253 ac->ac_b_ex.fe_len = 0; 2254 ac->ac_status = AC_STATUS_CONTINUE; 2255 ac->ac_flags |= EXT4_MB_HINT_FIRST; 2256 cr = 3; 2257 atomic_inc(&sbi->s_mb_lost_chunks); 2258 goto repeat; 2259 } 2260 } 2261 out: 2262 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) 2263 err = first_err; 2264 return err; 2265 } 2266 2267 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2268 { 2269 struct super_block *sb = seq->private; 2270 ext4_group_t group; 2271 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 void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2279 { 2280 struct super_block *sb = seq->private; 2281 ext4_group_t group; 2282 2283 ++*pos; 2284 if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2285 return NULL; 2286 group = *pos + 1; 2287 return (void *) ((unsigned long) group); 2288 } 2289 2290 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2291 { 2292 struct super_block *sb = seq->private; 2293 ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2294 int i; 2295 int err, buddy_loaded = 0; 2296 struct ext4_buddy e4b; 2297 struct ext4_group_info *grinfo; 2298 unsigned char blocksize_bits = min_t(unsigned char, 2299 sb->s_blocksize_bits, 2300 EXT4_MAX_BLOCK_LOG_SIZE); 2301 struct sg { 2302 struct ext4_group_info info; 2303 ext4_grpblk_t counters[blocksize_bits + 2]; 2304 } sg; 2305 2306 group--; 2307 if (group == 0) 2308 seq_puts(seq, "#group: free frags first [" 2309 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " 2310 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); 2311 2312 grinfo = ext4_get_group_info(sb, group); 2313 /* Load the group info in memory only if not already loaded. */ 2314 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { 2315 err = ext4_mb_load_buddy(sb, group, &e4b); 2316 if (err) { 2317 seq_printf(seq, "#%-5u: I/O error\n", group); 2318 return 0; 2319 } 2320 buddy_loaded = 1; 2321 } 2322 2323 memcpy(&sg, ext4_get_group_info(sb, group), sizeof(sg)); 2324 2325 if (buddy_loaded) 2326 ext4_mb_unload_buddy(&e4b); 2327 2328 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2329 sg.info.bb_fragments, sg.info.bb_first_free); 2330 for (i = 0; i <= 13; i++) 2331 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? 2332 sg.info.bb_counters[i] : 0); 2333 seq_printf(seq, " ]\n"); 2334 2335 return 0; 2336 } 2337 2338 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2339 { 2340 } 2341 2342 static const struct seq_operations ext4_mb_seq_groups_ops = { 2343 .start = ext4_mb_seq_groups_start, 2344 .next = ext4_mb_seq_groups_next, 2345 .stop = ext4_mb_seq_groups_stop, 2346 .show = ext4_mb_seq_groups_show, 2347 }; 2348 2349 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) 2350 { 2351 struct super_block *sb = PDE_DATA(inode); 2352 int rc; 2353 2354 rc = seq_open(file, &ext4_mb_seq_groups_ops); 2355 if (rc == 0) { 2356 struct seq_file *m = file->private_data; 2357 m->private = sb; 2358 } 2359 return rc; 2360 2361 } 2362 2363 const struct file_operations ext4_seq_mb_groups_fops = { 2364 .open = ext4_mb_seq_groups_open, 2365 .read = seq_read, 2366 .llseek = seq_lseek, 2367 .release = seq_release, 2368 }; 2369 2370 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 2371 { 2372 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2373 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 2374 2375 BUG_ON(!cachep); 2376 return cachep; 2377 } 2378 2379 /* 2380 * Allocate the top-level s_group_info array for the specified number 2381 * of groups 2382 */ 2383 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) 2384 { 2385 struct ext4_sb_info *sbi = EXT4_SB(sb); 2386 unsigned size; 2387 struct ext4_group_info ***new_groupinfo; 2388 2389 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> 2390 EXT4_DESC_PER_BLOCK_BITS(sb); 2391 if (size <= sbi->s_group_info_size) 2392 return 0; 2393 2394 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); 2395 new_groupinfo = kvzalloc(size, GFP_KERNEL); 2396 if (!new_groupinfo) { 2397 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 2398 return -ENOMEM; 2399 } 2400 if (sbi->s_group_info) { 2401 memcpy(new_groupinfo, sbi->s_group_info, 2402 sbi->s_group_info_size * sizeof(*sbi->s_group_info)); 2403 kvfree(sbi->s_group_info); 2404 } 2405 sbi->s_group_info = new_groupinfo; 2406 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); 2407 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 2408 sbi->s_group_info_size); 2409 return 0; 2410 } 2411 2412 /* Create and initialize ext4_group_info data for the given group. */ 2413 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 2414 struct ext4_group_desc *desc) 2415 { 2416 int i; 2417 int metalen = 0; 2418 struct ext4_sb_info *sbi = EXT4_SB(sb); 2419 struct ext4_group_info **meta_group_info; 2420 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2421 2422 /* 2423 * First check if this group is the first of a reserved block. 2424 * If it's true, we have to allocate a new table of pointers 2425 * to ext4_group_info structures 2426 */ 2427 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 2428 metalen = sizeof(*meta_group_info) << 2429 EXT4_DESC_PER_BLOCK_BITS(sb); 2430 meta_group_info = kmalloc(metalen, GFP_NOFS); 2431 if (meta_group_info == NULL) { 2432 ext4_msg(sb, KERN_ERR, "can't allocate mem " 2433 "for a buddy group"); 2434 goto exit_meta_group_info; 2435 } 2436 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 2437 meta_group_info; 2438 } 2439 2440 meta_group_info = 2441 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 2442 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 2443 2444 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); 2445 if (meta_group_info[i] == NULL) { 2446 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 2447 goto exit_group_info; 2448 } 2449 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 2450 &(meta_group_info[i]->bb_state)); 2451 2452 /* 2453 * initialize bb_free to be able to skip 2454 * empty groups without initialization 2455 */ 2456 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2457 meta_group_info[i]->bb_free = 2458 ext4_free_clusters_after_init(sb, group, desc); 2459 } else { 2460 meta_group_info[i]->bb_free = 2461 ext4_free_group_clusters(sb, desc); 2462 } 2463 2464 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2465 init_rwsem(&meta_group_info[i]->alloc_sem); 2466 meta_group_info[i]->bb_free_root = RB_ROOT; 2467 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 2468 2469 #ifdef DOUBLE_CHECK 2470 { 2471 struct buffer_head *bh; 2472 meta_group_info[i]->bb_bitmap = 2473 kmalloc(sb->s_blocksize, GFP_NOFS); 2474 BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 2475 bh = ext4_read_block_bitmap(sb, group); 2476 BUG_ON(IS_ERR_OR_NULL(bh)); 2477 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 2478 sb->s_blocksize); 2479 put_bh(bh); 2480 } 2481 #endif 2482 2483 return 0; 2484 2485 exit_group_info: 2486 /* If a meta_group_info table has been allocated, release it now */ 2487 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 2488 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 2489 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; 2490 } 2491 exit_meta_group_info: 2492 return -ENOMEM; 2493 } /* ext4_mb_add_groupinfo */ 2494 2495 static int ext4_mb_init_backend(struct super_block *sb) 2496 { 2497 ext4_group_t ngroups = ext4_get_groups_count(sb); 2498 ext4_group_t i; 2499 struct ext4_sb_info *sbi = EXT4_SB(sb); 2500 int err; 2501 struct ext4_group_desc *desc; 2502 struct kmem_cache *cachep; 2503 2504 err = ext4_mb_alloc_groupinfo(sb, ngroups); 2505 if (err) 2506 return err; 2507 2508 sbi->s_buddy_cache = new_inode(sb); 2509 if (sbi->s_buddy_cache == NULL) { 2510 ext4_msg(sb, KERN_ERR, "can't get new inode"); 2511 goto err_freesgi; 2512 } 2513 /* To avoid potentially colliding with an valid on-disk inode number, 2514 * use EXT4_BAD_INO for the buddy cache inode number. This inode is 2515 * not in the inode hash, so it should never be found by iget(), but 2516 * this will avoid confusion if it ever shows up during debugging. */ 2517 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 2518 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 2519 for (i = 0; i < ngroups; i++) { 2520 desc = ext4_get_group_desc(sb, i, NULL); 2521 if (desc == NULL) { 2522 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 2523 goto err_freebuddy; 2524 } 2525 if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 2526 goto err_freebuddy; 2527 } 2528 2529 return 0; 2530 2531 err_freebuddy: 2532 cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2533 while (i-- > 0) 2534 kmem_cache_free(cachep, ext4_get_group_info(sb, i)); 2535 i = sbi->s_group_info_size; 2536 while (i-- > 0) 2537 kfree(sbi->s_group_info[i]); 2538 iput(sbi->s_buddy_cache); 2539 err_freesgi: 2540 kvfree(sbi->s_group_info); 2541 return -ENOMEM; 2542 } 2543 2544 static void ext4_groupinfo_destroy_slabs(void) 2545 { 2546 int i; 2547 2548 for (i = 0; i < NR_GRPINFO_CACHES; i++) { 2549 if (ext4_groupinfo_caches[i]) 2550 kmem_cache_destroy(ext4_groupinfo_caches[i]); 2551 ext4_groupinfo_caches[i] = NULL; 2552 } 2553 } 2554 2555 static int ext4_groupinfo_create_slab(size_t size) 2556 { 2557 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 2558 int slab_size; 2559 int blocksize_bits = order_base_2(size); 2560 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2561 struct kmem_cache *cachep; 2562 2563 if (cache_index >= NR_GRPINFO_CACHES) 2564 return -EINVAL; 2565 2566 if (unlikely(cache_index < 0)) 2567 cache_index = 0; 2568 2569 mutex_lock(&ext4_grpinfo_slab_create_mutex); 2570 if (ext4_groupinfo_caches[cache_index]) { 2571 mutex_unlock(&ext4_grpinfo_slab_create_mutex); 2572 return 0; /* Already created */ 2573 } 2574 2575 slab_size = offsetof(struct ext4_group_info, 2576 bb_counters[blocksize_bits + 2]); 2577 2578 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 2579 slab_size, 0, SLAB_RECLAIM_ACCOUNT, 2580 NULL); 2581 2582 ext4_groupinfo_caches[cache_index] = cachep; 2583 2584 mutex_unlock(&ext4_grpinfo_slab_create_mutex); 2585 if (!cachep) { 2586 printk(KERN_EMERG 2587 "EXT4-fs: no memory for groupinfo slab cache\n"); 2588 return -ENOMEM; 2589 } 2590 2591 return 0; 2592 } 2593 2594 int ext4_mb_init(struct super_block *sb) 2595 { 2596 struct ext4_sb_info *sbi = EXT4_SB(sb); 2597 unsigned i, j; 2598 unsigned offset, offset_incr; 2599 unsigned max; 2600 int ret; 2601 2602 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); 2603 2604 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2605 if (sbi->s_mb_offsets == NULL) { 2606 ret = -ENOMEM; 2607 goto out; 2608 } 2609 2610 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); 2611 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2612 if (sbi->s_mb_maxs == NULL) { 2613 ret = -ENOMEM; 2614 goto out; 2615 } 2616 2617 ret = ext4_groupinfo_create_slab(sb->s_blocksize); 2618 if (ret < 0) 2619 goto out; 2620 2621 /* order 0 is regular bitmap */ 2622 sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2623 sbi->s_mb_offsets[0] = 0; 2624 2625 i = 1; 2626 offset = 0; 2627 offset_incr = 1 << (sb->s_blocksize_bits - 1); 2628 max = sb->s_blocksize << 2; 2629 do { 2630 sbi->s_mb_offsets[i] = offset; 2631 sbi->s_mb_maxs[i] = max; 2632 offset += offset_incr; 2633 offset_incr = offset_incr >> 1; 2634 max = max >> 1; 2635 i++; 2636 } while (i <= sb->s_blocksize_bits + 1); 2637 2638 spin_lock_init(&sbi->s_md_lock); 2639 spin_lock_init(&sbi->s_bal_lock); 2640 sbi->s_mb_free_pending = 0; 2641 INIT_LIST_HEAD(&sbi->s_freed_data_list); 2642 2643 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2644 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2645 sbi->s_mb_stats = MB_DEFAULT_STATS; 2646 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2647 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 2648 /* 2649 * The default group preallocation is 512, which for 4k block 2650 * sizes translates to 2 megabytes. However for bigalloc file 2651 * systems, this is probably too big (i.e, if the cluster size 2652 * is 1 megabyte, then group preallocation size becomes half a 2653 * gigabyte!). As a default, we will keep a two megabyte 2654 * group pralloc size for cluster sizes up to 64k, and after 2655 * that, we will force a minimum group preallocation size of 2656 * 32 clusters. This translates to 8 megs when the cluster 2657 * size is 256k, and 32 megs when the cluster size is 1 meg, 2658 * which seems reasonable as a default. 