1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * f2fs extent cache support 4 * 5 * Copyright (c) 2015 Motorola Mobility 6 * Copyright (c) 2015 Samsung Electronics 7 * Authors: Jaegeuk Kim <jaegeuk@kernel.org> 8 * Chao Yu <chao2.yu@samsung.com> 9 * 10 * block_age-based extent cache added by: 11 * Copyright (c) 2022 xiaomi Co., Ltd. 12 * http://www.xiaomi.com/ 13 */ 14 15 #include <linux/fs.h> 16 #include <linux/f2fs_fs.h> 17 18 #include "f2fs.h" 19 #include "node.h" 20 #include <trace/events/f2fs.h> 21 22 bool sanity_check_extent_cache(struct inode *inode, struct folio *ifolio) 23 { 24 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 25 struct f2fs_extent *i_ext = &F2FS_INODE(ifolio)->i_ext; 26 struct extent_info ei; 27 int devi; 28 29 get_read_extent_info(&ei, i_ext); 30 31 if (!ei.len) 32 return true; 33 34 if (!f2fs_is_valid_blkaddr(sbi, ei.blk, DATA_GENERIC_ENHANCE) || 35 !f2fs_is_valid_blkaddr(sbi, ei.blk + ei.len - 1, 36 DATA_GENERIC_ENHANCE)) { 37 f2fs_warn(sbi, "%s: inode (ino=%llx) extent info [%u, %u, %u] is incorrect, run fsck to fix", 38 __func__, inode->i_ino, 39 ei.blk, ei.fofs, ei.len); 40 return false; 41 } 42 43 if (!IS_DEVICE_ALIASING(inode)) 44 return true; 45 46 for (devi = 0; devi < sbi->s_ndevs; devi++) { 47 if (FDEV(devi).start_blk != ei.blk || 48 FDEV(devi).end_blk != ei.blk + ei.len - 1) 49 continue; 50 51 if (devi == 0) { 52 f2fs_warn(sbi, 53 "%s: inode (ino=%llx) is an alias of meta device", 54 __func__, inode->i_ino); 55 return false; 56 } 57 58 if (bdev_is_zoned(FDEV(devi).bdev)) { 59 f2fs_warn(sbi, 60 "%s: device alias inode (ino=%llx)'s extent info " 61 "[%u, %u, %u] maps to zoned block device", 62 __func__, inode->i_ino, ei.blk, ei.fofs, ei.len); 63 return false; 64 } 65 return true; 66 } 67 68 f2fs_warn(sbi, "%s: device alias inode (ino=%llx)'s extent info " 69 "[%u, %u, %u] is inconsistent w/ any devices", 70 __func__, inode->i_ino, ei.blk, ei.fofs, ei.len); 71 return false; 72 } 73 74 static void __set_extent_info(struct extent_info *ei, 75 unsigned int fofs, unsigned int len, 76 block_t blk, bool keep_clen, 77 unsigned long age, unsigned long last_blocks, 78 enum extent_type type) 79 { 80 ei->fofs = fofs; 81 ei->len = len; 82 83 if (type == EX_READ) { 84 ei->blk = blk; 85 if (keep_clen) 86 return; 87 #ifdef CONFIG_F2FS_FS_COMPRESSION 88 ei->c_len = 0; 89 #endif 90 } else if (type == EX_BLOCK_AGE) { 91 ei->age = age; 92 ei->last_blocks = last_blocks; 93 } 94 } 95 96 static bool __init_may_extent_tree(struct inode *inode, enum extent_type type) 97 { 98 if (type == EX_READ) 99 return test_opt(F2FS_I_SB(inode), READ_EXTENT_CACHE) && 100 S_ISREG(inode->i_mode); 101 if (type == EX_BLOCK_AGE) 102 return test_opt(F2FS_I_SB(inode), AGE_EXTENT_CACHE) && 103 (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)); 104 return false; 105 } 106 107 static bool __may_extent_tree(struct inode *inode, enum extent_type type) 108 { 109 if (IS_DEVICE_ALIASING(inode) && type == EX_READ) 110 return true; 111 112 /* 113 * for recovered files during mount do not create extents 114 * if shrinker is not registered. 115 */ 116 if (list_empty(&F2FS_I_SB(inode)->s_list)) 117 return false; 118 119 if (!__init_may_extent_tree(inode, type)) 120 return false; 121 122 if (type == EX_READ) { 123 if (is_inode_flag_set(inode, FI_NO_EXTENT)) 124 return false; 125 if (is_inode_flag_set(inode, FI_COMPRESSED_FILE) && 126 !f2fs_sb_has_readonly(F2FS_I_SB(inode))) 127 return false; 128 } else if (type == EX_BLOCK_AGE) { 129 if (is_inode_flag_set(inode, FI_COMPRESSED_FILE)) 130 return false; 131 if (file_is_cold(inode)) 132 return false; 133 } 134 return true; 135 } 136 137 static void __try_update_largest_extent(struct extent_tree *et, 138 struct extent_node *en) 139 { 140 if (et->type != EX_READ) 141 return; 142 if (en->ei.len <= et->largest.len) 143 return; 144 145 et->largest = en->ei; 146 et->largest_updated = true; 147 } 148 149 static bool __is_extent_mergeable(struct extent_info *back, 150 struct extent_info *front, enum extent_type type) 151 { 152 if (type == EX_READ) { 153 #ifdef CONFIG_F2FS_FS_COMPRESSION 154 if (back->c_len && back->len != back->c_len) 155 return false; 156 if (front->c_len && front->len != front->c_len) 157 return false; 158 #endif 159 return (back->fofs + back->len == front->fofs && 160 back->blk + back->len == front->blk); 161 } else if (type == EX_BLOCK_AGE) { 162 return (back->fofs + back->len == front->fofs && 163 abs(back->age - front->age) <= SAME_AGE_REGION && 164 abs(back->last_blocks - front->last_blocks) <= 165 SAME_AGE_REGION); 166 } 167 return false; 168 } 169 170 static bool __is_back_mergeable(struct extent_info *cur, 171 struct extent_info *back, enum extent_type type) 172 { 173 return __is_extent_mergeable(back, cur, type); 174 } 175 176 static bool __is_front_mergeable(struct extent_info *cur, 177 struct extent_info *front, enum extent_type type) 178 { 179 return __is_extent_mergeable(cur, front, type); 180 } 181 182 static struct extent_node *__lookup_extent_node(struct rb_root_cached *root, 183 struct extent_node *cached_en, unsigned int fofs) 184 { 185 struct rb_node *node = root->rb_root.rb_node; 186 struct extent_node *en; 187 188 /* check a cached entry */ 189 if (cached_en && cached_en->ei.fofs <= fofs && 190 cached_en->ei.fofs + cached_en->ei.len > fofs) 191 return cached_en; 192 193 /* check rb_tree */ 194 while (node) { 195 en = rb_entry(node, struct extent_node, rb_node); 196 197 if (fofs < en->ei.fofs) 198 node = node->rb_left; 199 else if (fofs >= en->ei.fofs + en->ei.len) 200 node = node->rb_right; 201 else 202 return en; 203 } 204 return NULL; 205 } 206 207 /* 208 * lookup rb entry in position of @fofs in rb-tree, 209 * if hit, return the entry, otherwise, return NULL 210 * @prev_ex: extent before fofs 211 * @next_ex: extent after fofs 212 * @insert_p: insert point for new extent at fofs 213 * in order to simplify the insertion after. 214 * tree must stay unchanged between lookup and insertion. 215 */ 216 static struct extent_node *__lookup_extent_node_ret(struct rb_root_cached *root, 217 struct extent_node *cached_en, 218 unsigned int fofs, 219 struct extent_node **prev_entry, 220 struct extent_node **next_entry, 221 struct rb_node ***insert_p, 222 struct rb_node **insert_parent, 223 bool *leftmost) 224 { 225 struct rb_node **pnode = &root->rb_root.rb_node; 226 struct rb_node *parent = NULL, *tmp_node; 227 struct extent_node *en = cached_en; 228 229 *insert_p = NULL; 230 *insert_parent = NULL; 231 *prev_entry = NULL; 232 *next_entry = NULL; 233 234 if (RB_EMPTY_ROOT(&root->rb_root)) 235 return NULL; 236 237 if (en && en->ei.fofs <= fofs && en->ei.fofs + en->ei.len > fofs) 238 goto lookup_neighbors; 239 240 *leftmost = true; 241 242 while (*pnode) { 243 parent = *pnode; 244 en = rb_entry(*pnode, struct extent_node, rb_node); 245 246 if (fofs < en->ei.fofs) { 247 pnode = &(*pnode)->rb_left; 248 } else if (fofs >= en->ei.fofs + en->ei.len) { 249 pnode = &(*pnode)->rb_right; 250 *leftmost = false; 251 } else { 252 goto lookup_neighbors; 253 } 254 } 255 256 *insert_p = pnode; 257 *insert_parent = parent; 258 259 en = rb_entry(parent, struct extent_node, rb_node); 260 tmp_node = parent; 261 if (parent && fofs > en->ei.fofs) 262 tmp_node = rb_next(parent); 263 *next_entry = rb_entry_safe(tmp_node, struct extent_node, rb_node); 264 265 tmp_node = parent; 266 if (parent && fofs < en->ei.fofs) 267 tmp_node = rb_prev(parent); 268 *prev_entry = rb_entry_safe(tmp_node, struct extent_node, rb_node); 269 return NULL; 270 271 lookup_neighbors: 272 if (fofs == en->ei.fofs) { 273 /* lookup prev node for merging backward later */ 274 tmp_node = rb_prev(&en->rb_node); 275 *prev_entry = rb_entry_safe(tmp_node, 276 struct extent_node, rb_node); 277 } 278 if (fofs == en->ei.fofs + en->ei.len - 1) { 279 /* lookup next node for merging frontward later */ 280 tmp_node = rb_next(&en->rb_node); 281 *next_entry = rb_entry_safe(tmp_node, 282 struct extent_node, rb_node); 283 } 284 return en; 285 } 286 287 static struct kmem_cache *extent_tree_slab; 288 static struct kmem_cache *extent_node_slab; 289 290 static struct extent_node *__attach_extent_node(struct f2fs_sb_info *sbi, 291 struct extent_tree *et, struct extent_info *ei, 292 struct rb_node *parent, struct rb_node **p, 293 bool leftmost) 294 { 295 struct extent_tree_info *eti = &sbi->extent_tree[et->type]; 296 struct extent_node *en; 297 298 en = f2fs_kmem_cache_alloc(extent_node_slab, GFP_ATOMIC, false, sbi); 299 if (!en) 300 return NULL; 301 302 en->ei = *ei; 303 INIT_LIST_HEAD(&en->list); 304 en->et = et; 305 306 rb_link_node(&en->rb_node, parent, p); 307 rb_insert_color_cached(&en->rb_node, &et->root, leftmost); 308 atomic_inc(&et->node_cnt); 309 atomic_inc(&eti->total_ext_node); 310 return en; 311 } 312 313 static void __detach_extent_node(struct f2fs_sb_info *sbi, 314 struct extent_tree *et, struct extent_node *en) 315 { 316 struct