1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * fs/f2fs/segment.c 4 * 5 * Copyright (c) 2012 Samsung Electronics Co., Ltd. 6 * http://www.samsung.com/ 7 */ 8 #include <linux/fs.h> 9 #include <linux/f2fs_fs.h> 10 #include <linux/bio.h> 11 #include <linux/blkdev.h> 12 #include <linux/sched/mm.h> 13 #include <linux/prefetch.h> 14 #include <linux/kthread.h> 15 #include <linux/swap.h> 16 #include <linux/timer.h> 17 #include <linux/freezer.h> 18 #include <linux/sched/signal.h> 19 #include <linux/random.h> 20 #include <linux/fserror.h> 21 22 #include "f2fs.h" 23 #include "segment.h" 24 #include "node.h" 25 #include "gc.h" 26 #include "iostat.h" 27 #include <trace/events/f2fs.h> 28 29 #define __reverse_ffz(x) __reverse_ffs(~(x)) 30 31 static struct kmem_cache *discard_entry_slab; 32 static struct kmem_cache *discard_cmd_slab; 33 static struct kmem_cache *sit_entry_set_slab; 34 static struct kmem_cache *revoke_entry_slab; 35 36 static unsigned long __reverse_ulong(unsigned char *str) 37 { 38 unsigned long tmp = 0; 39 int shift = 24, idx = 0; 40 41 #if BITS_PER_LONG == 64 42 shift = 56; 43 #endif 44 while (shift >= 0) { 45 tmp |= (unsigned long)str[idx++] << shift; 46 shift -= BITS_PER_BYTE; 47 } 48 return tmp; 49 } 50 51 /* 52 * __reverse_ffs is copied from include/asm-generic/bitops/__ffs.h since 53 * MSB and LSB are reversed in a byte by f2fs_set_bit. 54 */ 55 static inline unsigned long __reverse_ffs(unsigned long word) 56 { 57 int num = 0; 58 59 #if BITS_PER_LONG == 64 60 if ((word & 0xffffffff00000000UL) == 0) 61 num += 32; 62 else 63 word >>= 32; 64 #endif 65 if ((word & 0xffff0000) == 0) 66 num += 16; 67 else 68 word >>= 16; 69 70 if ((word & 0xff00) == 0) 71 num += 8; 72 else 73 word >>= 8; 74 75 if ((word & 0xf0) == 0) 76 num += 4; 77 else 78 word >>= 4; 79 80 if ((word & 0xc) == 0) 81 num += 2; 82 else 83 word >>= 2; 84 85 if ((word & 0x2) == 0) 86 num += 1; 87 return num; 88 } 89 90 /* 91 * __find_rev_next(_zero)_bit is copied from lib/find_next_bit.c because 92 * f2fs_set_bit makes MSB and LSB reversed in a byte. 93 * @size must be integral times of unsigned long. 94 * Example: 95 * MSB <--> LSB 96 * f2fs_set_bit(0, bitmap) => 1000 0000 97 * f2fs_set_bit(7, bitmap) => 0000 0001 98 */ 99 static unsigned long __find_rev_next_bit(const unsigned long *addr, 100 unsigned long size, unsigned long offset) 101 { 102 const unsigned long *p = addr + BIT_WORD(offset); 103 unsigned long result = size; 104 unsigned long tmp; 105 106 if (offset >= size) 107 return size; 108 109 size -= (offset & ~(BITS_PER_LONG - 1)); 110 offset %= BITS_PER_LONG; 111 112 while (1) { 113 if (*p == 0) 114 goto pass; 115 116 tmp = __reverse_ulong((unsigned char *)p); 117 118 tmp &= ~0UL >> offset; 119 if (size < BITS_PER_LONG) 120 tmp &= (~0UL << (BITS_PER_LONG - size)); 121 if (tmp) 122 goto found; 123 pass: 124 if (size <= BITS_PER_LONG) 125 break; 126 size -= BITS_PER_LONG; 127 offset = 0; 128 p++; 129 } 130 return result; 131 found: 132 return result - size + __reverse_ffs(tmp); 133 } 134 135 static unsigned long __find_rev_next_zero_bit(const unsigned long *addr, 136 unsigned long size, unsigned long offset) 137 { 138 const unsigned long *p = addr + BIT_WORD(offset); 139 unsigned long result = size; 140 unsigned long tmp; 141 142 if (offset >= size) 143 return size; 144 145 size -= (offset & ~(BITS_PER_LONG - 1)); 146 offset %= BITS_PER_LONG; 147 148 while (1) { 149 if (*p == ~0UL) 150 goto pass; 151 152 tmp = __reverse_ulong((unsigned char *)p); 153 154 if (offset) 155 tmp |= ~0UL << (BITS_PER_LONG - offset); 156 if (size < BITS_PER_LONG) 157 tmp |= ~0UL >> size; 158 if (tmp != ~0UL) 159 goto found; 160 pass: 161 if (size <= BITS_PER_LONG) 162 break; 163 size -= BITS_PER_LONG; 164 offset = 0; 165 p++; 166 } 167 return result; 168 found: 169 return result - size + __reverse_ffz(tmp); 170 } 171 172 bool f2fs_need_SSR(struct f2fs_sb_info *sbi) 173 { 174 int node_secs = get_blocktype_secs(sbi, F2FS_DIRTY_NODES); 175 int dent_secs = get_blocktype_secs(sbi, F2FS_DIRTY_DENTS); 176 int imeta_secs = get_blocktype_secs(sbi, F2FS_DIRTY_IMETA); 177 178 if (f2fs_lfs_mode(sbi)) 179 return