2659 */ 2660 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 2661 sbi->s_cluster_bits, 32); 2662 /* 2663 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 2664 * to the lowest multiple of s_stripe which is bigger than 2665 * the s_mb_group_prealloc as determined above. We want 2666 * the preallocation size to be an exact multiple of the 2667 * RAID stripe size so that preallocations don't fragment 2668 * the stripes. 2669 */ 2670 if (sbi->s_stripe > 1) { 2671 sbi->s_mb_group_prealloc = roundup( 2672 sbi->s_mb_group_prealloc, sbi->s_stripe); 2673 } 2674 2675 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2676 if (sbi->s_locality_groups == NULL) { 2677 ret = -ENOMEM; 2678 goto out; 2679 } 2680 for_each_possible_cpu(i) { 2681 struct ext4_locality_group *lg; 2682 lg = per_cpu_ptr(sbi->s_locality_groups, i); 2683 mutex_init(&lg->lg_mutex); 2684 for (j = 0; j < PREALLOC_TB_SIZE; j++) 2685 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2686 spin_lock_init(&lg->lg_prealloc_lock); 2687 } 2688 2689 /* init file for buddy data */ 2690 ret = ext4_mb_init_backend(sb); 2691 if (ret != 0) 2692 goto out_free_locality_groups; 2693 2694 return 0; 2695 2696 out_free_locality_groups: 2697 free_percpu(sbi->s_locality_groups); 2698 sbi->s_locality_groups = NULL; 2699 out: 2700 kfree(sbi->s_mb_offsets); 2701 sbi->s_mb_offsets = NULL; 2702 kfree(sbi->s_mb_maxs); 2703 sbi->s_mb_maxs = NULL; 2704 return ret; 2705 } 2706 2707 /* need to called with the ext4 group lock held */ 2708 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2709 { 2710 struct ext4_prealloc_space *pa; 2711 struct list_head *cur, *tmp; 2712 int count = 0; 2713 2714 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2715 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2716 list_del(&pa->pa_group_list); 2717 count++; 2718 kmem_cache_free(ext4_pspace_cachep, pa); 2719 } 2720 if (count) 2721 mb_debug(1, "mballoc: %u PAs left\n", count); 2722 2723 } 2724 2725 int ext4_mb_release(struct super_block *sb) 2726 { 2727 ext4_group_t ngroups = ext4_get_groups_count(sb); 2728 ext4_group_t i; 2729 int num_meta_group_infos; 2730 struct ext4_group_info *grinfo; 2731 struct ext4_sb_info *sbi = EXT4_SB(sb); 2732 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2733 2734 if (sbi->s_group_info) { 2735 for (i = 0; i < ngroups; i++) { 2736 grinfo = ext4_get_group_info(sb, i); 2737 #ifdef DOUBLE_CHECK 2738 kfree(grinfo->bb_bitmap); 2739 #endif 2740 ext4_lock_group(sb, i); 2741 ext4_mb_cleanup_pa(grinfo); 2742 ext4_unlock_group(sb, i); 2743 kmem_cache_free(cachep, grinfo); 2744 } 2745 num_meta_group_infos = (ngroups + 2746 EXT4_DESC_PER_BLOCK(sb) - 1) >> 2747 EXT4_DESC_PER_BLOCK_BITS(sb); 2748 for (i = 0; i < num_meta_group_infos; i++) 2749 kfree(sbi->s_group_info[i]); 2750 kvfree(sbi->s_group_info); 2751 } 2752 kfree(sbi->s_mb_offsets); 2753 kfree(sbi->s_mb_maxs); 2754 iput(sbi->s_buddy_cache); 2755 if (sbi->s_mb_stats) { 2756 ext4_msg(sb, KERN_INFO, 2757 "mballoc: %u blocks %u reqs (%u success)", 2758 atomic_read(&sbi->s_bal_allocated), 2759 atomic_read(&sbi->s_bal_reqs), 2760 atomic_read(&sbi->s_bal_success)); 2761 ext4_msg(sb, KERN_INFO, 2762 "mballoc: %u extents scanned, %u goal hits, " 2763 "%u 2^N hits, %u breaks, %u lost", 2764 atomic_read(&sbi->s_bal_ex_scanned), 2765 atomic_read(&sbi->s_bal_goals), 2766 atomic_read(&sbi->s_bal_2orders), 2767 atomic_read(&sbi->s_bal_breaks), 2768 atomic_read(&sbi->s_mb_lost_chunks)); 2769 ext4_msg(sb, KERN_INFO, 2770 "mballoc: %lu generated and it took %Lu", 2771 sbi->s_mb_buddies_generated, 2772 sbi->s_mb_generation_time); 2773 ext4_msg(sb, KERN_INFO, 2774 "mballoc: %u preallocated, %u discarded", 2775 atomic_read(&sbi->s_mb_preallocated), 2776 atomic_read(&sbi->s_mb_discarded)); 2777 } 2778 2779 free_percpu(sbi->s_locality_groups); 2780 2781 return 0; 2782 } 2783 2784 static inline int ext4_issue_discard(struct super_block *sb, 2785 ext4_group_t block_group, ext4_grpblk_t cluster, int count, 2786 struct bio **biop) 2787 { 2788 ext4_fsblk_t discard_block; 2789 2790 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 2791 ext4_group_first_block_no(sb, block_group)); 2792 count = EXT4_C2B(EXT4_SB(sb), count); 2793 trace_ext4_discard_blocks(sb, 2794 (unsigned long long) discard_block, count); 2795 if (biop) { 2796 return __blkdev_issue_discard(sb->s_bdev, 2797 (sector_t)discard_block << (sb->s_blocksize_bits - 9), 2798 (sector_t)count << (sb->s_blocksize_bits - 9), 2799 GFP_NOFS, 0, biop); 2800 } else 2801 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 2802 } 2803 2804 static void ext4_free_data_in_buddy(struct super_block *sb, 2805 struct ext4_free_data *entry) 2806 { 2807 struct ext4_buddy e4b; 2808 struct ext4_group_info *db; 2809 int err, count = 0, count2 = 0; 2810 2811 mb_debug(1, "gonna free %u blocks in group %u (0x%p):", 2812 entry->efd_count, entry->efd_group, entry); 2813 2814 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 2815 /* we expect to find existing buddy because it's pinned */ 2816 BUG_ON(err != 0); 2817 2818 spin_lock(&EXT4_SB(sb)->s_md_lock); 2819 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; 2820 spin_unlock(&EXT4_SB(sb)->s_md_lock); 2821 2822 db = e4b.bd_info; 2823 /* there are blocks to put in buddy to make them really free */ 2824 count += entry->efd_count; 2825 count2++; 2826 ext4_lock_group(sb, entry->efd_group); 2827 /* Take it out of per group rb tree */ 2828 rb_erase(&entry->efd_node, &(db->bb_free_root)); 2829 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 2830 2831 /* 2832 * Clear the trimmed flag for the group so that the next 2833 * ext4_trim_fs can trim it. 2834 * If the volume is mounted with -o discard, online discard 2835 * is supported and the free blocks will be trimmed online. 2836 */ 2837 if (!test_opt(sb, DISCARD)) 2838 EXT4_MB_GRP_CLEAR_TRIMMED(db); 2839 2840 if (!db->bb_free_root.rb_node) { 2841 /* No more items in the per group rb tree 2842 * balance refcounts from ext4_mb_free_metadata() 2843 */ 2844 put_page(e4b.bd_buddy_page); 2845 put_page(e4b.bd_bitmap_page); 2846 } 2847 ext4_unlock_group(sb, entry->efd_group); 2848 kmem_cache_free(ext4_free_data_cachep, entry); 2849 ext4_mb_unload_buddy(&e4b); 2850 2851 mb_debug(1, "freed %u blocks in %u structures\n", count, count2); 2852 } 2853 2854 /* 2855 * This function is called by the jbd2 layer once the commit has finished, 2856 * so we know we can free the blocks that were released with that commit. 2857 */ 2858 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) 2859 { 2860 struct ext4_sb_info *sbi = EXT4_SB(sb); 2861 struct ext4_free_data *entry, *tmp; 2862 struct bio *discard_bio = NULL; 2863 struct list_head freed_data_list; 2864 struct list_head *cut_pos = NULL; 2865 int err; 2866 2867 INIT_LIST_HEAD(&freed_data_list); 2868 2869 spin_lock(&sbi->s_md_lock); 2870 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) { 2871 if (entry->efd_tid != commit_tid) 2872 break; 2873 cut_pos = &entry->efd_list; 2874 } 2875 if (cut_pos) 2876 list_cut_position(&freed_data_list, &sbi->s_freed_data_list, 2877 cut_pos); 2878 spin_unlock(&sbi->s_md_lock); 2879 2880 if (test_opt(sb, DISCARD)) { 2881 list_for_each_entry(entry, &freed_data_list, efd_list) { 2882 err = ext4_issue_discard(sb, entry->efd_group, 2883 entry->efd_start_cluster, 2884 entry->efd_count, 2885 &discard_bio); 2886 if (err && err != -EOPNOTSUPP) { 2887 ext4_msg(sb, KERN_WARNING, "discard request in" 2888 " group:%d block:%d count:%d failed" 2889 " with %d", entry->efd_group, 2890 entry->efd_start_cluster, 2891 entry->efd_count, err); 2892 } else if (err == -EOPNOTSUPP) 2893 break; 2894 } 2895 2896 if (discard_bio) { 2897 submit_bio_wait(discard_bio); 2898 bio_put(discard_bio); 2899 } 2900 } 2901 2902 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) 2903 ext4_free_data_in_buddy(sb, entry); 2904 } 2905 2906 int __init ext4_init_mballoc(void) 2907 { 2908 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 2909 SLAB_RECLAIM_ACCOUNT); 2910 if (ext4_pspace_cachep == NULL) 2911 return -ENOMEM; 2912 2913 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 2914 SLAB_RECLAIM_ACCOUNT); 2915 if (ext4_ac_cachep == NULL) { 2916 kmem_cache_destroy(ext4_pspace_cachep); 2917 return -ENOMEM; 2918 } 2919 2920 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 2921 SLAB_RECLAIM_ACCOUNT); 2922 if (ext4_free_data_cachep == NULL) { 2923 kmem_cache_destroy(ext4_pspace_cachep); 2924 kmem_cache_destroy(ext4_ac_cachep); 2925 return -ENOMEM; 2926 } 2927 return 0; 2928 } 2929 2930 void ext4_exit_mballoc(void) 2931 { 2932 /* 2933 * Wait for completion of call_rcu()'s on ext4_pspace_cachep 2934 * before destroying the slab cache. 2935 */ 2936 rcu_barrier(); 2937 kmem_cache_destroy(ext4_pspace_cachep); 2938 kmem_cache_destroy(ext4_ac_cachep); 2939 kmem_cache_destroy(ext4_free_data_cachep); 2940 ext4_groupinfo_destroy_slabs(); 2941 } 2942 2943 2944 /* 2945 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 2946 * Returns 0 if success or error code 2947 */ 2948 static noinline_for_stack int 2949 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 2950 handle_t *handle, unsigned int reserv_clstrs) 2951 { 2952 struct buffer_head *bitmap_bh = NULL; 2953 struct ext4_group_desc *gdp; 2954 struct buffer_head *gdp_bh; 2955 struct ext4_sb_info *sbi; 2956 struct super_block *sb; 2957 ext4_fsblk_t block; 2958 int err, len; 2959 2960 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2961 BUG_ON(ac->ac_b_ex.fe_len <= 0); 2962 2963 sb = ac->ac_sb; 2964 sbi = EXT4_SB(sb); 2965 2966 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 2967 if (IS_ERR(bitmap_bh)) { 2968 err = PTR_ERR(bitmap_bh); 2969 bitmap_bh = NULL; 2970 goto out_err; 2971 } 2972 2973 BUFFER_TRACE(bitmap_bh, "getting write access"); 2974 err = ext4_journal_get_write_access(handle, bitmap_bh); 2975 if (err) 2976 goto out_err; 2977 2978 err = -EIO; 2979 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2980 if (!gdp) 2981 goto out_err; 2982 2983 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 2984 ext4_free_group_clusters(sb, gdp)); 2985 2986 BUFFER_TRACE(gdp_bh, "get_write_access"); 2987 err = ext4_journal_get_write_access(handle, gdp_bh); 2988 if (err) 2989 goto out_err; 2990 2991 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 2992 2993 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 2994 if (!ext4_data_block_valid(sbi, block, len)) { 2995 ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 2996 "fs metadata", block, block+len); 2997 /* File system mounted not to panic on error 2998 * Fix the bitmap and return EFSCORRUPTED 2999 * We leak some of the blocks here. 3000 */ 3001 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3002 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 3003 ac->ac_b_ex.fe_len); 3004 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3005 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 3006 if (!err) 3007 err = -EFSCORRUPTED; 3008 goto out_err; 3009 } 3010 3011 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3012 #ifdef AGGRESSIVE_CHECK 3013 { 3014 int i; 3015 for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 3016 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 3017 bitmap_bh->b_data)); 3018 } 3019 } 3020 #endif 3021 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 3022 ac->ac_b_ex.fe_len); 3023 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 3024 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 3025 ext4_free_group_clusters_set(sb, gdp, 3026 ext4_free_clusters_after_init(sb, 3027 ac->ac_b_ex.fe_group, gdp)); 3028 } 3029 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; 3030 ext4_free_group_clusters_set(sb, gdp, len); 3031 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); 3032 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); 3033 3034 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3035 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 3036 /* 3037 * Now reduce the dirty block count also. Should not go negative 3038 */ 3039 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 3040 /* release all the reserved blocks if non delalloc */ 3041 percpu_counter_sub(&sbi->s_dirtyclusters_counter, 3042 reserv_clstrs); 3043 3044 if (sbi->s_log_groups_per_flex) { 3045 ext4_group_t flex_group = ext4_flex_group(sbi, 3046 ac->ac_b_ex.fe_group); 3047 atomic64_sub(ac->ac_b_ex.fe_len, 3048 &sbi->s_flex_groups[flex_group].free_clusters); 3049 } 3050 3051 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 3052 if (err) 3053 goto out_err; 3054 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 3055 3056 out_err: 3057 brelse(bitmap_bh); 3058 return err; 3059 } 3060 3061 /* 3062 * here we normalize request for locality group 3063 * Group request are normalized to s_mb_group_prealloc, which goes to 3064 * s_strip if we set the same via mount option. 