extent_tree_info *eti = &sbi->extent_tree[et->type]; 317 318 rb_erase_cached(&en->rb_node, &et->root); 319 atomic_dec(&et->node_cnt); 320 atomic_dec(&eti->total_ext_node); 321 322 if (et->cached_en == en) 323 et->cached_en = NULL; 324 kmem_cache_free(extent_node_slab, en); 325 } 326 327 /* 328 * Flow to release an extent_node: 329 * 1. list_del_init 330 * 2. __detach_extent_node 331 * 3. kmem_cache_free. 332 */ 333 static void __release_extent_node(struct f2fs_sb_info *sbi, 334 struct extent_tree *et, struct extent_node *en) 335 { 336 struct extent_tree_info *eti = &sbi->extent_tree[et->type]; 337 338 spin_lock(&eti->extent_lock); 339 f2fs_bug_on(sbi, list_empty(&en->list)); 340 list_del_init(&en->list); 341 spin_unlock(&eti->extent_lock); 342 343 __detach_extent_node(sbi, et, en); 344 } 345 346 static struct extent_tree *__grab_extent_tree(struct inode *inode, 347 enum extent_type type) 348 { 349 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 350 struct extent_tree_info *eti = &sbi->extent_tree[type]; 351 struct extent_tree *et; 352 nid_t ino = inode->i_ino; 353 354 mutex_lock(&eti->extent_tree_lock); 355 et = radix_tree_lookup(&eti->extent_tree_root, ino); 356 if (!et) { 357 et = f2fs_kmem_cache_alloc(extent_tree_slab, 358 GFP_NOFS, true, NULL); 359 f2fs_radix_tree_insert(&eti->extent_tree_root, ino, et); 360 memset(et, 0, sizeof(struct extent_tree)); 361 et->ino = ino; 362 et->type = type; 363 et->root = RB_ROOT_CACHED; 364 et->cached_en = NULL; 365 rwlock_init(&et->lock); 366 INIT_LIST_HEAD(&et->list); 367 atomic_set(&et->node_cnt, 0); 368 atomic_inc(&eti->total_ext_tree); 369 } else { 370 atomic_dec(&eti->total_zombie_tree); 371 list_del_init(&et->list); 372 } 373 mutex_unlock(&eti->extent_tree_lock); 374 375 /* never died until evict_inode */ 376 F2FS_I(inode)->extent_tree[type] = et; 377 378 return et; 379 } 380 381 static unsigned int __free_extent_tree(struct f2fs_sb_info *sbi, 382 struct extent_tree *et, unsigned int nr_shrink) 383 { 384 struct rb_node *node, *next; 385 struct extent_node *en; 386 unsigned int count; 387 388 node = rb_first_cached(&et->root); 389 390 for (count = 0; node && count < nr_shrink; count++) { 391 next = rb_next(node); 392 en = rb_entry(node, struct extent_node, rb_node); 393 __release_extent_node(sbi, et, en); 394 node = next; 395 } 396 397 return count; 398 } 399 400 static void __drop_largest_extent(struct extent_tree *et, 401 pgoff_t fofs, unsigned int len) 402 { 403 if (fofs < (pgoff_t)et->largest.fofs + et->largest.len && 404 fofs + len > et->largest.fofs) { 405 et->largest.len = 0; 406 et->largest_updated = true; 407 } 408 } 409 410 void f2fs_init_read_extent_tree(struct inode *inode, struct folio *ifolio) 411 { 412 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 413 struct extent_tree_info *eti = &sbi->extent_tree[EX_READ]; 414 struct f2fs_extent *i_ext = &F2FS_INODE(ifolio)->i_ext; 415 struct extent_tree *et; 416 struct extent_node *en; 417 struct extent_info ei = {0}; 418 419 if (!__may_extent_tree(inode, EX_READ)) { 420 /* drop largest read extent */ 421 if (i_ext->len) { 422 f2fs_folio_wait_writeback(ifolio, NODE, true, true); 423 i_ext->len = 0; 424 folio_mark_dirty(ifolio); 425 } 426 set_inode_flag(inode, FI_NO_EXTENT); 427 return; 428 } 429 430 et = __grab_extent_tree(inode, EX_READ); 431 432 get_read_extent_info(&ei, i_ext); 433 434 write_lock(&et->lock); 435 if (atomic_read(&et->node_cnt) || !ei.len) 436 goto skip; 437 438 if (IS_DEVICE_ALIASING(inode)) { 439 et->largest = ei; 440 goto skip; 441 } 442 443 en = __attach_extent_node(sbi, et, &ei, NULL, 444 &et->root.rb_root.rb_node, true); 445 if (en) { 446 et->largest = en->ei; 447 et->cached_en = en; 448 449 spin_lock(&eti->extent_lock); 450 list_add_tail(&en->list, &eti->extent_list); 451 spin_unlock(&eti->extent_lock); 452 } 453 skip: 454 /* Let's drop, if checkpoint got corrupted. */ 455 if (f2fs_cp_error(sbi)) { 456 et->largest.len = 0; 457 et->largest_updated = true; 458 } 459 write_unlock(&et->lock); 460 } 461 462 void f2fs_init_age_extent_tree(struct inode *inode) 463 { 464 if (!