false; 180 if (sbi->gc_mode == GC_URGENT_HIGH) 181 return true; 182 if (unlikely(is_sbi_flag_set(sbi, SBI_CP_DISABLED))) 183 return true; 184 185 return free_sections(sbi) <= (node_secs + 2 * dent_secs + imeta_secs + 186 SM_I(sbi)->min_ssr_sections + reserved_sections(sbi)); 187 } 188 189 void f2fs_abort_atomic_write(struct inode *inode, bool clean) 190 { 191 struct f2fs_inode_info *fi = F2FS_I(inode); 192 193 if (!f2fs_is_atomic_file(inode)) 194 return; 195 196 if (clean) 197 truncate_inode_pages_final(inode->i_mapping); 198 199 release_atomic_write_cnt(inode); 200 clear_inode_flag(inode, FI_ATOMIC_COMMITTED); 201 clear_inode_flag(inode, FI_ATOMIC_REPLACE); 202 clear_inode_flag(inode, FI_ATOMIC_FILE); 203 if (is_inode_flag_set(inode, FI_ATOMIC_DIRTIED)) { 204 clear_inode_flag(inode, FI_ATOMIC_DIRTIED); 205 /* 206 * The vfs inode keeps clean during commit, but the f2fs inode 207 * doesn't. So clear the dirty state after commit and let 208 * f2fs_mark_inode_dirty_sync ensure a consistent dirty state. 209 */ 210 f2fs_inode_synced(inode); 211 f2fs_mark_inode_dirty_sync(inode, true); 212 } 213 stat_dec_atomic_inode(inode); 214 215 F2FS_I(inode)->atomic_write_task = NULL; 216 217 if (clean) { 218 f2fs_i_size_write(inode, fi->original_i_size); 219 fi->original_i_size = 0; 220 } 221 /* avoid stale dirty inode during eviction */ 222 sync_inode_metadata(inode, 0); 223 } 224 225 static int __replace_atomic_write_block(struct inode *inode, pgoff_t index, 226 block_t new_addr, block_t *old_addr, bool recover) 227 { 228 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 229 struct dnode_of_data dn; 230 struct node_info ni; 231 int err; 232 233 retry: 234 set_new_dnode(&dn, inode, NULL, NULL, 0); 235 err = f2fs_get_dnode_of_data(&dn, index, ALLOC_NODE); 236 if (err) { 237 if (err == -ENOMEM) { 238 memalloc_retry_wait(GFP_NOFS); 239 goto retry; 240 } 241 return err; 242 } 243 244 err = f2fs_get_node_info(sbi, dn.nid, &ni, false); 245 if (err) { 246 f2fs_put_dnode(&dn); 247 return err; 248 } 249 250 if (recover) { 251 /* dn.data_blkaddr is always valid */ 252 if (!__is_valid_data_blkaddr(new_addr)) { 253 if (new_addr == NULL_ADDR) 254 dec_valid_block_count(sbi, inode, 1); 255 f2fs_invalidate_blocks(sbi, dn.data_blkaddr, 1); 256 f2fs_update_data_blkaddr(&dn, new_addr); 257 } else { 258 f2fs_replace_block(sbi, &dn, dn.data_blkaddr, 259 new_addr, ni.version, true, true); 260 } 261 } else { 262 blkcnt_t count = 1; 263 264 err = inc_valid_block_count(sbi, inode, &count, true); 265 if (err) { 266 f2fs_put_dnode(&dn); 267 return err; 268 } 269 270 *old_addr = dn.data_blkaddr; 271 f2fs_truncate_data_blocks_range(&dn, 1); 272 dec_valid_block_count(sbi, F2FS_I(inode)->cow_inode, count); 273 274 f2fs_replace_block(sbi, &dn, dn.data_blkaddr, new_addr, 275 ni.version, true, false); 276 } 277 278 f2fs_put_dnode(&dn); 279 280 trace_f2fs_replace_atomic_write_block(inode, F2FS_I(inode)->cow_inode, 281 index, old_addr ? *old_addr : 0, new_addr, recover); 282 return 0; 283 } 284 285 static void __complete_revoke_list(struct inode *inode, struct list_head *head, 286 bool revoke) 287 { 288 struct revoke_entry *cur, *tmp; 289 pgoff_t start_index = 0; 290 bool truncate = is_inode_flag_set(inode, FI_ATOMIC_REPLACE); 291 292 list_for_each_entry_safe(cur, tmp, head, list) { 293 if (revoke) { 294 __replace_atomic_write_block(inode, cur->index, 295 cur->old_addr, NULL, true); 296 } else if (truncate) { 297 f2fs_truncate_hole(inode, start_index, cur->index); 298 start_index = cur->index + 1; 299 } 300 301 list_del(&cur->list); 302 kmem_cache_free(revoke_entry_slab, cur); 303 } 304 305 if (!revoke && truncate) 306 f2fs_do_truncate_blocks(inode, start_index * PAGE_SIZE, false); 307 } 308 309 static int __f2fs_commit_atomic_write(struct inode *inode) 310 { 311 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 312 struct f2fs_inode_info *fi = F2FS_I(inode); 313 struct inode *cow_inode = fi->cow_inode; 314 struct revoke_entry *new; 315 struct list_head revoke_list; 316 block_t blkaddr; 317 struct dnode_of_data dn; 318 pgoff_t len = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); 319 pgoff_t