3065 * s_mb_group_prealloc can be configured via 3066 * /sys/fs/ext4/<partition>/mb_group_prealloc 3067 * 3068 * XXX: should we try to preallocate more than the group has now? 3069 */ 3070 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 3071 { 3072 struct super_block *sb = ac->ac_sb; 3073 struct ext4_locality_group *lg = ac->ac_lg; 3074 3075 BUG_ON(lg == NULL); 3076 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 3077 mb_debug(1, "#%u: goal %u blocks for locality group\n", 3078 current->pid, ac->ac_g_ex.fe_len); 3079 } 3080 3081 /* 3082 * Normalization means making request better in terms of 3083 * size and alignment 3084 */ 3085 static noinline_for_stack void 3086 ext4_mb_normalize_request(struct ext4_allocation_context *ac, 3087 struct ext4_allocation_request *ar) 3088 { 3089 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3090 int bsbits, max; 3091 ext4_lblk_t end; 3092 loff_t size, start_off; 3093 loff_t orig_size __maybe_unused; 3094 ext4_lblk_t start; 3095 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3096 struct ext4_prealloc_space *pa; 3097 3098 /* do normalize only data requests, metadata requests 3099 do not need preallocation */ 3100 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3101 return; 3102 3103 /* sometime caller may want exact blocks */ 3104 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 3105 return; 3106 3107 /* caller may indicate that preallocation isn't 3108 * required (it's a tail, for example) */ 3109 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 3110 return; 3111 3112 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 3113 ext4_mb_normalize_group_request(ac); 3114 return ; 3115 } 3116 3117 bsbits = ac->ac_sb->s_blocksize_bits; 3118 3119 /* first, let's learn actual file size 3120 * given current request is allocated */ 3121 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 3122 size = size << bsbits; 3123 if (size < i_size_read(ac->ac_inode)) 3124 size = i_size_read(ac->ac_inode); 3125 orig_size = size; 3126 3127 /* max size of free chunks */ 3128 max = 2 << bsbits; 3129 3130 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 3131 (req <= (size) || max <= (chunk_size)) 3132 3133 /* first, try to predict filesize */ 3134 /* XXX: should this table be tunable? */ 3135 start_off = 0; 3136 if (size <= 16 * 1024) { 3137 size = 16 * 1024; 3138 } else if (size <= 32 * 1024) { 3139 size = 32 * 1024; 3140 } else if (size <= 64 * 1024) { 3141 size = 64 * 1024; 3142 } else if (size <= 128 * 1024) { 3143 size = 128 * 1024; 3144 } else if (size <= 256 * 1024) { 3145 size = 256 * 1024; 3146 } else if (size <= 512 * 1024) { 3147 size = 512 * 1024; 3148 } else if (size <= 1024 * 1024) { 3149 size = 1024 * 1024; 3150 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 3151 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3152 (21 - bsbits)) << 21; 3153 size = 2 * 1024 * 1024; 3154 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 3155 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3156 (22 - bsbits)) << 22; 3157 size = 4 * 1024 * 1024; 3158 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 3159 (8<<20)>>bsbits, max, 8 * 1024)) { 3160 start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3161 (23 - bsbits)) << 23; 3162 size = 8 * 1024 * 1024; 3163 } else { 3164 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; 3165 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), 3166 ac->ac_o_ex.fe_len) << bsbits; 3167 } 3168 size = size >> bsbits; 3169 start = start_off >> bsbits; 3170 3171 /* don't cover already allocated blocks in selected range */ 3172 if (ar->pleft && start <= ar->lleft) { 3173 size -= ar->lleft + 1 - start; 3174 start = ar->lleft + 1; 3175 } 3176 if (ar->pright && start + size - 1 >= ar->lright) 3177 size -= start + size - ar->lright; 3178 3179 /* 3180 * Trim allocation request for filesystems with artificially small 3181 * groups. 3182 */ 3183 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) 3184 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); 3185 3186 end = start + size; 3187 3188 /* check we don't cross already preallocated blocks */ 3189 rcu_read_lock(); 3190 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3191 ext4_lblk_t pa_end; 3192 3193 if (pa->pa_deleted) 3194 continue; 3195 spin_lock(&pa->pa_lock); 3196 if (pa->pa_deleted) { 3197 spin_unlock(&pa->pa_lock); 3198 continue; 3199 } 3200 3201 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 3202 pa->pa_len); 3203 3204 /* PA must not overlap original request */ 3205 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 3206 ac->ac_o_ex.fe_logical < pa->pa_lstart)); 3207 3208 /* skip PAs this normalized request doesn't overlap with */ 3209 if (pa->pa_lstart >= end || pa_end <= start) { 3210 spin_unlock(&pa->pa_lock); 3211 continue; 3212 } 3213 BUG_ON(pa->pa_lstart <= start && pa_end >= end); 3214 3215 /* adjust start or end to be adjacent to this pa */ 3216 if (pa_end <= ac->ac_o_ex.fe_logical) { 3217 BUG_ON(pa_end < start); 3218 start = pa_end; 3219 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 3220 BUG_ON(pa->pa_lstart > end); 3221 end = pa->pa_lstart; 3222 } 3223 spin_unlock(&pa->pa_lock); 3224 } 3225 rcu_read_unlock(); 3226 size = end - start; 3227 3228 /* XXX: extra loop to check we really don't overlap preallocations */ 3229 rcu_read_lock(); 3230 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3231 ext4_lblk_t pa_end; 3232 3233 spin_lock(&pa->pa_lock); 3234 if (pa->pa_deleted == 0) { 3235 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 3236 pa->pa_len); 3237 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 3238 } 3239 spin_unlock(&pa->pa_lock); 3240 } 3241 rcu_read_unlock(); 3242 3243 if (start + size <= ac->ac_o_ex.fe_logical && 3244 start > ac->ac_o_ex.fe_logical) { 3245 ext4_msg(ac->ac_sb, KERN_ERR, 3246 "start %lu, size %lu, fe_logical %lu", 3247 (unsigned long) start, (unsigned long) size, 3248 (unsigned long) ac->ac_o_ex.fe_logical); 3249 BUG(); 3250 } 3251 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 3252 3253 /* now prepare goal request */ 3254 3255 /* XXX: is it better to align blocks WRT to logical 3256 * placement or satisfy big request as is */ 3257 ac->ac_g_ex.fe_logical = start; 3258 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 3259 3260 /* define goal start in order to merge */ 3261 if (ar->pright && (ar->lright == (start + size))) { 3262 /* merge to the right */ 3263 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 3264 &ac->ac_f_ex.fe_group, 3265 &ac->ac_f_ex.fe_start); 3266 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3267 } 3268 if (ar->pleft && (ar->lleft + 1 == start)) { 3269 /* merge to the left */ 3270 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 3271 &ac->ac_f_ex.fe_group, 3272 &ac->ac_f_ex.fe_start); 3273 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3274 } 3275 3276 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, 3277 (unsigned) orig_size, (unsigned) start); 3278 } 3279 3280 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 3281 { 3282 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3283 3284 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 3285 atomic_inc(&sbi->s_bal_reqs); 3286 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 3287 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 3288 atomic_inc(&sbi->s_bal_success); 3289 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 3290 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 3291 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 3292 atomic_inc(&sbi->s_bal_goals); 3293 if (ac->ac_found > sbi->s_mb_max_to_scan) 3294 atomic_inc(&sbi->s_bal_breaks); 3295 } 3296 3297 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 3298 trace_ext4_mballoc_alloc(ac); 3299 else 3300 trace_ext4_mballoc_prealloc(ac); 3301 } 3302 3303 /* 3304 * Called on failure; free up any blocks from the inode PA for this 3305 * context. We don't need this for MB_GROUP_PA because we only change 3306 * pa_free in ext4_mb_release_context(), but on failure, we've already 3307 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 3308 */ 3309 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 3310 { 3311 struct ext4_prealloc_space *pa = ac->ac_pa; 3312 struct ext4_buddy e4b; 3313 int err; 3314 3315 if (pa == NULL) { 3316 if (ac->ac_f_ex.fe_len == 0) 3317 return; 3318 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); 3319 if (err) { 3320 /* 3321 * This should never happen since we pin the 3322 * pages in the ext4_allocation_context so 3323 * ext4_mb_load_buddy() should never fail. 3324 */ 3325 WARN(1, "mb_load_buddy failed (%d)", err); 3326 return; 3327 } 3328 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 3329 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, 3330 ac->ac_f_ex.fe_len); 3331 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 3332 ext4_mb_unload_buddy(&e4b); 3333 return; 3334 } 3335 if (pa->pa_type == MB_INODE_PA) 3336 pa->pa_free += ac->ac_b_ex.fe_len; 3337 } 3338 3339 /* 3340 * use blocks preallocated to inode 3341 */ 3342 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3343 struct ext4_prealloc_space *pa) 3344 { 3345 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3346 ext4_fsblk_t start; 3347 ext4_fsblk_t end; 3348 int len; 3349 3350 /* found preallocated blocks, use them */ 3351 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 3352 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 3353 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 3354 len = EXT4_NUM_B2C(sbi, end - start); 3355 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3356 &ac->ac_b_ex.fe_start); 3357 ac->ac_b_ex.fe_len = len; 3358 ac->ac_status = AC_STATUS_FOUND; 3359 ac->ac_pa = pa; 3360 3361 BUG_ON(start < pa->pa_pstart); 3362 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 3363 BUG_ON(pa->pa_free < len); 3364 pa->pa_free -= len; 3365 3366 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); 3367 } 3368 3369 /* 3370 * use blocks preallocated to locality group 3371 */ 3372 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3373 struct ext4_prealloc_space *pa) 3374 { 3375 unsigned int len = ac->ac_o_ex.fe_len; 3376 3377 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3378 &ac->ac_b_ex.fe_group, 3379 &ac->ac_b_ex.fe_start); 3380 ac->ac_b_ex.fe_len = len; 3381 ac->ac_status = AC_STATUS_FOUND; 3382 ac->ac_pa = pa; 3383 3384 /* we don't correct pa_pstart or pa_plen here to avoid 3385 * possible race when the group is being loaded concurrently 3386 * instead we correct pa later, after blocks are marked 3387 * in on-disk bitmap -- see ext4_mb_release_context() 3388 * Other CPUs are prevented from allocating from this pa by lg_mutex 3389 */ 3390 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3391 } 3392 3393 /* 3394 * Return the prealloc space that have minimal distance 3395 * from the goal block. @cpa is the prealloc 3396 * space that is having currently known minimal distance 3397 * from the goal block. 