__init_may_extent_tree(inode, EX_BLOCK_AGE)) 465 return; 466 __grab_extent_tree(inode, EX_BLOCK_AGE); 467 } 468 469 void f2fs_init_extent_tree(struct inode *inode) 470 { 471 /* initialize read cache */ 472 if (__init_may_extent_tree(inode, EX_READ)) 473 __grab_extent_tree(inode, EX_READ); 474 475 /* initialize block age cache */ 476 if (__init_may_extent_tree(inode, EX_BLOCK_AGE)) 477 __grab_extent_tree(inode, EX_BLOCK_AGE); 478 } 479 480 static bool __lookup_extent_tree(struct inode *inode, pgoff_t pgofs, 481 struct extent_info *ei, enum extent_type type) 482 { 483 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 484 struct extent_tree_info *eti = &sbi->extent_tree[type]; 485 struct extent_tree *et = F2FS_I(inode)->extent_tree[type]; 486 struct extent_node *en; 487 bool ret = false; 488 489 if (!et) 490 return false; 491 492 trace_f2fs_lookup_extent_tree_start(inode, pgofs, type); 493 494 read_lock(&et->lock); 495 496 if (type == EX_READ && 497 et->largest.fofs <= pgofs && 498 (pgoff_t)et->largest.fofs + et->largest.len > pgofs) { 499 *ei = et->largest; 500 ret = true; 501 stat_inc_largest_node_hit(sbi); 502 goto out; 503 } 504 505 if (IS_DEVICE_ALIASING(inode)) { 506 ret = false; 507 goto out; 508 } 509 510 en = __lookup_extent_node(&et->root, et->cached_en, pgofs); 511 if (!en) 512 goto out; 513 514 if (en == et->cached_en) 515 stat_inc_cached_node_hit(sbi, type); 516 else 517 stat_inc_rbtree_node_hit(sbi, type); 518 519 *ei = en->ei; 520 spin_lock(&eti->extent_lock); 521 if (!list_empty(&en->list)) { 522 list_move_tail(&en->list, &eti->extent_list); 523 et->cached_en = en; 524 } 525 spin_unlock(&eti->extent_lock); 526 ret = true; 527 out: 528 stat_inc_total_hit(sbi, type); 529 read_unlock(&et->lock); 530 531 if (type == EX_READ) 532 trace_f2fs_lookup_read_extent_tree_end(inode, pgofs, ei); 533 else if (type == EX_BLOCK_AGE) 534 trace_f2fs_lookup_age_extent_tree_end(inode, pgofs, ei); 535 return ret; 536 } 537 538 static struct extent_node *__try_merge_extent_node(struct f2fs_sb_info *sbi, 539 struct extent_tree *et, struct extent_info *ei, 540 struct extent_node *prev_ex, 541 struct extent_node *next_ex) 542 { 543 struct extent_tree_info *eti = &sbi->extent_tree[et->type]; 544 struct extent_node *en = NULL; 545 546 if (prev_ex && __is_back_mergeable(ei, &prev_ex->ei, et->type)) { 547 prev_ex->ei.len += ei->len; 548 ei = &prev_ex->ei; 549 en = prev_ex; 550 } 551 552 if (next_ex && __is_front_mergeable(ei, &next_ex->ei, et->type)) { 553 next_ex->ei.fofs = ei->fofs; 554 next_ex->ei.len += ei->len; 555 if (et->type == EX_READ) 556 next_ex->ei.blk = ei->blk; 557 if (en) 558 __release_extent_node(sbi, et, prev_ex); 559 560 en = next_ex; 561 } 562 563 if (!en) 564 return NULL; 565 566 __try_update_largest_extent(et, en); 567 568 spin_lock(&eti->extent_lock); 569 if (!list_empty(&en->list)) { 570 list_move_tail(&en->list, &eti->extent_list); 571 et->cached_en = en; 572 } 573 spin_unlock(&eti->extent_lock); 574 return en; 575 } 576 577 static struct extent_node *__insert_extent_tree(struct f2fs_sb_info *sbi, 578 struct extent_tree *et, struct extent_info *ei, 579 struct rb_node **insert_p, 580 struct rb_node *insert_parent, 581 bool leftmost) 582 { 583 struct extent_tree_info *eti = &sbi->extent_tree[et->type]; 584 struct rb_node **p = &et->root.rb_root.rb_node; 585 struct rb_node *parent = NULL; 586 struct extent_node *en = NULL; 587 588 if (insert_p && insert_parent) { 589 parent = insert_parent; 590 p = insert_p; 591 goto do_insert; 592 } 593 594 leftmost = true; 595 596 /* look up extent_node in the rb tree */ 597 while (*p) { 598 parent = *p; 599 en = rb_entry(parent, struct extent_node, rb_node); 600 601 if (ei->fofs < en->ei.fofs) { 602 p = &(*p)->rb_left; 603 } else if (ei->fofs >= en->ei.fofs + en->ei.len) { 604 p = &(*p)->rb_right; 605 leftmost = false; 606 } else { 607 f2fs_err_ratelimited(sbi, "%s: corrupted extent, type: %d, " 608 "extent node in rb tree [%u, %u, %u], age [%llu, %llu], " 609 "extent node to insert [%u, %u, %u], age [%llu, %llu]", 610 __func__, et->type, en->ei.fofs, en->ei.blk, en->ei.len, en->ei.age, 611 en->ei.last_blocks, ei->fofs, ei->blk, ei->len, ei->age, ei->last_blocks); 612 f2fs_bug_on(sbi, 1); 613 return NULL; 614 } 615 } 616 617 do_insert: 618 en = __attach_extent_node(sbi, et, ei, parent, p, leftmost); 619 if (!en) 620 return NULL; 621 622 __try_update_largest_extent(et, en); 623 624 /* update in global extent list */ 625 spin_lock(&eti->extent_lock); 626 list_add_tail(&en->list, &eti->extent_list); 627 et->cached_en = en; 628 spin_unlock(&eti->extent_lock); 629 return en; 630 } 631 632 static unsigned int __destroy_extent_node(struct inode *inode, 633 enum extent_type type) 634 { 635 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 636 struct extent_tree *et = F2FS_I(inode)->extent_tree[type]; 637 unsigned int nr_shrink = type == EX_READ ? 