off = 0, blen, index; 320 int ret = 0, i; 321 322 INIT_LIST_HEAD(&revoke_list); 323 324 while (len) { 325 blen = min_t(pgoff_t, ADDRS_PER_BLOCK(cow_inode), len); 326 327 set_new_dnode(&dn, cow_inode, NULL, NULL, 0); 328 ret = f2fs_get_dnode_of_data(&dn, off, LOOKUP_NODE_RA); 329 if (ret && ret != -ENOENT) { 330 goto out; 331 } else if (ret == -ENOENT) { 332 ret = 0; 333 if (dn.max_level == 0) 334 goto out; 335 goto next; 336 } 337 338 blen = min((pgoff_t)ADDRS_PER_PAGE(dn.node_folio, cow_inode), 339 len); 340 index = off; 341 for (i = 0; i < blen; i++, dn.ofs_in_node++, index++) { 342 blkaddr = f2fs_data_blkaddr(&dn); 343 344 if (!__is_valid_data_blkaddr(blkaddr)) { 345 continue; 346 } else if (!f2fs_is_valid_blkaddr(sbi, blkaddr, 347 DATA_GENERIC_ENHANCE)) { 348 f2fs_put_dnode(&dn); 349 ret = -EFSCORRUPTED; 350 goto out; 351 } 352 353 new = f2fs_kmem_cache_alloc(revoke_entry_slab, GFP_NOFS, 354 true, NULL); 355 356 ret = __replace_atomic_write_block(inode, index, blkaddr, 357 &new->old_addr, false); 358 if (ret) { 359 f2fs_put_dnode(&dn); 360 kmem_cache_free(revoke_entry_slab, new); 361 goto out; 362 } 363 364 f2fs_update_data_blkaddr(&dn, NULL_ADDR); 365 new->index = index; 366 list_add_tail(&new->list, &revoke_list); 367 } 368 f2fs_put_dnode(&dn); 369 next: 370 off += blen; 371 len -= blen; 372 } 373 374 out: 375 if (time_to_inject(sbi, FAULT_ATOMIC_TIMEOUT)) 376 f2fs_schedule_timeout_killable(DEFAULT_FAULT_TIMEOUT, true); 377 378 if (ret) { 379 sbi->revoked_atomic_block += fi->atomic_write_cnt; 380 } else { 381 sbi->committed_atomic_block += fi->atomic_write_cnt; 382 set_inode_flag(inode, FI_ATOMIC_COMMITTED); 383 384 /* 385 * inode may has no FI_ATOMIC_DIRTIED flag due to no write 386 * before commit. 387 */ 388 if (is_inode_flag_set(inode, FI_ATOMIC_DIRTIED)) { 389 /* clear atomic dirty status and set vfs dirty status */ 390 clear_inode_flag(inode, FI_ATOMIC_DIRTIED); 391 f2fs_mark_inode_dirty_sync(inode, true); 392 } 393 } 394 395 __complete_revoke_list(inode, &revoke_list, ret ? true : false); 396 397 return ret; 398 } 399 400 int f2fs_commit_atomic_write(struct inode *inode) 401 { 402 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 403 struct f2fs_inode_info *fi = F2FS_I(inode); 404 struct f2fs_lock_context lc; 405 int err; 406 407 err = filemap_write_and_wait_range(inode->i_mapping, 0, LLONG_MAX); 408 if (err) 409 return err; 410 411 f2fs_down_write(&fi->i_gc_rwsem[WRITE]); 412 f2fs_lock_op(sbi, &lc); 413 414 err = __f2fs_commit_atomic_write(inode); 415 416 f2fs_unlock_op(sbi, &lc); 417 f2fs_up_write(&fi->i_gc_rwsem[WRITE]); 418 419 return err; 420 } 421 422 /* 423 * This function balances dirty node and dentry pages. 424 * In addition, it controls garbage collection. 425 */ 426 void f2fs_balance_fs(struct f2fs_sb_info *sbi, bool need) 427 { 428 if (f2fs_cp_error(sbi)) 429 return; 430 431 if (time_to_inject(sbi, FAULT_CHECKPOINT)) 432 f2fs_stop_checkpoint(sbi, false, STOP_CP_REASON_FAULT_INJECT); 433 434 /* balance_fs_bg is able to be pending */ 435 if (need && excess_cached_nats(sbi)) 436 f2fs_balance_fs_bg(sbi, false); 437 438 if (unlikely(is_sbi_flag_set(sbi, SBI_CP_DISABLED))) 439 return; 440 441 /* 442 * We should do GC or end up with checkpoint, if there are so many dirty 443 * dir/node pages without enough free segments. 444 */ 445 if (has_enough_free_secs(sbi, 0, 0)) 446 return; 447 448 /* 449 * Submit all cached OPU/IPU DATA bios before triggering 450 * foreground GC to avoid potential deadlocks. 451 */ 452 f2fs_submit_merged_write(sbi, DATA); 453 f2fs_submit_all_merged_ipu_writes(sbi); 454 455 if (test_opt(sbi, GC_MERGE) && sbi->gc_thread && 456 sbi->gc_thread->f2fs_gc_task) { 457 DEFINE_WAIT(wait); 458 459 prepare_to_wait(&sbi->gc_thread->fggc_wq, &wait, 460 TASK_UNINTERRUPTIBLE); 461 wake_up(&sbi->gc_thread->gc_wait_queue_head); 462 io_schedule(); 463 finish_wait(&sbi->gc_thread->fggc_wq, &wait); 464 } else { 465 struct f2fs_gc_control gc_control = { 466 .victim_segno = NULL_SEGNO, 467 .init_gc_type = f2fs_sb_has_blkzoned(sbi) ? 