3398 */ 3399 static struct ext4_prealloc_space * 3400 ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 3401 struct ext4_prealloc_space *pa, 3402 struct ext4_prealloc_space *cpa) 3403 { 3404 ext4_fsblk_t cur_distance, new_distance; 3405 3406 if (cpa == NULL) { 3407 atomic_inc(&pa->pa_count); 3408 return pa; 3409 } 3410 cur_distance = abs(goal_block - cpa->pa_pstart); 3411 new_distance = abs(goal_block - pa->pa_pstart); 3412 3413 if (cur_distance <= new_distance) 3414 return cpa; 3415 3416 /* drop the previous reference */ 3417 atomic_dec(&cpa->pa_count); 3418 atomic_inc(&pa->pa_count); 3419 return pa; 3420 } 3421 3422 /* 3423 * search goal blocks in preallocated space 3424 */ 3425 static noinline_for_stack int 3426 ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3427 { 3428 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3429 int order, i; 3430 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3431 struct ext4_locality_group *lg; 3432 struct ext4_prealloc_space *pa, *cpa = NULL; 3433 ext4_fsblk_t goal_block; 3434 3435 /* only data can be preallocated */ 3436 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3437 return 0; 3438 3439 /* first, try per-file preallocation */ 3440 rcu_read_lock(); 3441 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3442 3443 /* all fields in this condition don't change, 3444 * so we can skip locking for them */ 3445 if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 3446 ac->ac_o_ex.fe_logical >= (pa->pa_lstart + 3447 EXT4_C2B(sbi, pa->pa_len))) 3448 continue; 3449 3450 /* non-extent files can't have physical blocks past 2^32 */ 3451 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 3452 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > 3453 EXT4_MAX_BLOCK_FILE_PHYS)) 3454 continue; 3455 3456 /* found preallocated blocks, use them */ 3457 spin_lock(&pa->pa_lock); 3458 if (pa->pa_deleted == 0 && pa->pa_free) { 3459 atomic_inc(&pa->pa_count); 3460 ext4_mb_use_inode_pa(ac, pa); 3461 spin_unlock(&pa->pa_lock); 3462 ac->ac_criteria = 10; 3463 rcu_read_unlock(); 3464 return 1; 3465 } 3466 spin_unlock(&pa->pa_lock); 3467 } 3468 rcu_read_unlock(); 3469 3470 /* can we use group allocation? */ 3471 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3472 return 0; 3473 3474 /* inode may have no locality group for some reason */ 3475 lg = ac->ac_lg; 3476 if (lg == NULL) 3477 return 0; 3478 order = fls(ac->ac_o_ex.fe_len) - 1; 3479 if (order > PREALLOC_TB_SIZE - 1) 3480 /* The max size of hash table is PREALLOC_TB_SIZE */ 3481 order = PREALLOC_TB_SIZE - 1; 3482 3483 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 3484 /* 3485 * search for the prealloc space that is having 3486 * minimal distance from the goal block. 3487 */ 3488 for (i = order; i < PREALLOC_TB_SIZE; i++) { 3489 rcu_read_lock(); 3490 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 3491 pa_inode_list) { 3492 spin_lock(&pa->pa_lock); 3493 if (pa->pa_deleted == 0 && 3494 pa->pa_free >= ac->ac_o_ex.fe_len) { 3495 3496 cpa = ext4_mb_check_group_pa(goal_block, 3497 pa, cpa); 3498 } 3499 spin_unlock(&pa->pa_lock); 3500 } 3501 rcu_read_unlock(); 3502 } 3503 if (cpa) { 3504 ext4_mb_use_group_pa(ac, cpa); 3505 ac->ac_criteria = 20; 3506 return 1; 3507 } 3508 return 0; 3509 } 3510 3511 /* 3512 * the function goes through all block freed in the group 3513 * but not yet committed and marks them used in in-core bitmap. 3514 * buddy must be generated from this bitmap 3515 * Need to be called with the ext4 group lock held 3516 */ 3517 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 3518 ext4_group_t group) 3519 { 3520 struct rb_node *n; 3521 struct ext4_group_info *grp; 3522 struct ext4_free_data *entry; 3523 3524 grp = ext4_get_group_info(sb, group); 3525 n = rb_first(&(grp->bb_free_root)); 3526 3527 while (n) { 3528 entry = rb_entry(n, struct ext4_free_data, efd_node); 3529 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); 3530 n = rb_next(n); 3531 } 3532 return; 3533 } 3534 3535 /* 3536 * the function goes through all preallocation in this group and marks them 3537 * used in in-core bitmap. buddy must be generated from this bitmap 3538 * Need to be called with ext4 group lock held 3539 */ 3540 static noinline_for_stack 3541 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3542 ext4_group_t group) 3543 { 3544 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3545 struct ext4_prealloc_space *pa; 3546 struct list_head *cur; 3547 ext4_group_t groupnr; 3548 ext4_grpblk_t start; 3549 int preallocated = 0; 3550 int len; 3551 3552 /* all form of preallocation discards first load group, 3553 * so the only competing code is preallocation use. 3554 * we don't need any locking here 3555 * notice we do NOT ignore preallocations with pa_deleted 3556 * otherwise we could leave used blocks available for 3557 * allocation in buddy when concurrent ext4_mb_put_pa() 3558 * is dropping preallocation 3559 */ 3560 list_for_each(cur, &grp->bb_prealloc_list) { 3561 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3562 spin_lock(&pa->pa_lock); 3563 ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3564 &groupnr, &start); 3565 len = pa->pa_len; 3566 spin_unlock(&pa->pa_lock); 3567 if (unlikely(len == 0)) 3568 continue; 3569 BUG_ON(groupnr != group); 3570 ext4_set_bits(bitmap, start, len); 3571 preallocated += len; 3572 } 3573 mb_debug(1, "preallocated %u for group %u\n", preallocated, group); 3574 } 3575 3576 static void ext4_mb_pa_callback(struct rcu_head *head) 3577 { 3578 struct ext4_prealloc_space *pa; 3579 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 3580 3581 BUG_ON(atomic_read(&pa->pa_count)); 3582 BUG_ON(pa->pa_deleted == 0); 3583 kmem_cache_free(ext4_pspace_cachep, pa); 3584 } 3585 3586 /* 3587 * drops a reference to preallocated space descriptor 3588 * if this was the last reference and the space is consumed 3589 */ 3590 static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3591 struct super_block *sb, struct ext4_prealloc_space *pa) 3592 { 3593 ext4_group_t grp; 3594 ext4_fsblk_t grp_blk; 3595 3596 /* in this short window concurrent discard can set pa_deleted */ 3597 spin_lock(&pa->pa_lock); 3598 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { 3599 spin_unlock(&pa->pa_lock); 3600 return; 3601 } 3602 3603 if (pa->pa_deleted == 1) { 3604 spin_unlock(&pa->pa_lock); 3605 return; 3606 } 3607 3608 pa->pa_deleted = 1; 3609 spin_unlock(&pa->pa_lock); 3610 3611 grp_blk = pa->pa_pstart; 3612 /* 3613 * If doing group-based preallocation, pa_pstart may be in the 3614 * next group when pa is used up 3615 */ 3616 if (pa->pa_type == MB_GROUP_PA) 3617 grp_blk--; 3618 3619 grp = ext4_get_group_number(sb, grp_blk); 3620 3621 /* 3622 * possible race: 3623 * 3624 * P1 (buddy init) P2 (regular allocation) 3625 * find block B in PA 3626 * copy on-disk bitmap to buddy 3627 * mark B in on-disk bitmap 3628 * drop PA from group 3629 * mark all PAs in buddy 3630 * 3631 * thus, P1 initializes buddy with B available. to prevent this 3632 * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3633 * against that pair 3634 */ 3635 ext4_lock_group(sb, grp); 3636 list_del(&pa->pa_group_list); 3637 ext4_unlock_group(sb, grp); 3638 3639 spin_lock(pa->pa_obj_lock); 3640 list_del_rcu(&pa->pa_inode_list); 3641 spin_unlock(pa->pa_obj_lock); 3642 3643 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3644 } 3645 3646 /* 3647 * creates new preallocated space for given inode 3648 */ 3649 static noinline_for_stack int 3650 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3651 { 3652 struct super_block *sb = ac->ac_sb; 3653 struct ext4_sb_info *sbi = EXT4_SB(sb); 3654 struct ext4_prealloc_space *pa; 3655 struct ext4_group_info *grp; 3656 struct ext4_inode_info *ei; 3657 3658 /* preallocate only when found space is larger then requested */ 3659 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3660 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3661 BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3662 3663 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3664 if (pa == NULL) 3665 return -ENOMEM; 3666 3667 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3668 int winl; 3669 int wins; 3670 int win; 3671 int offs; 3672 3673 /* we can't allocate as much as normalizer wants. 3674 * so, found space must get proper lstart 3675 * to cover original request */ 3676 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3677 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3678 3679 /* we're limited by original request in that 3680 * logical block must be covered any way 3681 * winl is window we can move our chunk within */ 3682 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3683 3684 /* also, we should cover whole original request */ 3685 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); 3686 3687 /* the smallest one defines real window */ 3688 win = min(winl, wins); 3689 3690 offs = ac->ac_o_ex.fe_logical % 3691 EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 3692 if (offs && offs < win) 3693 win = offs; 3694 3695 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - 3696 EXT4_NUM_B2C(sbi, win); 3697 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3698 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3699 } 3700 3701 /* preallocation can change ac_b_ex, thus we store actually 3702 * allocated blocks for history */ 3703 ac->ac_f_ex = ac->ac_b_ex; 3704 3705 pa->pa_lstart = ac->ac_b_ex.fe_logical; 3706 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3707 pa->pa_len = ac->ac_b_ex.fe_len; 3708 pa->pa_free = pa->pa_len; 3709 atomic_set(&pa->pa_count, 1); 3710 spin_lock_init(&pa->pa_lock); 3711 INIT_LIST_HEAD(&pa->pa_inode_list); 3712 INIT_LIST_HEAD(&pa->pa_group_list); 3713 pa->pa_deleted = 0; 3714 pa->pa_type = MB_INODE_PA; 3715 3716 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, 3717 pa->pa_pstart, pa->pa_len, pa->pa_lstart); 3718 trace_ext4_mb_new_inode_pa(ac, pa); 3719 3720 ext4_mb_use_inode_pa(ac, pa); 3721 atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 3722 3723 ei = EXT4_I(ac->ac_inode); 3724 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3725 3726 pa->pa_obj_lock = &ei->i_prealloc_lock; 3727 pa->pa_inode = ac->ac_inode; 3728 3729 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3730 list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3731 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3732 3733 spin_lock(pa->pa_obj_lock); 3734 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3735 spin_unlock(pa->pa_obj_lock); 3736 3737 return 0; 3738 } 3739 3740 /* 3741 * creates new preallocated space for locality group inodes belongs to 3742 */ 3743 static noinline_for_stack int 3744 ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3745 { 3746 struct super_block *sb = ac->ac_sb; 3747 struct ext4_locality_group *lg; 3748 struct ext4_prealloc_space *pa; 3749 struct ext4_group_info *grp; 3750 3751 /* preallocate only when found space is larger then requested */ 3752 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3753 BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3754 BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3755 3756 BUG_ON(ext4_pspace_cachep == NULL); 3757 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3758 if (pa == NULL) 3759 return -ENOMEM; 3760 3761 /* preallocation can change ac_b_ex, thus we store actually 3762 * allocated blocks for history */ 3763 ac->ac_f_ex = ac->ac_b_ex; 3764 3765 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3766 pa->pa_lstart = pa->pa_pstart; 3767 pa->pa_len = ac->ac_b_ex.fe_len; 3768 pa->pa_free = pa->pa_len; 3769 atomic_set(&pa->pa_count, 1); 3770 spin_lock_init(&pa->pa_lock); 3771 INIT_LIST_HEAD(&pa->pa_inode_list); 3772 INIT_LIST_HEAD(&pa->pa_group_list); 3773 pa->pa_deleted = 0; 3774 pa->pa_type = MB_GROUP_PA; 3775 3776 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, 3777 pa->pa_pstart, pa->pa_len, pa->pa_lstart); 3778 trace_ext4_mb_new_group_pa(ac, pa); 3779 3780 ext4_mb_use_group_pa(ac, pa); 3781 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3782 3783 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3784 lg = ac->ac_lg; 3785 BUG_ON(lg == NULL); 3786 3787 pa->pa_obj_lock = &lg->lg_prealloc_lock; 3788 pa->pa_inode = NULL; 3789 3790 ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3791 list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3792 ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3793 3794 /* 3795 * We will later add the new pa to the right bucket 3796 * after updating the pa_free in ext4_mb_release_context 3797 */ 3798 return 0; 3799 } 3800 3801 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3802 { 3803 int err; 3804 3805 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3806 err = ext4_mb_new_group_pa(ac); 3807 else 3808 err = ext4_mb_new_inode_pa(ac); 3809 return err; 3810 } 3811 3812 /* 3813 * finds all unused blocks in on-disk bitmap, frees them in 3814 * in-core bitmap and buddy. 