638 READ_EXTENT_CACHE_SHRINK_NUMBER : 639 AGE_EXTENT_CACHE_SHRINK_NUMBER; 640 unsigned int node_cnt = 0; 641 642 if (!et || !atomic_read(&et->node_cnt)) 643 return 0; 644 645 while (atomic_read(&et->node_cnt)) { 646 write_lock(&et->lock); 647 node_cnt += __free_extent_tree(sbi, et, nr_shrink); 648 write_unlock(&et->lock); 649 } 650 651 return node_cnt; 652 } 653 654 static void __update_extent_tree_range(struct inode *inode, 655 struct extent_info *tei, enum extent_type type) 656 { 657 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 658 struct extent_tree *et = F2FS_I(inode)->extent_tree[type]; 659 struct extent_node *en = NULL, *en1 = NULL; 660 struct extent_node *prev_en = NULL, *next_en = NULL; 661 struct extent_info ei, dei, prev; 662 struct rb_node **insert_p = NULL, *insert_parent = NULL; 663 unsigned int fofs = tei->fofs, len = tei->len; 664 unsigned int end = fofs + len; 665 bool updated = false; 666 bool leftmost = false; 667 668 if (!et) 669 return; 670 671 if (unlikely(len == 0)) { 672 f2fs_err_ratelimited(sbi, "%s: extent len is zero, type: %d, " 673 "extent [%u, %u, %u], age [%llu, %llu]", 674 __func__, type, tei->fofs, tei->blk, tei->len, 675 tei->age, tei->last_blocks); 676 f2fs_bug_on(sbi, 1); 677 return; 678 } 679 680 if (type == EX_READ) 681 trace_f2fs_update_read_extent_tree_range(inode, fofs, len, 682 tei->blk, 0); 683 else if (type == EX_BLOCK_AGE) 684 trace_f2fs_update_age_extent_tree_range(inode, fofs, len, 685 tei->age, tei->last_blocks); 686 687 write_lock(&et->lock); 688 689 if (type == EX_READ) { 690 if (is_inode_flag_set(inode, FI_NO_EXTENT)) { 691 write_unlock(&et->lock); 692 return; 693 } 694 695 prev = et->largest; 696 dei.len = 0; 697 698 /* 699 * drop largest extent before lookup, in case it's already 700 * been shrunk from extent tree 701 */ 702 __drop_largest_extent(et, fofs, len); 703 } 704 705 /* 1. lookup first extent node in range [fofs, fofs + len - 1] */ 706 en = __lookup_extent_node_ret(&et->root, 707 et->cached_en, fofs, 708 &prev_en, &next_en, 709 &insert_p, &insert_parent, 710 &leftmost); 711 if (!en) 712 en = next_en; 713 714 /* 2. invalidate all extent nodes in range [fofs, fofs + len - 1] */ 715 while (en && en->ei.fofs < end) { 716 unsigned int org_end; 717 int parts = 0; /* # of parts current extent split into */ 718 719 next_en = en1 = NULL; 720 721 dei = en->ei; 722 org_end = dei.fofs + dei.len; 723 f2fs_bug_on(sbi, fofs >= org_end); 724 725 if (fofs > dei.fofs && (type != EX_READ || 726 fofs - dei.fofs >= F2FS_MIN_EXTENT_LEN)) { 727 en->ei.len = fofs - en->ei.fofs; 728 prev_en = en; 729 parts = 1; 730 } 731 732 if (end < org_end && (type != EX_READ || 733 (org_end - end >= F2FS_MIN_EXTENT_LEN && 734 atomic_read(&et->node_cnt) < 735 sbi->max_read_extent_count))) { 736 if (parts) { 737 __set_extent_info(&ei, 738 end, org_end - end, 739 end - dei.fofs + dei.blk, false, 740 dei.age, dei.last_blocks, 741 type); 742 en1 = __insert_extent_tree(sbi, et, &ei, 743 NULL, NULL, true); 744 next_en = en1; 745 } else { 746 __set_extent_info(&en->ei, 747 end, en->ei.len - (end - dei.fofs), 748 en->ei.blk + (end - dei.fofs), true, 749 dei.age, dei.last_blocks, 750 type); 751 next_en = en; 752 } 753 parts++; 754 } 755 756 if (!next_en) { 757 struct rb_node *node = rb_next(&en->rb_node); 758 759 next_en = rb_entry_safe(node, struct extent_node, 760 rb_node); 761 } 762 763 if (parts) 764 __try_update_largest_extent(et, en); 765 else 766 __release_extent_node(sbi, et, en); 767 768 /* 769 * if original extent is split into zero or two parts, extent 770 * tree has been altered by deletion or insertion, therefore 771 * invalidate pointers regard to tree. 772 */ 773 if (parts != 1) { 774 insert_p = NULL; 775 insert_parent = NULL; 776 } 777 en = next_en; 778 } 779 780 if (type == EX_BLOCK_AGE) 781 goto update_age_extent_cache; 782 783 /* 3. update extent in read extent cache */ 784 BUG_ON(type != EX_READ); 785 786 if (tei->blk) { 787 __set_extent_info(&ei, fofs, len, tei->blk, false, 788 0, 0, EX_READ); 789 if (!