468 FG_GC : BG_GC, 469 .no_bg_gc = true, 470 .should_migrate_blocks = false, 471 .err_gc_skipped = false, 472 .nr_free_secs = 1 }; 473 474 f2fs_down_write_trace(&sbi->gc_lock, &gc_control.lc); 475 stat_inc_gc_call_count(sbi, FOREGROUND); 476 f2fs_gc(sbi, &gc_control); 477 } 478 } 479 480 static inline bool excess_dirty_threshold(struct f2fs_sb_info *sbi) 481 { 482 int factor = f2fs_rwsem_is_locked(&sbi->cp_rwsem) ? 3 : 2; 483 unsigned int dents = get_pages(sbi, F2FS_DIRTY_DENTS); 484 unsigned int qdata = get_pages(sbi, F2FS_DIRTY_QDATA); 485 unsigned int nodes = get_pages(sbi, F2FS_DIRTY_NODES); 486 unsigned int meta = get_pages(sbi, F2FS_DIRTY_META); 487 unsigned int imeta = get_pages(sbi, F2FS_DIRTY_IMETA); 488 unsigned int threshold = 489 SEGS_TO_BLKS(sbi, (factor * DEFAULT_DIRTY_THRESHOLD)); 490 unsigned int global_threshold = threshold * 3 / 2; 491 492 if (dents >= threshold || qdata >= threshold || 493 nodes >= threshold || meta >= threshold || 494 imeta >= threshold) 495 return true; 496 return dents + qdata + nodes + meta + imeta > global_threshold; 497 } 498 499 void f2fs_balance_fs_bg(struct f2fs_sb_info *sbi, bool from_bg) 500 { 501 if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING))) 502 return; 503 504 /* try to shrink extent cache when there is no enough memory */ 505 if (!f2fs_available_free_memory(sbi, READ_EXTENT_CACHE)) 506 f2fs_shrink_read_extent_tree(sbi, 507 READ_EXTENT_CACHE_SHRINK_NUMBER); 508 509 /* try to shrink age extent cache when there is no enough memory */ 510 if (!f2fs_available_free_memory(sbi, AGE_EXTENT_CACHE)) 511 f2fs_shrink_age_extent_tree(sbi, 512 AGE_EXTENT_CACHE_SHRINK_NUMBER); 513 514 /* check the # of cached NAT entries */ 515 if (!f2fs_available_free_memory(sbi, NAT_ENTRIES)) 516 f2fs_try_to_free_nats(sbi, NAT_ENTRY_PER_BLOCK); 517 518 if (!f2fs_available_free_memory(sbi, FREE_NIDS)) 519 f2fs_try_to_free_nids(sbi, MAX_FREE_NIDS); 520 else 521 f2fs_build_free_nids(sbi, false, false); 522 523 if (excess_dirty_nats(sbi) || excess_dirty_threshold(sbi) || 524 excess_prefree_segs(sbi) || !f2fs_space_for_roll_forward(sbi)) 525 goto do_sync; 526 527 /* there is background inflight IO or foreground operation recently */ 528 if (is_inflight_io(sbi, REQ_TIME) || 529 (!f2fs_time_over(sbi, REQ_TIME) && f2fs_rwsem_is_locked(&sbi->cp_rwsem))) 530 return; 531 532 /* exceed periodical checkpoint timeout threshold */ 533 if (f2fs_time_over(sbi, CP_TIME)) 534 goto do_sync; 535 536 /* checkpoint is the only way to shrink partial cached entries */ 537 if (f2fs_available_free_memory(sbi, NAT_ENTRIES) && 538 f2fs_available_free_memory(sbi, INO_ENTRIES)) 539 return; 540 541 do_sync: 542 if (test_opt(sbi, DATA_FLUSH) && from_bg) { 543 struct blk_plug plug; 544 545 mutex_lock(&sbi->flush_lock); 546 547 blk_start_plug(&plug); 548 f2fs_sync_dirty_inodes(sbi, FILE_INODE, false); 549 blk_finish_plug(&plug); 550 551 mutex_unlock(&sbi->flush_lock); 552 } 553 stat_inc_cp_call_count(sbi, BACKGROUND); 554 f2fs_sync_fs(sbi->sb, 1); 555 } 556 557 static int __submit_flush_wait(struct f2fs_sb_info *sbi, 558 struct block_device *bdev) 559 { 560 int ret = blkdev_issue_flush(bdev); 561 562 trace_f2fs_issue_flush(bdev, test_opt(sbi, NOBARRIER), 563 test_opt(sbi, FLUSH_MERGE), ret); 564 if (!ret) 565 f2fs_update_iostat(sbi, NULL, FS_FLUSH_IO, 0); 566 return ret; 567 } 568 569 static int submit_flush_wait(struct f2fs_sb_info *sbi, nid_t ino) 570 { 571 int ret = 0; 572 int i; 573 574 if (!f2fs_is_multi_device(sbi)) 575 return __submit_flush_wait(sbi, sbi->sb->s_bdev); 576 577 for (i = 0; i < sbi->s_ndevs; i++) { 578 if (!f2fs_is_dirty_device(sbi, ino, i, FLUSH_INO)) 579 continue; 580 ret = __submit_flush_wait(sbi, FDEV(i).bdev); 581 if (ret) 582 break; 583 } 584 return ret; 585 } 586 587 static int issue_flush_thread(void *data) 588 { 589 struct f2fs_sb_info *sbi = data; 590 struct flush_cmd_control *fcc = SM_I(sbi)->fcc_info; 591 wait_queue_head_t *q = &fcc->flush_wait_queue; 592 repeat: 593 if (kthread_should_stop()) 594 return 0; 595 596 if (!llist_empty(&fcc->issue_list)) { 597 struct flush_cmd *cmd, *next; 598 int ret; 599 600 fcc->dispatch_list = llist_del_all(&fcc->issue_list); 601 fcc->dispatch_list = llist_reverse_order(fcc->dispatch_list); 