3815 * @pa must be unlinked from inode and group lists, so that 3816 * nobody else can find/use it. 3817 * the caller MUST hold group/inode locks. 3818 * TODO: optimize the case when there are no in-core structures yet 3819 */ 3820 static noinline_for_stack int 3821 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 3822 struct ext4_prealloc_space *pa) 3823 { 3824 struct super_block *sb = e4b->bd_sb; 3825 struct ext4_sb_info *sbi = EXT4_SB(sb); 3826 unsigned int end; 3827 unsigned int next; 3828 ext4_group_t group; 3829 ext4_grpblk_t bit; 3830 unsigned long long grp_blk_start; 3831 int err = 0; 3832 int free = 0; 3833 3834 BUG_ON(pa->pa_deleted == 0); 3835 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3836 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 3837 BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3838 end = bit + pa->pa_len; 3839 3840 while (bit < end) { 3841 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3842 if (bit >= end) 3843 break; 3844 next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 3845 mb_debug(1, " free preallocated %u/%u in group %u\n", 3846 (unsigned) ext4_group_first_block_no(sb, group) + bit, 3847 (unsigned) next - bit, (unsigned) group); 3848 free += next - bit; 3849 3850 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 3851 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 3852 EXT4_C2B(sbi, bit)), 3853 next - bit); 3854 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3855 bit = next + 1; 3856 } 3857 if (free != pa->pa_free) { 3858 ext4_msg(e4b->bd_sb, KERN_CRIT, 3859 "pa %p: logic %lu, phys. %lu, len %lu", 3860 pa, (unsigned long) pa->pa_lstart, 3861 (unsigned long) pa->pa_pstart, 3862 (unsigned long) pa->pa_len); 3863 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 3864 free, pa->pa_free); 3865 /* 3866 * pa is already deleted so we use the value obtained 3867 * from the bitmap and continue. 3868 */ 3869 } 3870 atomic_add(free, &sbi->s_mb_discarded); 3871 3872 return err; 3873 } 3874 3875 static noinline_for_stack int 3876 ext4_mb_release_group_pa(struct ext4_buddy *e4b, 3877 struct ext4_prealloc_space *pa) 3878 { 3879 struct super_block *sb = e4b->bd_sb; 3880 ext4_group_t group; 3881 ext4_grpblk_t bit; 3882 3883 trace_ext4_mb_release_group_pa(sb, pa); 3884 BUG_ON(pa->pa_deleted == 0); 3885 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3886 BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3887 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3888 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 3889 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 3890 3891 return 0; 3892 } 3893 3894 /* 3895 * releases all preallocations in given group 3896 * 3897 * first, we need to decide discard policy: 3898 * - when do we discard 3899 * 1) ENOSPC 3900 * - how many do we discard 3901 * 1) how many requested 3902 */ 3903 static noinline_for_stack int 3904 ext4_mb_discard_group_preallocations(struct super_block *sb, 3905 ext4_group_t group, int needed) 3906 { 3907 struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3908 struct buffer_head *bitmap_bh = NULL; 3909 struct ext4_prealloc_space *pa, *tmp; 3910 struct list_head list; 3911 struct ext4_buddy e4b; 3912 int err; 3913 int busy = 0; 3914 int free = 0; 3915 3916 mb_debug(1, "discard preallocation for group %u\n", group); 3917 3918 if (list_empty(&grp->bb_prealloc_list)) 3919 return 0; 3920 3921 bitmap_bh = ext4_read_block_bitmap(sb, group); 3922 if (IS_ERR(bitmap_bh)) { 3923 err = PTR_ERR(bitmap_bh); 3924 ext4_error(sb, "Error %d reading block bitmap for %u", 3925 err, group); 3926 return 0; 3927 } 3928 3929 err = ext4_mb_load_buddy(sb, group, &e4b); 3930 if (err) { 3931 ext4_warning(sb, "Error %d loading buddy information for %u", 3932 err, group); 3933 put_bh(bitmap_bh); 3934 return 0; 3935 } 3936 3937 if (needed == 0) 3938 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 3939 3940 INIT_LIST_HEAD(&list); 3941 repeat: 3942 ext4_lock_group(sb, group); 3943 list_for_each_entry_safe(pa, tmp, 3944 &grp->bb_prealloc_list, pa_group_list) { 3945 spin_lock(&pa->pa_lock); 3946 if (atomic_read(&pa->pa_count)) { 3947 spin_unlock(&pa->pa_lock); 3948 busy = 1; 3949 continue; 3950 } 3951 if (pa->pa_deleted) { 3952 spin_unlock(&pa->pa_lock); 3953 continue; 3954 } 3955 3956 /* seems this one can be freed ... */ 3957 pa->pa_deleted = 1; 3958 3959 /* we can trust pa_free ... */ 3960 free += pa->pa_free; 3961 3962 spin_unlock(&pa->pa_lock); 3963 3964 list_del(&pa->pa_group_list); 3965 list_add(&pa->u.pa_tmp_list, &list); 3966 } 3967 3968 /* if we still need more blocks and some PAs were used, try again */ 3969 if (free < needed && busy) { 3970 busy = 0; 3971 ext4_unlock_group(sb, group); 3972 cond_resched(); 3973 goto repeat; 3974 } 3975 3976 /* found anything to free? */ 3977 if (list_empty(&list)) { 3978 BUG_ON(free != 0); 3979 goto out; 3980 } 3981 3982 /* now free all selected PAs */ 3983 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3984 3985 /* remove from object (inode or locality group) */ 3986 spin_lock(pa->pa_obj_lock); 3987 list_del_rcu(&pa->pa_inode_list); 3988 spin_unlock(pa->pa_obj_lock); 3989 3990 if (pa->pa_type == MB_GROUP_PA) 3991 ext4_mb_release_group_pa(&e4b, pa); 3992 else 3993 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3994 3995 list_del(&pa->u.pa_tmp_list); 3996 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3997 } 3998 3999 out: 4000 ext4_unlock_group(sb, group); 4001 ext4_mb_unload_buddy(&e4b); 4002 put_bh(bitmap_bh); 4003 return free; 4004 } 4005 4006 /* 4007 * releases all non-used preallocated blocks for given inode 4008 * 4009 * It's important to discard preallocations under i_data_sem 4010 * We don't want another block to be served from the prealloc 4011 * space when we are discarding the inode prealloc space. 4012 * 4013 * FIXME!! Make sure it is valid at all the call sites 4014 */ 4015 void ext4_discard_preallocations(struct inode *inode) 4016 { 4017 struct ext4_inode_info *ei = EXT4_I(inode); 4018 struct super_block *sb = inode->i_sb; 4019 struct buffer_head *bitmap_bh = NULL; 4020 struct ext4_prealloc_space *pa, *tmp; 4021 ext4_group_t group = 0; 4022 struct list_head list; 4023 struct ext4_buddy e4b; 4024 int err; 4025 4026 if (!S_ISREG(inode->i_mode)) { 4027 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 4028 return; 4029 } 4030 4031 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); 4032 trace_ext4_discard_preallocations(inode); 4033 4034 INIT_LIST_HEAD(&list); 4035 4036 repeat: 4037 /* first, collect all pa's in the inode */ 4038 spin_lock(&ei->i_prealloc_lock); 4039 while (!list_empty(&ei->i_prealloc_list)) { 4040 pa = list_entry(ei->i_prealloc_list.next, 4041 struct ext4_prealloc_space, pa_inode_list); 4042 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 4043 spin_lock(&pa->pa_lock); 4044 if (atomic_read(&pa->pa_count)) { 4045 /* this shouldn't happen often - nobody should 4046 * use preallocation while we're discarding it */ 4047 spin_unlock(&pa->pa_lock); 4048 spin_unlock(&ei->i_prealloc_lock); 4049 ext4_msg(sb, KERN_ERR, 4050 "uh-oh! used pa while discarding"); 4051 WARN_ON(1); 4052 schedule_timeout_uninterruptible(HZ); 4053 goto repeat; 4054 4055 } 4056 if (pa->pa_deleted == 0) { 4057 pa->pa_deleted = 1; 4058 spin_unlock(&pa->pa_lock); 4059 list_del_rcu(&pa->pa_inode_list); 4060 list_add(&pa->u.pa_tmp_list, &list); 4061 continue; 4062 } 4063 4064 /* someone is deleting pa right now */ 4065 spin_unlock(&pa->pa_lock); 4066 spin_unlock(&ei->i_prealloc_lock); 4067 4068 /* we have to wait here because pa_deleted 4069 * doesn't mean pa is already unlinked from 4070 * the list. as we might be called from 4071 * ->clear_inode() the inode will get freed 4072 * and concurrent thread which is unlinking 4073 * pa from inode's list may access already 4074 * freed memory, bad-bad-bad */ 4075 4076 /* XXX: if this happens too often, we can 4077 * add a flag to force wait only in case 4078 * of ->clear_inode(), but not in case of 4079 * regular truncate */ 4080 schedule_timeout_uninterruptible(HZ); 4081 goto repeat; 4082 } 4083 spin_unlock(&ei->i_prealloc_lock); 4084 4085 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 4086 BUG_ON(pa->pa_type != MB_INODE_PA); 4087 group = ext4_get_group_number(sb, pa->pa_pstart); 4088 4089 err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 4090 GFP_NOFS|__GFP_NOFAIL); 4091 if (err) { 4092 ext4_error(sb, "Error %d loading buddy information for %u", 4093 err, group); 4094 continue; 4095 } 4096 4097 bitmap_bh = ext4_read_block_bitmap(sb, group); 4098 if (IS_ERR(bitmap_bh)) { 4099 err = PTR_ERR(bitmap_bh); 4100 ext4_error(sb, "Error %d reading block bitmap for %u", 4101 err, group); 4102 ext4_mb_unload_buddy(&e4b); 4103 continue; 4104 } 4105 4106 ext4_lock_group(sb, group); 4107 list_del(&pa->pa_group_list); 4108 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 4109 ext4_unlock_group(sb, group); 4110 4111 ext4_mb_unload_buddy(&e4b); 4112 put_bh(bitmap_bh); 4113 4114 list_del(&pa->u.pa_tmp_list); 4115 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 4116 } 4117 } 4118 4119 #ifdef CONFIG_EXT4_DEBUG 4120 static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4121 { 4122 struct super_block *sb = ac->ac_sb; 4123 ext4_group_t ngroups, i; 4124 4125 if (!ext4_mballoc_debug || 4126 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) 4127 return; 4128 4129 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" 4130 " Allocation context details:"); 4131 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", 4132 ac->ac_status, ac->ac_flags); 4133 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " 4134 "goal %lu/%lu/%lu@%lu, " 4135 "best %lu/%lu/%lu@%lu cr %d", 4136 (unsigned long)ac->ac_o_ex.fe_group, 4137 (unsigned long)ac->ac_o_ex.fe_start, 4138 (unsigned long)ac->ac_o_ex.fe_len, 4139 (unsigned long)ac->ac_o_ex.fe_logical, 4140 (unsigned long)ac->ac_g_ex.fe_group, 4141 (unsigned long)ac->ac_g_ex.fe_start, 4142 (unsigned long)ac->ac_g_ex.fe_len, 4143 (unsigned long)ac->ac_g_ex.fe_logical, 4144 (unsigned long)ac->ac_b_ex.fe_group, 4145 (unsigned long)ac->ac_b_ex.fe_start, 4146 (unsigned long)ac->ac_b_ex.fe_len, 4147 (unsigned long)ac->ac_b_ex.fe_logical, 4148 (int)ac->ac_criteria); 4149 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); 4150 ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); 4151 ngroups = ext4_get_groups_count(sb); 4152 for (i = 0; i < ngroups; i++) { 4153 struct ext4_group_info *grp = ext4_get_group_info(sb, i); 4154 struct ext4_prealloc_space *pa; 4155 ext4_grpblk_t start; 4156 struct list_head *cur; 4157 ext4_lock_group(sb, i); 4158 list_for_each(cur, &grp->bb_prealloc_list) { 4159 pa = list_entry(cur, struct ext4_prealloc_space, 4160 pa_group_list); 4161 spin_lock(&pa->pa_lock); 4162 ext4_get_group_no_and_offset(sb, pa->pa_pstart, 4163 NULL, &start); 4164 spin_unlock(&pa->pa_lock); 4165 printk(KERN_ERR "PA:%u:%d:%u \n", i, 4166 start, pa->pa_len); 4167 } 4168 ext4_unlock_group(sb, i); 4169 4170 if (grp->bb_free == 0) 4171 continue; 4172 printk(KERN_ERR "%u: %d/%d \n", 4173 i, grp->bb_free, grp->bb_fragments); 4174 } 4175 printk(KERN_ERR "\n"); 4176 } 4177 #else 4178 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4179 { 4180 return; 4181 } 4182 #endif 4183 4184 /* 4185 * We use locality group preallocation for small size file. The size of the 4186 * file is determined by the current size or the resulting size after 4187 * allocation which ever is larger 4188 * 4189 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 4190 */ 4191 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 4192 { 4193 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4194 int bsbits = ac->ac_sb->s_blocksize_bits; 4195 loff_t size, isize; 4196 4197 if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4198 return; 4199 4200 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 4201 return; 4202 4203 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 4204 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 4205 >> bsbits; 4206 4207 if ((size == isize) && 4208 !ext4_fs_is_busy(sbi) && 4209 (atomic_read(&ac->ac_inode->i_writecount) == 0)) { 4210 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 4211 return; 4212 } 4213 4214 if (sbi->s_mb_group_prealloc <= 0) { 4215 ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4216 return; 4217 } 4218 4219 /* don't use group allocation for large files */ 4220 size = max(size, isize); 4221 if (size > sbi->s_mb_stream_request) { 4222 ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4223 return; 4224 } 4225 4226 BUG_ON(ac->ac_lg != NULL); 4227 /* 4228 * locality group prealloc space are per cpu. The reason for having 4229 * per cpu locality group is to reduce the contention between block 4230 * request from multiple CPUs. 4231 */ 4232 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); 4233 4234 /* we're going to use group allocation */ 4235 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 4236 4237 /* serialize all allocations in the group */ 4238 mutex_lock(&ac->ac_lg->lg_mutex); 4239 } 4240 4241 static noinline_for_stack int 4242 ext4_mb_initialize_context(struct ext4_allocation_context *ac, 4243 struct ext4_allocation_request *ar) 4244 { 4245 struct super_block *sb = ar->inode->i_sb; 4246 struct ext4_sb_info *sbi = EXT4_SB(sb); 4247 struct ext4_super_block *es = sbi->s_es; 4248 ext4_group_t group; 4249 unsigned int len; 4250 ext4_fsblk_t goal; 4251 ext4_grpblk_t block; 4252 4253 /* we can't allocate > group size */ 4254 len = ar->len; 4255 4256 /* just a dirty hack to filter too big requests */ 4257 if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) 4258 len = EXT4_CLUSTERS_PER_GROUP(sb); 4259 4260 /* start searching from the goal */ 4261 goal = ar->goal; 4262 if (goal < le32_to_cpu(es->s_first_data_block) || 4263 goal >= ext4_blocks_count(es)) 4264 goal = le32_to_cpu(es->s_first_data_block); 4265 ext4_get_group_no_and_offset(sb, goal, &group, &block); 4266 4267 /* set up allocation goals */ 4268 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); 4269 ac->ac_status = AC_STATUS_CONTINUE; 4270 ac->ac_sb = sb; 4271 ac->ac_inode = ar->inode; 4272 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 4273 ac->ac_o_ex.fe_group = group; 4274 ac->ac_o_ex.fe_start = block; 4275 ac->ac_o_ex.fe_len = len; 4276 ac->ac_g_ex = ac->ac_o_ex; 4277 ac->ac_flags = ar->flags; 4278 4279 /* we have to define context: we'll we work with a file or 4280 * locality group. this is a policy, actually */ 4281 ext4_mb_group_or_file(ac); 4282 4283 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 4284 "left: %u/%u, right %u/%u to %swritable\n", 4285 (unsigned) ar->len, (unsigned) ar->logical, 4286 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 4287 (unsigned) ar->lleft, (unsigned) ar->pleft, 4288 (unsigned) ar->lright, (unsigned) ar->pright, 4289 atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 4290 return 0; 4291 4292 } 4293 4294 static noinline_for_stack void 4295 ext4_mb_discard_lg_preallocations(struct super_block *sb, 4296 struct ext4_locality_group *lg, 4297 int order, int total_entries) 4298 { 4299 ext4_group_t group = 0; 4300 struct ext4_buddy e4b; 4301 struct list_head discard_list; 4302 struct ext4_prealloc_space *pa, *tmp; 4303 4304 mb_debug(1, "discard locality group preallocation\n"); 4305 4306 INIT_LIST_HEAD(&discard_list); 4307 4308 spin_lock(&lg->lg_prealloc_lock); 4309 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 4310 pa_inode_list) { 4311 spin_lock(&pa->pa_lock); 4312 if (atomic_read(&pa->pa_count)) { 4313 /* 4314 * This is the pa that we just used 4315 * for block allocation. So don't 4316 * free that 4317 */ 4318 spin_unlock(&pa->pa_lock); 4319 continue; 4320 } 4321 if (pa->pa_deleted) { 4322 spin_unlock(&pa->pa_lock); 4323 continue; 4324 } 4325 /* only lg prealloc space */ 4326 BUG_ON(pa->pa_type != MB_GROUP_PA); 4327 4328 /* seems this one can be freed ... */ 4329 pa->pa_deleted = 1; 4330 spin_unlock(&pa->pa_lock); 4331 4332 list_del_rcu(&pa->pa_inode_list); 4333 list_add(&pa->u.pa_tmp_list, &discard_list); 4334 4335 total_entries--; 4336 if (total_entries <= 5) { 4337 /* 4338 * we want to keep only 5 entries 4339 * allowing it to grow to 8. This 4340 * mak sure we don't call discard 4341 * soon for this list. 4342 */ 4343 break; 4344 } 4345 } 4346 spin_unlock(&lg->lg_prealloc_lock); 4347 4348 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 4349 int err; 4350 4351 group = ext4_get_group_number(sb, pa->pa_pstart); 4352 err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 4353 GFP_NOFS|__GFP_NOFAIL); 4354 if (err) { 4355 ext4_error(sb, "Error %d loading buddy information for %u", 4356 err, group); 4357 continue; 4358 } 4359 ext4_lock_group(sb, group); 4360 list_del(&pa->pa_group_list); 4361 ext4_mb_release_group_pa(&e4b, pa); 4362 ext4_unlock_group(sb, group); 4363 4364 ext4_mb_unload_buddy(&e4b); 4365 list_del(&pa->u.pa_tmp_list); 4366 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 4367 } 4368 } 4369 4370 /* 4371 * We have incremented pa_count. So it cannot be freed at this 4372 * point. Also we hold lg_mutex. So no parallel allocation is 4373 * possible from this lg. That means pa_free cannot be updated. 4374 * 4375 * A parallel ext4_mb_discard_group_preallocations is possible. 4376 * which can cause the lg_prealloc_list to be updated. 4377 */ 4378 4379 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 4380 { 4381 int order, added = 0, lg_prealloc_count = 1; 4382 struct super_block *sb = ac->ac_sb; 4383 struct ext4_locality_group *lg = ac->ac_lg; 4384 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 4385 4386 order = fls(pa->pa_free) - 1; 4387 if (order > PREALLOC_TB_SIZE - 1) 4388 /* The max size of hash table is PREALLOC_TB_SIZE */ 4389 order = PREALLOC_TB_SIZE - 1; 4390 /* Add the prealloc space to lg */ 4391 spin_lock(&lg->lg_prealloc_lock); 4392 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 4393 pa_inode_list) { 4394 spin_lock(&tmp_pa->pa_lock); 4395 if (tmp_pa->pa_deleted) { 4396 spin_unlock(&tmp_pa->pa_lock); 4397 continue; 4398 } 4399 if (!added && pa->pa_free < tmp_pa->pa_free) { 4400 /* Add to the tail of the previous entry */ 4401 list_add_tail_rcu(&pa->pa_inode_list, 4402 &tmp_pa->pa_inode_list); 4403 added = 1; 4404 /* 4405 * we want to count the total 4406 * number of entries in the list 4407 */ 4408 } 4409 spin_unlock(&tmp_pa->pa_lock); 4410 lg_prealloc_count++; 4411 } 4412 if (!added) 4413 list_add_tail_rcu(&pa->pa_inode_list, 4414 &lg->lg_prealloc_list[order]); 4415 spin_unlock(&lg->lg_prealloc_lock); 4416 4417 /* Now trim the list to be not more than 8 elements */ 4418 if (lg_prealloc_count > 8) { 4419 ext4_mb_discard_lg_preallocations(sb, lg, 4420 order, lg_prealloc_count); 4421 return; 4422 } 4423 return ; 4424 } 4425 4426 /* 4427 * release all resource we used in allocation 4428 */ 4429 static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4430 { 4431 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4432 struct ext4_prealloc_space *pa = ac->ac_pa; 4433 if (pa) { 4434 if (pa->pa_type == MB_GROUP_PA) { 4435 /* see comment in ext4_mb_use_group_pa() */ 4436 spin_lock(&pa->pa_lock); 4437 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 4438 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 4439 pa->pa_free -= ac->ac_b_ex.fe_len; 4440 pa->pa_len -= ac->ac_b_ex.fe_len; 4441 spin_unlock(&pa->pa_lock); 4442 } 4443 } 4444 if (pa) { 4445 /* 4446 * We want to add the pa to the right bucket. 4447 * Remove it from the list and while adding 4448 * make sure the list to which we are adding 4449 * doesn't grow big. 4450 */ 4451 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { 4452 spin_lock(pa->pa_obj_lock); 4453 list_del_rcu(&pa->pa_inode_list); 4454 spin_unlock(pa->pa_obj_lock); 4455 ext4_mb_add_n_trim(ac); 4456 } 4457 ext4_mb_put_pa(ac, ac->ac_sb, pa); 4458 } 4459 if (ac->ac_bitmap_page) 4460 put_page(ac->ac_bitmap_page); 4461 if (ac->ac_buddy_page) 4462 put_page(ac->ac_buddy_page); 4463 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4464 mutex_unlock(&ac->ac_lg->lg_mutex); 4465 ext4_mb_collect_stats(ac); 4466 return 0; 4467 } 4468 4469 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4470 { 4471 ext4_group_t i, ngroups = ext4_get_groups_count(sb); 4472 int ret; 4473 int freed = 0; 4474 4475 trace_ext4_mb_discard_preallocations(sb, needed); 4476 for (i = 0; i < ngroups && needed > 0; i++) { 4477 ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4478 freed += ret; 4479 needed -= ret; 4480 } 4481 4482 return freed; 4483 } 4484 4485 /* 4486 * Main entry point into mballoc to allocate blocks 4487 * it tries to use preallocation first, then falls back 4488 * to usual allocation 4489 */ 4490 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4491 struct ext4_allocation_request *ar, int *errp) 4492 { 4493 int freed; 4494 struct ext4_allocation_context *ac = NULL; 4495 struct ext4_sb_info *sbi; 4496 struct super_block *sb; 4497 ext4_fsblk_t block = 0; 4498 unsigned int inquota = 0; 4499 unsigned int reserv_clstrs = 0; 4500 4501 might_sleep(); 4502 sb = ar->inode->i_sb; 4503 sbi = EXT4_SB(sb); 4504 4505 trace_ext4_request_blocks(ar); 4506 4507 /* Allow to use superuser reservation for quota file */ 4508 if (ext4_is_quota_file(ar->inode)) 4509 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 4510 4511 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { 4512 /* Without delayed allocation we need to verify 4513 * there is enough free blocks to do block allocation 4514 * and verify allocation doesn't exceed the quota limits. 4515 */ 4516 while (ar->len && 4517 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 4518 4519 /* let others to free the space */ 4520 cond_resched(); 4521 ar->len = ar->len >> 1; 4522 } 4523 if (!ar->len) { 4524 *errp = -ENOSPC; 4525 return 0; 4526 } 4527 reserv_clstrs = ar->len; 4528 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 4529 dquot_alloc_block_nofail(ar->inode, 4530 EXT4_C2B(sbi, ar->len)); 4531 } else { 4532 while (ar->len && 4533 dquot_alloc_block(ar->inode, 4534 EXT4_C2B(sbi, ar->len))) { 4535 4536 ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4537 ar->len--; 4538 } 4539 } 4540 inquota = ar->len; 4541 if (ar->len == 0) { 4542 *errp = -EDQUOT; 4543 goto out; 4544 } 4545 } 4546 4547 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); 4548 if (!ac) { 4549 ar->len = 0; 4550 *errp = -ENOMEM; 4551 goto out; 4552 } 4553 4554 *errp = ext4_mb_initialize_context(ac, ar); 4555 if (*errp) { 4556 ar->len = 0; 4557 goto out; 4558 } 4559 4560 ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4561 if (!ext4_mb_use_preallocated(ac)) { 4562 ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4563 ext4_mb_normalize_request(ac, ar); 4564 repeat: 4565 /* allocate space in core */ 4566 *errp = ext4_mb_regular_allocator(ac); 4567 if (*errp) 4568 goto discard_and_exit; 4569 4570 /* as we've just preallocated more space than 4571 * user requested originally, we store allocated 4572 * space in a special descriptor */ 4573 if (ac->ac_status == AC_STATUS_FOUND && 4574 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 4575 *errp = ext4_mb_new_preallocation(ac); 4576 if (*errp) { 4577 discard_and_exit: 4578 ext4_discard_allocated_blocks(ac); 4579 goto errout; 4580 } 4581 } 4582 if (likely(ac->ac_status == AC_STATUS_FOUND)) { 4583 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 4584 if (*errp) { 4585 ext4_discard_allocated_blocks(ac); 4586 goto errout; 4587 } else { 4588 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4589 ar->len = ac->ac_b_ex.fe_len; 4590 } 4591 } else { 4592 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4593 if (freed) 4594 goto repeat; 4595 *errp = -ENOSPC; 4596 } 4597 4598 errout: 4599 if (*errp) { 4600 ac->ac_b_ex.fe_len = 0; 4601 ar->len = 0; 4602 ext4_mb_show_ac(ac); 4603 } 4604 ext4_mb_release_context(ac); 4605 out: 4606 if (ac) 4607 kmem_cache_free(ext4_ac_cachep, ac); 4608 if (inquota && ar->len < inquota) 4609 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 4610 if (!ar->len) { 4611 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) 4612 /* release all the reserved blocks if non delalloc */ 4613 percpu_counter_sub(&sbi->s_dirtyclusters_counter, 4614 reserv_clstrs); 4615 } 4616 4617 trace_ext4_allocate_blocks(ar, (unsigned long long)block); 4618 4619 return block; 4620 } 4621 4622 /* 4623 * We can merge two free data extents only if the physical blocks 4624 * are contiguous, AND the extents were freed by the same transaction, 4625 * AND the blocks are associated with the same group. 