__try_merge_extent_node(sbi, et, &ei, prev_en, next_en)) 790 __insert_extent_tree(sbi, et, &ei, 791 insert_p, insert_parent, leftmost); 792 793 /* give up extent_cache, if split and small updates happen */ 794 if (dei.len >= 1 && 795 prev.len < F2FS_MIN_EXTENT_LEN && 796 et->largest.len < F2FS_MIN_EXTENT_LEN) { 797 et->largest.len = 0; 798 et->largest_updated = true; 799 set_inode_flag(inode, FI_NO_EXTENT); 800 } 801 } 802 803 if (et->largest_updated) { 804 et->largest_updated = false; 805 updated = true; 806 } 807 goto out_read_extent_cache; 808 update_age_extent_cache: 809 if (tei->last_blocks == F2FS_EXTENT_AGE_INVALID) 810 goto out_read_extent_cache; 811 812 __set_extent_info(&ei, fofs, len, 0, false, 813 tei->age, tei->last_blocks, EX_BLOCK_AGE); 814 if (!__try_merge_extent_node(sbi, et, &ei, prev_en, next_en)) 815 __insert_extent_tree(sbi, et, &ei, 816 insert_p, insert_parent, leftmost); 817 out_read_extent_cache: 818 write_unlock(&et->lock); 819 820 if (is_inode_flag_set(inode, FI_NO_EXTENT)) 821 __destroy_extent_node(inode, EX_READ); 822 823 if (updated) 824 f2fs_mark_inode_dirty_sync(inode, true); 825 } 826 827 #ifdef CONFIG_F2FS_FS_COMPRESSION 828 void f2fs_update_read_extent_tree_range_compressed(struct inode *inode, 829 pgoff_t fofs, block_t blkaddr, unsigned int llen, 830 unsigned int c_len) 831 { 832 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 833 struct extent_tree *et = F2FS_I(inode)->extent_tree[EX_READ]; 834 struct extent_node *en = NULL; 835 struct extent_node *prev_en = NULL, *next_en = NULL; 836 struct extent_info ei; 837 struct rb_node **insert_p = NULL, *insert_parent = NULL; 838 bool leftmost = false; 839 840 trace_f2fs_update_read_extent_tree_range(inode, fofs, llen, 841 blkaddr, c_len); 842 843 /* it is safe here to check FI_NO_EXTENT w/o et->lock in ro image */ 844 if (is_inode_flag_set(inode, FI_NO_EXTENT)) 845 return; 846 847 write_lock(&et->lock); 848 849 en = __lookup_extent_node_ret(&et->root, 850 et->cached_en, fofs, 851 &prev_en, &next_en, 852 &insert_p, &insert_parent, 853 &leftmost); 854 if (en) 855 goto unlock_out; 856 857 __set_extent_info(&ei, fofs, llen, blkaddr, true, 0, 0, EX_READ); 858 ei.c_len = c_len; 859 860 if (!__try_merge_extent_node(sbi, et, &ei, prev_en, next_en)) 861 __insert_extent_tree(sbi, et, &ei, 862 insert_p, insert_parent, leftmost); 863 unlock_out: 864 write_unlock(&et->lock); 865 } 866 #endif 867 868 static unsigned long long __calculate_block_age(struct f2fs_sb_info *sbi, 869 unsigned long long new, 870 unsigned long long old) 871 { 872 unsigned int rem_old, rem_new; 873 unsigned long long res; 874 unsigned int weight = sbi->last_age_weight; 875 876 res = div_u64_rem(new, 100, &rem_new) * (100 - weight) 877 + div_u64_rem(old, 100, &rem_old) * weight; 878 879 if (rem_new) 880 res += rem_new * (100 - weight) / 100; 881 if (rem_old) 882 res += rem_old * weight / 100; 883 884 return res; 885 } 886 887 /* This returns a new age and allocated blocks in ei */ 888 static int __get_new_block_age(struct inode *inode, struct extent_info *ei, 889 block_t blkaddr) 890 { 891 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 892 loff_t f_size = i_size_read(inode); 893 unsigned long long cur_blocks = 894 atomic64_read(&sbi->allocated_data_blocks); 895 struct extent_info tei = *ei; /* only fofs and len are valid */ 896 897 /* 898 * When I/O is not aligned to a PAGE_SIZE, update will happen to the last 899 * file block even in seq write. So don't record age for newly last file 900 * block here. 901 */ 902 if ((f_size >> PAGE_SHIFT) == ei->fofs && f_size & (PAGE_SIZE - 1) && 903 blkaddr == NEW_ADDR) 904 return -EINVAL; 905 906 if (__lookup_extent_tree(inode, ei->fofs, &tei, EX_BLOCK_AGE)) { 907 unsigned long long cur_age; 908 909 if (cur_blocks >= tei.last_blocks) 910 cur_age = cur_blocks - tei.last_blocks; 911 else 912 /* allocated_data_blocks overflow */ 913 cur_age = (ULLONG_MAX - 1) - tei.last_blocks + cur_blocks; 914 915 if (tei.age) 916 ei->age = __calculate_block_age(sbi, cur_age, tei.age); 917 else 918 ei->age = cur_age; 919 ei->last_blocks = cur_blocks; 920 WARN_ON(ei->age > cur_blocks); 921 return 0; 922 } 923 924 f2fs_bug_on(sbi, blkaddr == NULL_ADDR); 925 926 /* the data block was allocated for the first time */ 927 if (blkaddr == NEW_ADDR) 928 goto out; 929 930 if (__is_valid_data_blkaddr(blkaddr) && 931 !f2fs_is_valid_blkaddr(sbi, blkaddr, DATA_GENERIC_ENHANCE)) 932 return -EINVAL; 933 out: 934 /* 935 * init block age with zero, this can happen when the block age extent 936 * was reclaimed due to memory constraint or system reboot 937 */ 938 ei->age = 0; 939 ei->last_blocks = cur_blocks; 940 return 0; 941 } 942 943 static void __update_extent_cache(struct dnode_of_data *dn, enum extent_type type) 944 { 945 struct extent_info ei = {}; 946 947 if (!