602 603 cmd = llist_entry(fcc->dispatch_list, struct flush_cmd, llnode); 604 605 ret = submit_flush_wait(sbi, cmd->ino); 606 atomic_inc(&fcc->issued_flush); 607 608 llist_for_each_entry_safe(cmd, next, 609 fcc->dispatch_list, llnode) { 610 cmd->ret = ret; 611 complete(&cmd->wait); 612 } 613 fcc->dispatch_list = NULL; 614 } 615 616 wait_event_interruptible(*q, 617 kthread_should_stop() || !llist_empty(&fcc->issue_list)); 618 goto repeat; 619 } 620 621 int f2fs_issue_flush(struct f2fs_sb_info *sbi, nid_t ino) 622 { 623 struct flush_cmd_control *fcc = SM_I(sbi)->fcc_info; 624 struct flush_cmd cmd; 625 int ret; 626 627 if (test_opt(sbi, NOBARRIER)) 628 return 0; 629 630 if (!test_opt(sbi, FLUSH_MERGE)) { 631 atomic_inc(&fcc->queued_flush); 632 ret = submit_flush_wait(sbi, ino); 633 atomic_dec(&fcc->queued_flush); 634 atomic_inc(&fcc->issued_flush); 635 return ret; 636 } 637 638 if (atomic_inc_return(&fcc->queued_flush) == 1 || 639 f2fs_is_multi_device(sbi)) { 640 ret = submit_flush_wait(sbi, ino); 641 atomic_dec(&fcc->queued_flush); 642 643 atomic_inc(&fcc->issued_flush); 644 return ret; 645 } 646 647 cmd.ino = ino; 648 init_completion(&cmd.wait); 649 650 llist_add(&cmd.llnode, &fcc->issue_list); 651 652 /* 653 * update issue_list before we wake up issue_flush thread, this 654 * smp_mb() pairs with another barrier in ___wait_event(), see 655 * more details in comments of waitqueue_active(). 656 */ 657 smp_mb(); 658 659 if (waitqueue_active(&fcc->flush_wait_queue)) 660 wake_up(&fcc->flush_wait_queue); 661 662 if (fcc->f2fs_issue_flush) { 663 wait_for_completion(&cmd.wait); 664 atomic_dec(&fcc->queued_flush); 665 } else { 666 struct llist_node *list; 667 668 list = llist_del_all(&fcc->issue_list); 669 if (!list) { 670 wait_for_completion(&cmd.wait); 671 atomic_dec(&fcc->queued_flush); 672 } else { 673 struct flush_cmd *tmp, *next; 674 675 ret = submit_flush_wait(sbi, ino); 676 677 llist_for_each_entry_safe(tmp, next, list, llnode) { 678 if (tmp == &cmd) { 679 cmd.ret = ret; 680 atomic_dec(&fcc->queued_flush); 681 continue; 682 } 683 tmp->ret = ret; 684 complete(&tmp->wait); 685 } 686 } 687 } 688 689 return cmd.ret; 690 } 691 692 int f2fs_create_flush_cmd_control(struct f2fs_sb_info *sbi) 693 { 694 dev_t dev = sbi->sb->s_bdev->bd_dev; 695 struct flush_cmd_control *fcc; 696 697 if (SM_I(sbi)->fcc_info) { 698 fcc = SM_I(sbi)->fcc_info; 699 if (fcc->f2fs_issue_flush) 700 return 0; 701 goto init_thread; 702 } 703 704 fcc = f2fs_kzalloc(sbi, sizeof(struct flush_cmd_control), GFP_KERNEL); 705 if (!fcc) 706 return -ENOMEM; 707 atomic_set(&fcc->issued_flush, 0); 708 atomic_set(&fcc->queued_flush, 0); 709 init_waitqueue_head(&fcc->flush_wait_queue); 710 init_llist_head(&fcc->issue_list); 711 SM_I(sbi)->fcc_info = fcc; 712 if (!test_opt(sbi, FLUSH_MERGE)) 713 return 0; 714 715 init_thread: 716 fcc->f2fs_issue_flush = kthread_run(issue_flush_thread, sbi, 717 "f2fs_flush-%u:%u", MAJOR(dev), MINOR(dev)); 718 if (IS_ERR(fcc->f2fs_issue_flush)) { 719 int err = PTR_ERR(fcc->f2fs_issue_flush); 720 721 fcc->f2fs_issue_flush = NULL; 722 return err; 723 } 724 725 return 0; 726 } 727 728 void f2fs_destroy_flush_cmd_control(struct f2fs_sb_info *sbi, bool free) 729 { 730 struct flush_cmd_control *fcc = SM_I(sbi)->fcc_info; 731 732 if (fcc && fcc->f2fs_issue_flush) { 733 struct task_struct *flush_thread = fcc->f2fs_issue_flush; 734 735 fcc->f2fs_issue_flush = NULL; 736 kthread_stop(flush_thread); 737 } 738 if (free) { 739 kfree(fcc); 740 SM_I(sbi)->fcc_info = NULL; 741 } 742 } 743 744 int f2fs_flush_device_cache(struct f2fs_sb_info *sbi) 745 { 746 int ret = 0, i; 747 748 if (!f2fs_is_multi_device(sbi)) 749 return 0; 750 751 if (test_opt(sbi, NOBARRIER)) 752 return 0; 753 754 for (i = 1; i < sbi->s_ndevs; i++) { 755 int count = DEFAULT_RETRY_IO_COUNT; 756 757 if (!f2fs_test_bit(i, (char *)&sbi->dirty_device)) 758 continue; 759 760 do { 761 ret = __submit_flush_wait(sbi, FDEV(i).bdev); 762 if (ret) 763 f2fs_schedule_timeout(DEFAULT_SCHEDULE_TIMEOUT); 764 } while (ret && --count); 765 766 if (ret) { 767 f2fs_stop_checkpoint(sbi, false, 768 STOP_CP_REASON_FLUSH_FAIL); 769 break; 770 } 