4626 */ 4627 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, 4628 struct ext4_free_data *entry, 4629 struct ext4_free_data *new_entry, 4630 struct rb_root *entry_rb_root) 4631 { 4632 if ((entry->efd_tid != new_entry->efd_tid) || 4633 (entry->efd_group != new_entry->efd_group)) 4634 return; 4635 if (entry->efd_start_cluster + entry->efd_count == 4636 new_entry->efd_start_cluster) { 4637 new_entry->efd_start_cluster = entry->efd_start_cluster; 4638 new_entry->efd_count += entry->efd_count; 4639 } else if (new_entry->efd_start_cluster + new_entry->efd_count == 4640 entry->efd_start_cluster) { 4641 new_entry->efd_count += entry->efd_count; 4642 } else 4643 return; 4644 spin_lock(&sbi->s_md_lock); 4645 list_del(&entry->efd_list); 4646 spin_unlock(&sbi->s_md_lock); 4647 rb_erase(&entry->efd_node, entry_rb_root); 4648 kmem_cache_free(ext4_free_data_cachep, entry); 4649 } 4650 4651 static noinline_for_stack int 4652 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 4653 struct ext4_free_data *new_entry) 4654 { 4655 ext4_group_t group = e4b->bd_group; 4656 ext4_grpblk_t cluster; 4657 ext4_grpblk_t clusters = new_entry->efd_count; 4658 struct ext4_free_data *entry; 4659 struct ext4_group_info *db = e4b->bd_info; 4660 struct super_block *sb = e4b->bd_sb; 4661 struct ext4_sb_info *sbi = EXT4_SB(sb); 4662 struct rb_node **n = &db->bb_free_root.rb_node, *node; 4663 struct rb_node *parent = NULL, *new_node; 4664 4665 BUG_ON(!ext4_handle_valid(handle)); 4666 BUG_ON(e4b->bd_bitmap_page == NULL); 4667 BUG_ON(e4b->bd_buddy_page == NULL); 4668 4669 new_node = &new_entry->efd_node; 4670 cluster = new_entry->efd_start_cluster; 4671 4672 if (!*n) { 4673 /* first free block exent. We need to 4674 protect buddy cache from being freed, 4675 * otherwise we'll refresh it from 4676 * on-disk bitmap and lose not-yet-available 4677 * blocks */ 4678 get_page(e4b->bd_buddy_page); 4679 get_page(e4b->bd_bitmap_page); 4680 } 4681 while (*n) { 4682 parent = *n; 4683 entry = rb_entry(parent, struct ext4_free_data, efd_node); 4684 if (cluster < entry->efd_start_cluster) 4685 n = &(*n)->rb_left; 4686 else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 4687 n = &(*n)->rb_right; 4688 else { 4689 ext4_grp_locked_error(sb, group, 0, 4690 ext4_group_first_block_no(sb, group) + 4691 EXT4_C2B(sbi, cluster), 4692 "Block already on to-be-freed list"); 4693 return 0; 4694 } 4695 } 4696 4697 rb_link_node(new_node, parent, n); 4698 rb_insert_color(new_node, &db->bb_free_root); 4699 4700 /* Now try to see the extent can be merged to left and right */ 4701 node = rb_prev(new_node); 4702 if (node) { 4703 entry = rb_entry(node, struct ext4_free_data, efd_node); 4704 ext4_try_merge_freed_extent(sbi, entry, new_entry, 4705 &(db->bb_free_root)); 4706 } 4707 4708 node = rb_next(new_node); 4709 if (node) { 4710 entry = rb_entry(node, struct ext4_free_data, efd_node); 4711 ext4_try_merge_freed_extent(sbi, entry, new_entry, 4712 &(db->bb_free_root)); 4713 } 4714 4715 spin_lock(&sbi->s_md_lock); 4716 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list); 4717 sbi->s_mb_free_pending += clusters; 4718 spin_unlock(&sbi->s_md_lock); 4719 return 0; 4720 } 4721 4722 /** 4723 * ext4_free_blocks() -- Free given blocks and update quota 4724 * @handle: handle for this transaction 4725 * @inode: inode 4726 * @block: start physical block to free 4727 * @count: number of blocks to count 4728 * @flags: flags used by ext4_free_blocks 4729 */ 4730 void ext4_free_blocks(handle_t *handle, struct inode *inode, 4731 struct buffer_head *bh, ext4_fsblk_t block, 4732 unsigned long count, int flags) 4733 { 4734 struct buffer_head *bitmap_bh = NULL; 4735 struct super_block *sb = inode->i_sb; 4736 struct ext4_group_desc *gdp; 4737 unsigned int overflow; 4738 ext4_grpblk_t bit; 4739 struct buffer_head *gd_bh; 4740 ext4_group_t block_group; 4741 struct ext4_sb_info *sbi; 4742 struct ext4_buddy e4b; 4743 unsigned int count_clusters; 4744 int err = 0; 4745 int ret; 4746 4747 might_sleep(); 4748 if (bh) { 4749 if (block) 4750 BUG_ON(block != bh->b_blocknr); 4751 else 4752 block = bh->b_blocknr; 4753 } 4754 4755 sbi = EXT4_SB(sb); 4756 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 4757 !ext4_data_block_valid(sbi, block, count)) { 4758 ext4_error(sb, "Freeing blocks not in datazone - " 4759 "block = %llu, count = %lu", block, count); 4760 goto error_return; 4761 } 4762 4763 ext4_debug("freeing block %llu\n", block); 4764 trace_ext4_free_blocks(inode, block, count, flags); 4765 4766 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 4767 BUG_ON(count > 1); 4768 4769 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 4770 inode, bh, block); 4771 } 4772 4773 /* 4774 * If the extent to be freed does not begin on a cluster 4775 * boundary, we need to deal with partial clusters at the 4776 * beginning and end of the extent. Normally we will free 4777 * blocks at the beginning or the end unless we are explicitly 4778 * requested to avoid doing so. 4779 */ 4780 overflow = EXT4_PBLK_COFF(sbi, block); 4781 if (overflow) { 4782 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 4783 overflow = sbi->s_cluster_ratio - overflow; 4784 block += overflow; 4785 if (count > overflow) 4786 count -= overflow; 4787 else 4788 return; 4789 } else { 4790 block -= overflow; 4791 count += overflow; 4792 } 4793 } 4794 overflow = EXT4_LBLK_COFF(sbi, count); 4795 if (overflow) { 4796 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 4797 if (count > overflow) 4798 count -= overflow; 4799 else 4800 return; 4801 } else 4802 count += sbi->s_cluster_ratio - overflow; 4803 } 4804 4805 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 4806 int i; 4807 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; 4808 4809 for (i = 0; i < count; i++) { 4810 cond_resched(); 4811 if (is_metadata) 4812 bh = sb_find_get_block(inode->i_sb, block + i); 4813 ext4_forget(handle, is_metadata, inode, bh, block + i); 4814 } 4815 } 4816 4817 do_more: 4818 overflow = 0; 4819 ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4820 4821 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( 4822 ext4_get_group_info(sb, block_group)))) 4823 return; 4824 4825 /* 4826 * Check to see if we are freeing blocks across a group 4827 * boundary. 4828 */ 4829 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 4830 overflow = EXT4_C2B(sbi, bit) + count - 4831 EXT4_BLOCKS_PER_GROUP(sb); 4832 count -= overflow; 4833 } 4834 count_clusters = EXT4_NUM_B2C(sbi, count); 4835 bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4836 if (IS_ERR(bitmap_bh)) { 4837 err = PTR_ERR(bitmap_bh); 4838 bitmap_bh = NULL; 4839 goto error_return; 4840 } 4841 gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4842 if (!gdp) { 4843 err = -EIO; 4844 goto error_return; 4845 } 4846 4847 if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4848 in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4849 in_range(block, ext4_inode_table(sb, gdp), 4850 EXT4_SB(sb)->s_itb_per_group) || 4851 in_range(block + count - 1, ext4_inode_table(sb, gdp), 4852 EXT4_SB(sb)->s_itb_per_group)) { 4853 4854 ext4_error(sb, "Freeing blocks in system zone - " 4855 "Block = %llu, count = %lu", block, count); 4856 /* err = 0. ext4_std_error should be a no op */ 4857 goto error_return; 4858 } 4859 4860 BUFFER_TRACE(bitmap_bh, "getting write access"); 4861 err = ext4_journal_get_write_access(handle, bitmap_bh); 4862 if (err) 4863 goto error_return; 4864 4865 /* 4866 * We are about to modify some metadata. Call the journal APIs 4867 * to unshare ->b_data if a currently-committing transaction is 4868 * using it 4869 */ 4870 BUFFER_TRACE(gd_bh, "get_write_access"); 4871 err = ext4_journal_get_write_access(handle, gd_bh); 4872 if (err) 4873 goto error_return; 4874 #ifdef AGGRESSIVE_CHECK 4875 { 4876 int i; 4877 for (i = 0; i < count_clusters; i++) 4878 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4879 } 4880 #endif 4881 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 4882 4883 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ 4884 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, 4885 GFP_NOFS|__GFP_NOFAIL); 4886 if (err) 4887 goto error_return; 4888 4889 /* 4890 * We need to make sure we don't reuse the freed block until after the 4891 * transaction is committed. We make an exception if the inode is to be 4892 * written in writeback mode since writeback mode has weak data 4893 * consistency guarantees. 4894 */ 4895 if (ext4_handle_valid(handle) && 4896 ((flags & EXT4_FREE_BLOCKS_METADATA) || 4897 !ext4_should_writeback_data(inode))) { 4898 struct ext4_free_data *new_entry; 4899 /* 4900 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed 4901 * to fail. 4902 */ 4903 new_entry = kmem_cache_alloc(ext4_free_data_cachep, 4904 GFP_NOFS|__GFP_NOFAIL); 4905 new_entry->efd_start_cluster = bit; 4906 new_entry->efd_group = block_group; 4907 new_entry->efd_count = count_clusters; 4908 new_entry->efd_tid = handle->h_transaction->t_tid; 4909 4910 ext4_lock_group(sb, block_group); 4911 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 4912 ext4_mb_free_metadata(handle, &e4b, new_entry); 4913 } else { 4914 /* need to update group_info->bb_free and bitmap 4915 * with group lock held. generate_buddy look at 4916 * them with group lock_held 4917 */ 4918 if (test_opt(sb, DISCARD)) { 4919 err = ext4_issue_discard(sb, block_group, bit, count, 4920 NULL); 4921 if (err && err != -EOPNOTSUPP) 4922 ext4_msg(sb, KERN_WARNING, "discard request in" 4923 " group:%d block:%d count:%lu failed" 4924 " with %d", block_group, bit, count, 4925 err); 4926 } else 4927 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); 4928 4929 ext4_lock_group(sb, block_group); 4930 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 4931 mb_free_blocks(inode, &e4b, bit, count_clusters); 4932 } 4933 4934 ret = ext4_free_group_clusters(sb, gdp) + count_clusters; 4935 ext4_free_group_clusters_set(sb, gdp, ret); 4936 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); 4937 ext4_group_desc_csum_set(sb, block_group, gdp); 4938 ext4_unlock_group(sb, block_group); 4939 4940 if (sbi->s_log_groups_per_flex) { 4941 ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 4942 atomic64_add(count_clusters, 4943 &sbi->s_flex_groups[flex_group].free_clusters); 4944 } 4945 4946 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 4947 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 4948 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); 4949 4950 ext4_mb_unload_buddy(&e4b); 4951 4952 /* We dirtied the bitmap block */ 4953 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 4954 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4955 4956 /* And the group descriptor block */ 4957 BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 4958 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4959 if (!err) 4960 err = ret; 4961 4962 if (overflow && !err) { 4963 block += count; 4964 count = overflow; 4965 put_bh(bitmap_bh); 4966 goto do_more; 4967 } 4968 error_return: 4969 brelse(bitmap_bh); 4970 ext4_std_error(sb, err); 4971 return; 4972 } 4973 4974 /** 4975 * ext4_group_add_blocks() -- Add given blocks to an existing group 4976 * @handle: handle to this transaction 4977 * @sb: super block 4978 * @block: start physical block to add to the block group 4979 * @count: number of blocks to free 4980 * 4981 * This marks the blocks as free in the bitmap and buddy. 