__may_extent_tree(dn->inode, type)) 948 return; 949 950 ei.fofs = f2fs_start_bidx_of_node(ofs_of_node(dn->node_folio), dn->inode) + 951 dn->ofs_in_node; 952 ei.len = 1; 953 954 if (type == EX_READ) { 955 if (dn->data_blkaddr == NEW_ADDR) 956 ei.blk = NULL_ADDR; 957 else 958 ei.blk = dn->data_blkaddr; 959 } else if (type == EX_BLOCK_AGE) { 960 if (__get_new_block_age(dn->inode, &ei, dn->data_blkaddr)) 961 return; 962 } 963 __update_extent_tree_range(dn->inode, &ei, type); 964 } 965 966 static unsigned int __shrink_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink, 967 enum extent_type type) 968 { 969 struct extent_tree_info *eti = &sbi->extent_tree[type]; 970 struct extent_tree *et, *next; 971 struct extent_node *en; 972 unsigned int node_cnt = 0, tree_cnt = 0; 973 int remained; 974 975 if (!atomic_read(&eti->total_zombie_tree)) 976 goto free_node; 977 978 if (!mutex_trylock(&eti->extent_tree_lock)) 979 goto out; 980 981 /* 1. remove unreferenced extent tree */ 982 list_for_each_entry_safe(et, next, &eti->zombie_list, list) { 983 if (atomic_read(&et->node_cnt)) { 984 write_lock(&et->lock); 985 node_cnt += __free_extent_tree(sbi, et, 986 nr_shrink - node_cnt - tree_cnt); 987 write_unlock(&et->lock); 988 } 989 990 if (atomic_read(&et->node_cnt)) 991 goto unlock_out; 992 993 list_del_init(&et->list); 994 radix_tree_delete(&eti->extent_tree_root, et->ino); 995 kmem_cache_free(extent_tree_slab, et); 996 atomic_dec(&eti->total_ext_tree); 997 atomic_dec(&eti->total_zombie_tree); 998 tree_cnt++; 999 1000 if (node_cnt + tree_cnt >= nr_shrink) 1001 goto unlock_out; 1002 cond_resched(); 1003 } 1004 mutex_unlock(&eti->extent_tree_lock); 1005 1006 free_node: 1007 /* 2. remove LRU extent entries */ 1008 if (!mutex_trylock(&eti->extent_tree_lock)) 1009 goto out; 1010 1011 remained = nr_shrink - (node_cnt + tree_cnt); 1012 1013 spin_lock(&eti->extent_lock); 1014 for (; remained > 0; remained--) { 1015 if (list_empty(&eti->extent_list)) 1016 break; 1017 en = list_first_entry(&eti->extent_list, 1018 struct extent_node, list); 1019 et = en->et; 1020 if (!write_trylock(&et->lock)) { 1021 /* refresh this extent node's position in extent list */ 1022 list_move_tail(&en->list, &eti->extent_list); 1023 continue; 1024 } 1025 1026 list_del_init(&en->list); 1027 spin_unlock(&eti->extent_lock); 1028 1029 __detach_extent_node(sbi, et, en); 1030 1031 write_unlock(&et->lock); 1032 node_cnt++; 1033 spin_lock(&eti->extent_lock); 1034 } 1035 spin_unlock(&eti->extent_lock); 1036 1037 unlock_out: 1038 mutex_unlock(&eti->extent_tree_lock); 1039 out: 1040 trace_f2fs_shrink_extent_tree(sbi, node_cnt, tree_cnt, type); 1041 1042 return node_cnt + tree_cnt; 1043 } 1044 1045 /* read extent cache operations */ 1046 bool f2fs_lookup_read_extent_cache(struct inode *inode, pgoff_t pgofs, 1047 struct extent_info *ei) 1048 { 1049 if (!__may_extent_tree(inode, EX_READ)) 1050 return false; 1051 1052 return __lookup_extent_tree(inode, pgofs, ei, EX_READ); 1053 } 1054 1055 bool f2fs_lookup_read_extent_cache_block(struct inode *inode, pgoff_t index, 1056 block_t *blkaddr) 1057 { 1058 struct extent_info ei = {}; 1059 1060 if (!f2fs_lookup_read_extent_cache(inode, index, &ei)) 1061 return false; 1062 *blkaddr = ei.blk + index - ei.fofs; 1063 return true; 1064 } 1065 1066 void f2fs_update_read_extent_cache(struct dnode_of_data *dn) 1067 { 1068 return __update_extent_cache(dn, EX_READ); 1069 } 1070 1071 void f2fs_update_read_extent_cache_range(struct dnode_of_data *dn, 1072 pgoff_t fofs, block_t blkaddr, unsigned int len) 1073 { 1074 struct extent_info ei = { 1075 .fofs = fofs, 1076 .len = len, 1077 .blk = blkaddr, 1078 }; 1079 1080 if (!__may_extent_tree(dn->inode, EX_READ)) 1081 return; 1082 1083 __update_extent_tree_range(dn->inode, &ei, EX_READ); 1084 } 1085 1086 unsigned int f2fs_shrink_read_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink) 1087 { 1088 if (!test_opt(sbi, READ_EXTENT_CACHE)) 1089 return 0; 1090 1091 return __shrink_extent_tree(sbi, nr_shrink, EX_READ); 1092 } 1093 1094 /* block age extent cache operations */ 1095 bool f2fs_lookup_age_extent_cache(struct inode *inode, pgoff_t pgofs, 1096 struct extent_info *ei) 1097 { 1098 if (!__may_extent_tree(inode, EX_BLOCK_AGE)) 1099 return false; 1100 1101 return __lookup_extent_tree(inode, pgofs, ei, EX_BLOCK_AGE); 1102 } 1103 1104 void f2fs_update_age_extent_cache(struct dnode_of_data *dn) 1105 { 1106 return __update_extent_cache(dn, EX_BLOCK_AGE); 1107 } 1108 1109 void f2fs_update_age_extent_cache_range(struct dnode_of_data *dn, 1110 pgoff_t fofs, unsigned int len) 1111 { 1112 struct extent_info ei = { 1113 .fofs = fofs, 1114 .len = len, 1115 .last_blocks = F2FS_EXTENT_AGE_INVALID, 1116 }; 1117 1118 if (!