771 772 spin_lock(&sbi->dev_lock); 773 f2fs_clear_bit(i, (char *)&sbi->dirty_device); 774 spin_unlock(&sbi->dev_lock); 775 } 776 777 return ret; 778 } 779 780 static void __locate_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno, 781 enum dirty_type dirty_type) 782 { 783 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 784 785 /* need not be added */ 786 if (is_curseg(sbi, segno)) 787 return; 788 789 if (!test_and_set_bit(segno, dirty_i->dirty_segmap[dirty_type])) 790 dirty_i->nr_dirty[dirty_type]++; 791 792 if (dirty_type == DIRTY) { 793 struct seg_entry *sentry = get_seg_entry(sbi, segno); 794 enum dirty_type t = sentry->type; 795 796 if (unlikely(t >= DIRTY)) { 797 f2fs_bug_on(sbi, 1); 798 return; 799 } 800 if (!test_and_set_bit(segno, dirty_i->dirty_segmap[t])) 801 dirty_i->nr_dirty[t]++; 802 803 if (__is_large_section(sbi)) { 804 unsigned int secno = GET_SEC_FROM_SEG(sbi, segno); 805 block_t valid_blocks = 806 get_valid_blocks(sbi, segno, true); 807 808 f2fs_bug_on(sbi, 809 (!is_sbi_flag_set(sbi, SBI_CP_DISABLED) && 810 !valid_blocks) || 811 valid_blocks == CAP_BLKS_PER_SEC(sbi)); 812 813 if (!is_cursec(sbi, secno)) 814 set_bit(secno, dirty_i->dirty_secmap); 815 } 816 } 817 } 818 819 static void __remove_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno, 820 enum dirty_type dirty_type) 821 { 822 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 823 block_t valid_blocks; 824 825 if (test_and_clear_bit(segno, dirty_i->dirty_segmap[dirty_type])) 826 dirty_i->nr_dirty[dirty_type]--; 827 828 if (dirty_type == DIRTY) { 829 struct seg_entry *sentry = get_seg_entry(sbi, segno); 830 enum dirty_type t = sentry->type; 831 832 if (test_and_clear_bit(segno, dirty_i->dirty_segmap[t])) 833 dirty_i->nr_dirty[t]--; 834 835 valid_blocks = get_valid_blocks(sbi, segno, true); 836 if (valid_blocks == 0) { 837 clear_bit(GET_SEC_FROM_SEG(sbi, segno), 838 dirty_i->victim_secmap); 839 #ifdef CONFIG_F2FS_CHECK_FS 840 clear_bit(segno, SIT_I(sbi)->invalid_segmap); 841 #endif 842 } 843 if (__is_large_section(sbi)) { 844 unsigned int secno = GET_SEC_FROM_SEG(sbi, segno); 845 846 if (!valid_blocks || 847 valid_blocks == CAP_BLKS_PER_SEC(sbi)) { 848 clear_bit(secno, dirty_i->dirty_secmap); 849 return; 850 } 851 852 if (!is_cursec(sbi, secno)) 853 set_bit(secno, dirty_i->dirty_secmap); 854 } 855 } 856 } 857 858 /* 859 * Should not occur error such as -ENOMEM. 860 * Adding dirty entry into seglist is not critical operation. 861 * If a given segment is one of current working segments, it won't be added. 862 */ 863 static void locate_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno) 864 { 865 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 866 unsigned short valid_blocks, ckpt_valid_blocks; 867 unsigned int usable_blocks; 868 869 if (segno == NULL_SEGNO || is_curseg(sbi, segno)) 870 return; 871 872 usable_blocks = f2fs_usable_blks_in_seg(sbi, segno); 873 mutex_lock(&dirty_i->seglist_lock); 874 875 valid_blocks = get_valid_blocks(sbi, segno, false); 876 ckpt_valid_blocks = get_ckpt_valid_blocks(sbi, segno, false); 877 878 if (valid_blocks == 0 && (!is_sbi_flag_set(sbi, SBI_CP_DISABLED) || 879 ckpt_valid_blocks == usable_blocks)) { 880 __locate_dirty_segment(sbi, segno, PRE); 881 __remove_dirty_segment(sbi, segno, DIRTY); 882 } else if (valid_blocks < usable_blocks) { 883 __locate_dirty_segment(sbi, segno, DIRTY); 884 } else { 885 /* Recovery routine with SSR needs this */ 886 __remove_dirty_segment(sbi, segno, DIRTY); 887 } 888 889 mutex_unlock(&dirty_i->seglist_lock); 890 } 891 892 /* This moves currently empty dirty blocks to prefree. Must hold seglist_lock */ 893 void f2fs_dirty_to_prefree(struct f2fs_sb_info *sbi) 894 { 895 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 896 unsigned int segno; 897 898 mutex_lock(&dirty_i->seglist_lock); 899 for_each_set_bit(segno, dirty_i->dirty_segmap[DIRTY], MAIN_SEGS(sbi)) { 900 if (get_valid_blocks(sbi, segno, false)) 901 continue; 902 if (is_curseg(sbi, segno)) 903 continue; 904 __locate_dirty_segment(sbi, segno, PRE); 905 __remove_dirty_segment(sbi, segno, DIRTY); 906 } 907 mutex_unlock(&dirty_i->seglist_lock); 