4982 */ 4983 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 4984 ext4_fsblk_t block, unsigned long count) 4985 { 4986 struct buffer_head *bitmap_bh = NULL; 4987 struct buffer_head *gd_bh; 4988 ext4_group_t block_group; 4989 ext4_grpblk_t bit; 4990 unsigned int i; 4991 struct ext4_group_desc *desc; 4992 struct ext4_sb_info *sbi = EXT4_SB(sb); 4993 struct ext4_buddy e4b; 4994 int err = 0, ret, blk_free_count; 4995 ext4_grpblk_t blocks_freed; 4996 4997 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 4998 4999 if (count == 0) 5000 return 0; 5001 5002 ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 5003 /* 5004 * Check to see if we are freeing blocks across a group 5005 * boundary. 5006 */ 5007 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { 5008 ext4_warning(sb, "too much blocks added to group %u", 5009 block_group); 5010 err = -EINVAL; 5011 goto error_return; 5012 } 5013 5014 bitmap_bh = ext4_read_block_bitmap(sb, block_group); 5015 if (IS_ERR(bitmap_bh)) { 5016 err = PTR_ERR(bitmap_bh); 5017 bitmap_bh = NULL; 5018 goto error_return; 5019 } 5020 5021 desc = ext4_get_group_desc(sb, block_group, &gd_bh); 5022 if (!desc) { 5023 err = -EIO; 5024 goto error_return; 5025 } 5026 5027 if (in_range(ext4_block_bitmap(sb, desc), block, count) || 5028 in_range(ext4_inode_bitmap(sb, desc), block, count) || 5029 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || 5030 in_range(block + count - 1, ext4_inode_table(sb, desc), 5031 sbi->s_itb_per_group)) { 5032 ext4_error(sb, "Adding blocks in system zones - " 5033 "Block = %llu, count = %lu", 5034 block, count); 5035 err = -EINVAL; 5036 goto error_return; 5037 } 5038 5039 BUFFER_TRACE(bitmap_bh, "getting write access"); 5040 err = ext4_journal_get_write_access(handle, bitmap_bh); 5041 if (err) 5042 goto error_return; 5043 5044 /* 5045 * We are about to modify some metadata. Call the journal APIs 5046 * to unshare ->b_data if a currently-committing transaction is 5047 * using it 5048 */ 5049 BUFFER_TRACE(gd_bh, "get_write_access"); 5050 err = ext4_journal_get_write_access(handle, gd_bh); 5051 if (err) 5052 goto error_return; 5053 5054 for (i = 0, blocks_freed = 0; i < count; i++) { 5055 BUFFER_TRACE(bitmap_bh, "clear bit"); 5056 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { 5057 ext4_error(sb, "bit already cleared for block %llu", 5058 (ext4_fsblk_t)(block + i)); 5059 BUFFER_TRACE(bitmap_bh, "bit already cleared"); 5060 } else { 5061 blocks_freed++; 5062 } 5063 } 5064 5065 err = ext4_mb_load_buddy(sb, block_group, &e4b); 5066 if (err) 5067 goto error_return; 5068 5069 /* 5070 * need to update group_info->bb_free and bitmap 5071 * with group lock held. generate_buddy look at 5072 * them with group lock_held 5073 */ 5074 ext4_lock_group(sb, block_group); 5075 mb_clear_bits(bitmap_bh->b_data, bit, count); 5076 mb_free_blocks(NULL, &e4b, bit, count); 5077 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); 5078 ext4_free_group_clusters_set(sb, desc, blk_free_count); 5079 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); 5080 ext4_group_desc_csum_set(sb, block_group, desc); 5081 ext4_unlock_group(sb, block_group); 5082 percpu_counter_add(&sbi->s_freeclusters_counter, 5083 EXT4_NUM_B2C(sbi, blocks_freed)); 5084 5085 if (sbi->s_log_groups_per_flex) { 5086 ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 5087 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed), 5088 &sbi->s_flex_groups[flex_group].free_clusters); 5089 } 5090 5091 ext4_mb_unload_buddy(&e4b); 5092 5093 /* We dirtied the bitmap block */ 5094 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 5095 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 5096 5097 /* And the group descriptor block */ 5098 BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 5099 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 5100 if (!err) 5101 err = ret; 5102 5103 error_return: 5104 brelse(bitmap_bh); 5105 ext4_std_error(sb, err); 5106 return err; 5107 } 5108 5109 /** 5110 * ext4_trim_extent -- function to TRIM one single free extent in the group 5111 * @sb: super block for the file system 5112 * @start: starting block of the free extent in the alloc. group 5113 * @count: number of blocks to TRIM 5114 * @group: alloc. group we are working with 5115 * @e4b: ext4 buddy for the group 5116 * 5117 * Trim "count" blocks starting at "start" in the "group". To assure that no 5118 * one will allocate those blocks, mark it as used in buddy bitmap. This must 5119 * be called with under the group lock. 5120 */ 5121 static int ext4_trim_extent(struct super_block *sb, int start, int count, 5122 ext4_group_t group, struct ext4_buddy *e4b) 5123 __releases(bitlock) 5124 __acquires(bitlock) 5125 { 5126 struct ext4_free_extent ex; 5127 int ret = 0; 5128 5129 trace_ext4_trim_extent(sb, group, start, count); 5130 5131 assert_spin_locked(ext4_group_lock_ptr(sb, group)); 5132 5133 ex.fe_start = start; 5134 ex.fe_group = group; 5135 ex.fe_len = count; 5136 5137 /* 5138 * Mark blocks used, so no one can reuse them while 5139 * being trimmed. 5140 */ 5141 mb_mark_used(e4b, &ex); 5142 ext4_unlock_group(sb, group); 5143 ret = ext4_issue_discard(sb, group, start, count, NULL); 5144 ext4_lock_group(sb, group); 5145 mb_free_blocks(NULL, e4b, start, ex.fe_len); 5146 return ret; 5147 } 5148 5149 /** 5150 * ext4_trim_all_free -- function to trim all free space in alloc. group 5151 * @sb: super block for file system 5152 * @group: group to be trimmed 5153 * @start: first group block to examine 5154 * @max: last group block to examine 5155 * @minblocks: minimum extent block count 5156 * 5157 * ext4_trim_all_free walks through group's buddy bitmap searching for free 5158 * extents. When the free block is found, ext4_trim_extent is called to TRIM 5159 * the extent. 5160 * 5161 * 5162 * ext4_trim_all_free walks through group's block bitmap searching for free 5163 * extents. When the free extent is found, mark it as used in group buddy 5164 * bitmap. Then issue a TRIM command on this extent and free the extent in 5165 * the group buddy bitmap. This is done until whole group is scanned. 5166 */ 5167 static ext4_grpblk_t 5168 ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 5169 ext4_grpblk_t start, ext4_grpblk_t max, 5170 ext4_grpblk_t minblocks) 5171 { 5172 void *bitmap; 5173 ext4_grpblk_t next, count = 0, free_count = 0; 5174 struct ext4_buddy e4b; 5175 int ret = 0; 5176 5177 trace_ext4_trim_all_free(sb, group, start, max); 5178 5179 ret = ext4_mb_load_buddy(sb, group, &e4b); 5180 if (ret) { 5181 ext4_warning(sb, "Error %d loading buddy information for %u", 5182 ret, group); 5183 return ret; 5184 } 5185 bitmap = e4b.bd_bitmap; 5186 5187 ext4_lock_group(sb, group); 5188 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && 5189 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) 5190 goto out; 5191 5192 start = (e4b.bd_info->bb_first_free > start) ? 5193 e4b.bd_info->bb_first_free : start; 5194 5195 while (start <= max) { 5196 start = mb_find_next_zero_bit(bitmap, max + 1, start); 5197 if (start > max) 5198 break; 5199 next = mb_find_next_bit(bitmap, max + 1, start); 5200 5201 if ((next - start) >= minblocks) { 5202 ret = ext4_trim_extent(sb, start, 5203 next - start, group, &e4b); 5204 if (ret && ret != -EOPNOTSUPP) 5205 break; 5206 ret = 0; 5207 count += next - start; 5208 } 5209 free_count += next - start; 5210 start = next + 1; 5211 5212 if (fatal_signal_pending(current)) { 5213 count = -ERESTARTSYS; 5214 break; 5215 } 5216 5217 if (need_resched()) { 5218 ext4_unlock_group(sb, group); 5219 cond_resched(); 5220 ext4_lock_group(sb, group); 5221 } 5222 5223 if ((e4b.bd_info->bb_free - free_count) < minblocks) 5224 break; 5225 } 5226 5227 if (!ret) { 5228 ret = count; 5229 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); 5230 } 5231 out: 5232 ext4_unlock_group(sb, group); 5233 ext4_mb_unload_buddy(&e4b); 5234 5235 ext4_debug("trimmed %d blocks in the group %d\n", 5236 count, group); 5237 5238 return ret; 5239 } 5240 5241 /** 5242 * ext4_trim_fs() -- trim ioctl handle function 5243 * @sb: superblock for filesystem 5244 * @range: fstrim_range structure 5245 * 5246 * start: First Byte to trim 5247 * len: number of Bytes to trim from start 5248 * minlen: minimum extent length in Bytes 5249 * ext4_trim_fs goes through all allocation groups containing Bytes from 5250 * start to start+len. For each such a group ext4_trim_all_free function 5251 * is invoked to trim all free space. 5252 */ 5253 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 5254 { 5255 struct ext4_group_info *grp; 5256 ext4_group_t group, first_group, last_group; 5257 ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 5258 uint64_t start, end, minlen, trimmed = 0; 5259 ext4_fsblk_t first_data_blk = 5260 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 5261 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 5262 int ret = 0; 5263 5264 start = range->start >> sb->s_blocksize_bits; 5265 end = start + (range->len >> sb->s_blocksize_bits) - 1; 5266 minlen = EXT4_NUM_B2C(EXT4_SB(sb), 5267 range->minlen >> sb->s_blocksize_bits); 5268 5269 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || 5270 start >= max_blks || 5271 range->len < sb->s_blocksize) 5272 return -EINVAL; 5273 if (end >= max_blks) 5274 end = max_blks - 1; 5275 if (end <= first_data_blk) 5276 goto out; 5277 if (start < first_data_blk) 5278 start = first_data_blk; 5279 5280 /* Determine first and last group to examine based on start and end */ 5281 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 5282 &first_group, &first_cluster); 5283 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 5284 &last_group, &last_cluster); 5285 5286 /* end now represents the last cluster to discard in this group */ 5287 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 5288 5289 for (group = first_group; group <= last_group; group++) { 5290 grp = ext4_get_group_info(sb, group); 5291 /* We only do this if the grp has never been initialized */ 5292 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 5293 ret = ext4_mb_init_group(sb, group, GFP_NOFS); 5294 if (ret) 5295 break; 5296 } 5297 5298 /* 5299 * For all the groups except the last one, last cluster will 5300 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 5301 * change it for the last group, note that last_cluster is 5302 * already computed earlier by ext4_get_group_no_and_offset() 5303 */ 5304 if (group == last_group) 5305 end = last_cluster; 5306 5307 if (grp->bb_free >= minlen) { 5308 cnt = ext4_trim_all_free(sb, group, first_cluster, 5309 end, minlen); 5310 if (cnt < 0) { 5311 ret = cnt; 5312 break; 5313 } 5314 trimmed += cnt; 5315 } 5316 5317 /* 5318 * For every group except the first one, we are sure 5319 * that the first cluster to discard will be cluster #0. 5320 */ 5321 first_cluster = 0; 5322 } 5323 5324 if (!ret) 5325 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); 5326 5327 out: 5328 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; 5329 return ret; 5330 } 5331 5332 /* Iterate all the free extents in the group. */ 5333 int 5334 ext4_mballoc_query_range( 5335 struct super_block *sb, 5336 ext4_group_t group, 5337 ext4_grpblk_t start, 5338 ext4_grpblk_t end, 5339 ext4_mballoc_query_range_fn formatter, 5340 void *priv) 5341 { 5342 void *bitmap; 5343 ext4_grpblk_t next; 5344 struct ext4_buddy e4b; 5345 int error; 5346 5347 error = ext4_mb_load_buddy(sb, group, &e4b); 5348 if (error) 5349 return error; 5350 bitmap = e4b.bd_bitmap; 5351 5352 ext4_lock_group(sb, group); 5353 5354 start = (e4b.bd_info->bb_first_free > start) ? 5355 e4b.bd_info->bb_first_free : start; 5356 if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) 5357 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 5358 5359 while (start <= end) { 5360 start = mb_find_next_zero_bit(bitmap, end + 1, start); 5361 if (start > end) 5362 break; 5363 next = mb_find_next_bit(bitmap, end + 1, start); 5364 5365 ext4_unlock_group(sb, group); 5366 error = formatter(sb, group, start, next - start, priv); 5367 if (error) 5368 goto out_unload; 5369 ext4_lock_group(sb, group); 5370 5371 start = next + 1; 5372 } 5373 5374 ext4_unlock_group(sb, group); 5375 out_unload: 5376 ext4_mb_unload_buddy(&e4b); 5377 5378 return error; 5379 } 5380