__may_extent_tree(dn->inode, EX_BLOCK_AGE)) 1119 return; 1120 1121 __update_extent_tree_range(dn->inode, &ei, EX_BLOCK_AGE); 1122 } 1123 1124 unsigned int f2fs_shrink_age_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink) 1125 { 1126 if (!test_opt(sbi, AGE_EXTENT_CACHE)) 1127 return 0; 1128 1129 return __shrink_extent_tree(sbi, nr_shrink, EX_BLOCK_AGE); 1130 } 1131 1132 void f2fs_destroy_extent_node(struct inode *inode) 1133 { 1134 __destroy_extent_node(inode, EX_READ); 1135 __destroy_extent_node(inode, EX_BLOCK_AGE); 1136 } 1137 1138 static void __drop_extent_tree(struct inode *inode, enum extent_type type) 1139 { 1140 struct extent_tree *et = F2FS_I(inode)->extent_tree[type]; 1141 bool updated = false; 1142 1143 if (!__may_extent_tree(inode, type)) 1144 return; 1145 1146 write_lock(&et->lock); 1147 if (type == EX_READ) { 1148 set_inode_flag(inode, FI_NO_EXTENT); 1149 if (et->largest.len) { 1150 et->largest.len = 0; 1151 updated = true; 1152 } 1153 } 1154 write_unlock(&et->lock); 1155 1156 __destroy_extent_node(inode, type); 1157 1158 if (updated) 1159 f2fs_mark_inode_dirty_sync(inode, true); 1160 } 1161 1162 void f2fs_drop_extent_tree(struct inode *inode) 1163 { 1164 __drop_extent_tree(inode, EX_READ); 1165 __drop_extent_tree(inode, EX_BLOCK_AGE); 1166 } 1167 1168 static void __destroy_extent_tree(struct inode *inode, enum extent_type type) 1169 { 1170 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 1171 struct extent_tree_info *eti = &sbi->extent_tree[type]; 1172 struct extent_tree *et = F2FS_I(inode)->extent_tree[type]; 1173 unsigned int node_cnt = 0; 1174 1175 if (!et) 1176 return; 1177 1178 if (inode->i_nlink && !is_bad_inode(inode) && 1179 atomic_read(&et->node_cnt)) { 1180 mutex_lock(&eti->extent_tree_lock); 1181 list_add_tail(&et->list, &eti->zombie_list); 1182 atomic_inc(&eti->total_zombie_tree); 1183 mutex_unlock(&eti->extent_tree_lock); 1184 return; 1185 } 1186 1187 /* free all extent info belong to this extent tree */ 1188 node_cnt = __destroy_extent_node(inode, type); 1189 1190 /* delete extent tree entry in radix tree */ 1191 mutex_lock(&eti->extent_tree_lock); 1192 f2fs_bug_on(sbi, atomic_read(&et->node_cnt)); 1193 radix_tree_delete(&eti->extent_tree_root, inode->i_ino); 1194 kmem_cache_free(extent_tree_slab, et); 1195 atomic_dec(&eti->total_ext_tree); 1196 mutex_unlock(&eti->extent_tree_lock); 1197 1198 F2FS_I(inode)->extent_tree[type] = NULL; 1199 1200 trace_f2fs_destroy_extent_tree(inode, node_cnt, type); 1201 } 1202 1203 void f2fs_destroy_extent_tree(struct inode *inode) 1204 { 1205 __destroy_extent_tree(inode, EX_READ); 1206 __destroy_extent_tree(inode, EX_BLOCK_AGE); 1207 } 1208 1209 static void __init_extent_tree_info(struct extent_tree_info *eti) 1210 { 1211 INIT_RADIX_TREE(&eti->extent_tree_root, GFP_NOIO); 1212 mutex_init(&eti->extent_tree_lock); 1213 INIT_LIST_HEAD(&eti->extent_list); 1214 spin_lock_init(&eti->extent_lock); 1215 atomic_set(&eti->total_ext_tree, 0); 1216 INIT_LIST_HEAD(&eti->zombie_list); 1217 atomic_set(&eti->total_zombie_tree, 0); 1218 atomic_set(&eti->total_ext_node, 0); 1219 } 1220 1221 void f2fs_init_extent_cache_info(struct f2fs_sb_info *sbi) 1222 { 1223 __init_extent_tree_info(&sbi->extent_tree[EX_READ]); 1224 __init_extent_tree_info(&sbi->extent_tree[EX_BLOCK_AGE]); 1225 1226 /* initialize for block age extents */ 1227 atomic64_set(&sbi->allocated_data_blocks, 0); 1228 sbi->hot_data_age_threshold = DEF_HOT_DATA_AGE_THRESHOLD; 1229 sbi->warm_data_age_threshold = DEF_WARM_DATA_AGE_THRESHOLD; 1230 sbi->last_age_weight = LAST_AGE_WEIGHT; 1231 sbi->max_read_extent_count = DEF_MAX_READ_EXTENT_COUNT; 1232 } 1233 1234 int __init f2fs_create_extent_cache(void) 1235 { 1236 extent_tree_slab = f2fs_kmem_cache_create("f2fs_extent_tree", 1237 sizeof(struct extent_tree)); 1238 if (!extent_tree_slab) 1239 return -ENOMEM; 1240 extent_node_slab = f2fs_kmem_cache_create("f2fs_extent_node", 1241 sizeof(struct extent_node)); 1242 if (!extent_node_slab) { 1243 kmem_cache_destroy(extent_tree_slab); 1244 return -ENOMEM; 1245 } 1246 return 0; 1247 } 1248 1249 void f2fs_destroy_extent_cache(void) 1250 { 1251 kmem_cache_destroy(extent_node_slab); 1252 kmem_cache_destroy(extent_tree_slab); 1253 } 1254