908 } 909 910 block_t f2fs_get_unusable_blocks(struct f2fs_sb_info *sbi) 911 { 912 int ovp_hole_segs = 913 (overprovision_segments(sbi) - reserved_segments(sbi)); 914 block_t ovp_holes = SEGS_TO_BLKS(sbi, ovp_hole_segs); 915 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 916 block_t holes[2] = {0, 0}; /* DATA and NODE */ 917 block_t unusable; 918 struct seg_entry *se; 919 unsigned int segno; 920 921 mutex_lock(&dirty_i->seglist_lock); 922 for_each_set_bit(segno, dirty_i->dirty_segmap[DIRTY], MAIN_SEGS(sbi)) { 923 se = get_seg_entry(sbi, segno); 924 if (IS_NODESEG(se->type)) 925 holes[NODE] += f2fs_usable_blks_in_seg(sbi, segno) - 926 se->valid_blocks; 927 else 928 holes[DATA] += f2fs_usable_blks_in_seg(sbi, segno) - 929 se->valid_blocks; 930 } 931 mutex_unlock(&dirty_i->seglist_lock); 932 933 unusable = max(holes[DATA], holes[NODE]); 934 if (unusable > ovp_holes) 935 return unusable - ovp_holes; 936 return 0; 937 } 938 939 int f2fs_disable_cp_again(struct f2fs_sb_info *sbi, block_t unusable) 940 { 941 int ovp_hole_segs = 942 (overprovision_segments(sbi) - reserved_segments(sbi)); 943 944 if (F2FS_OPTION(sbi).unusable_cap_perc == 100) 945 return 0; 946 if (unusable > F2FS_OPTION(sbi).unusable_cap) 947 return -EAGAIN; 948 if (is_sbi_flag_set(sbi, SBI_CP_DISABLED_QUICK) && 949 dirty_segments(sbi) > ovp_hole_segs) 950 return -EAGAIN; 951 if (has_not_enough_free_secs(sbi, 0, 0)) 952 return -EAGAIN; 953 return 0; 954 } 955 956 /* This is only used by SBI_CP_DISABLED */ 957 static unsigned int get_free_segment(struct f2fs_sb_info *sbi) 958 { 959 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 960 unsigned int segno = 0; 961 962 mutex_lock(&dirty_i->seglist_lock); 963 for_each_set_bit(segno, dirty_i->dirty_segmap[DIRTY], MAIN_SEGS(sbi)) { 964 if (get_valid_blocks(sbi, segno, false)) 965 continue; 966 if (get_ckpt_valid_blocks(sbi, segno, false)) 967 continue; 968 mutex_unlock(&dirty_i->seglist_lock); 969 return segno; 970 } 971 mutex_unlock(&dirty_i->seglist_lock); 972 return NULL_SEGNO; 973 } 974 975 static struct discard_cmd *__create_discard_cmd(struct f2fs_sb_info *sbi, 976 struct block_device *bdev, block_t lstart, 977 block_t start, block_t len) 978 { 979 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 980 struct list_head *pend_list; 981 struct discard_cmd *dc; 982 983 f2fs_bug_on(sbi, !len); 984 985 pend_list = &dcc->pend_list[plist_idx(len)]; 986 987 dc = f2fs_kmem_cache_alloc(discard_cmd_slab, GFP_NOFS, true, NULL); 988 INIT_LIST_HEAD(&dc->list); 989 dc->bdev = bdev; 990 dc->di.lstart = lstart; 991 dc->di.start = start; 992 dc->di.len = len; 993 dc->ref = 0; 994 dc->state = D_PREP; 995 dc->queued = 0; 996 dc->error = 0; 997 init_completion(&dc->wait); 998 list_add_tail(&dc->list, pend_list); 999 spin_lock_init(&dc->lock); 1000 dc->bio_ref = 0; 1001 atomic_inc(&dcc->discard_cmd_cnt); 1002 dcc->undiscard_blks += len; 1003 1004 return dc; 1005 } 1006 1007 static bool f2fs_check_discard_tree(struct f2fs_sb_info *sbi) 1008 { 1009 #ifdef CONFIG_F2FS_CHECK_FS 1010 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1011 struct rb_node *cur = rb_first_cached(&dcc->root), *next; 1012 struct discard_cmd *cur_dc, *next_dc; 1013 1014 while (cur) { 1015 next = rb_next(cur); 1016 if (!next) 1017 return true; 1018 1019 cur_dc = rb_entry(cur, struct discard_cmd, rb_node); 1020 next_dc = rb_entry(next, struct discard_cmd, rb_node); 1021 1022 if (cur_dc->di.lstart + cur_dc->di.len > next_dc->di.lstart) { 1023 f2fs_info(sbi, "broken discard_rbtree, " 1024 "cur(%u, %u) next(%u, %u)", 1025 cur_dc->di.lstart, cur_dc->di.len, 1026 next_dc->di.lstart, next_dc->di.len); 1027 return false; 1028 } 1029 cur = next; 1030 } 1031 #endif 1032 return true; 1033 } 1034 1035 static struct discard_cmd *__lookup_discard_cmd(struct f2fs_sb_info *sbi, 1036 block_t blkaddr) 1037 { 1038 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1039 struct rb_node *node = dcc->root.rb_root.rb_node; 1040 struct discard_cmd *dc; 1041 1042 while (node) { 1043 dc = rb_entry(node, struct discard_cmd, rb_node); 1044 1045 if (blkaddr < dc->di.lstart) 1046 node = node->rb_left; 1047 else if (blkaddr >= dc->di.lstart + dc->di.len) 1048 node = node->rb_right; 1049 else 1050 return dc; 1051 } 1052 return NULL; 1053 } 1054 1055 static struct discard_cmd *__lookup_discard_cmd_ret(struct rb_root_cached *root, 1056 block_t blkaddr, 1057 struct discard_cmd **prev_entry, 1058 struct discard_cmd **next_entry, 1059 struct rb_node ***insert_p, 1060 struct rb_node **insert_parent) 1061 { 1062 struct rb_node **pnode = &root->rb_root.rb_node; 1063 struct rb_node *parent = NULL, *tmp_node; 1064 struct discard_cmd *dc; 1065 1066 *insert_p = NULL; 1067 *insert_parent = NULL; 1068 *prev_entry = NULL; 1069 *next_entry = NULL; 1070 1071 if (RB_EMPTY_ROOT(&root->rb_root)) 1072 return NULL; 1073 1074 while (*pnode) { 1075 parent = *pnode; 1076 dc = rb_entry(*pnode, struct discard_cmd, rb_node); 1077 1078 if (blkaddr < dc->di.lstart) 1079 pnode = &(*pnode)->rb_left; 1080 else if (blkaddr >= dc->di.lstart + dc->di.len) 1081 pnode = &(*pnode)->rb_right; 1082 else 1083 goto lookup_neighbors; 1084 } 1085 1086 *insert_p = pnode; 1087 *insert_parent = parent; 1088 1089 dc = rb_entry(parent, struct discard_cmd, rb_node); 1090 tmp_node = parent; 1091 if (parent && blkaddr > dc->di.lstart) 1092 tmp_node = rb_next(parent); 1093 *next_entry = rb_entry_safe(tmp_node, struct discard_cmd, rb_node); 1094 1095 tmp_node = parent; 1096 if (parent && blkaddr < dc->di.lstart) 1097 tmp_node = rb_prev(parent); 1098 *prev_entry = rb_entry_safe(tmp_node, struct discard_cmd, rb_node); 1099 return NULL; 1100 1101 lookup_neighbors: 1102 /* lookup prev node for merging backward later */ 1103 tmp_node = rb_prev(&dc->rb_node); 1104 *prev_entry = rb_entry_safe(tmp_node, struct discard_cmd, rb_node); 1105 1106 /* lookup next node for merging frontward later */ 1107 tmp_node = rb_next(&dc->rb_node); 1108 *next_entry = rb_entry_safe(tmp_node, struct discard_cmd, rb_node); 1109 return dc; 1110 } 1111 1112 static void __detach_discard_cmd(struct discard_cmd_control *dcc, 1113 struct discard_cmd *dc) 1114 { 1115 if (dc->state == D_DONE) 1116 atomic_sub(dc->queued, &dcc->queued_discard); 1117 1118 list_del(&dc->list); 1119 rb_erase_cached(&dc->rb_node, &dcc->root); 1120 dcc->undiscard_blks -= dc->di.len; 1121 1122 kmem_cache_free(discard_cmd_slab, dc); 1123 1124 atomic_dec(&dcc->discard_cmd_cnt); 1125 } 1126 1127 static void __remove_discard_cmd(struct f2fs_sb_info *sbi, 1128 struct discard_cmd *dc) 1129 { 1130 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1131 unsigned long flags; 1132 1133 trace_f2fs_remove_discard(dc->bdev, dc->di.start, dc->di.len); 1134 1135 spin_lock_irqsave(&dc->lock, flags); 1136 if (dc->bio_ref) { 1137 spin_unlock_irqrestore(&dc->lock, flags); 1138 return; 1139 } 1140 spin_unlock_irqrestore(&dc->lock, flags); 1141 1142 f2fs_bug_on(sbi, dc->ref); 1143 1144 if (dc->error == -EOPNOTSUPP) 1145 dc->error = 0; 1146 1147 if (dc->error) 1148 f2fs_info_ratelimited(sbi, 1149 "Issue discard(%u, %u, %u) failed, ret: %d", 1150 dc->di.lstart, dc->di.start, dc->di.len, dc->error); 1151 __detach_discard_cmd(dcc, dc); 1152 } 1153 1154 static void f2fs_submit_discard_endio(struct bio *bio) 1155 { 1156 struct discard_cmd *dc = (struct discard_cmd *)bio->bi_private; 1157 unsigned long flags; 1158 1159 spin_lock_irqsave(&dc->lock, flags); 1160 if (!dc->error) 1161 dc->error = blk_status_to_errno(bio->bi_status); 1162 dc->bio_ref--; 1163 if (!dc->bio_ref && dc->state == D_SUBMIT) { 1164 dc->state = D_DONE; 1165 complete_all(&dc->wait); 1166 } 1167 spin_unlock_irqrestore(&dc->lock, flags); 1168 bio_put(bio); 1169 } 1170 1171 static void __check_sit_bitmap(struct f2fs_sb_info *sbi, 1172 block_t start, block_t end) 1173 { 1174 #ifdef CONFIG_F2FS_CHECK_FS 1175 struct seg_entry *sentry; 1176 unsigned int segno; 1177 block_t blk = start; 1178 unsigned long offset, size, *map; 1179 1180 while (blk < end) { 1181 segno = GET_SEGNO(sbi, blk); 1182 sentry = get_seg_entry(sbi, segno); 1183 offset = GET_BLKOFF_FROM_SEG0(sbi, blk); 1184 1185 if (end < START_BLOCK(sbi, segno + 1)) 1186 size = GET_BLKOFF_FROM_SEG0(sbi, end); 1187 else 1188 size = BLKS_PER_SEG(sbi); 1189 map = (unsigned long *)(sentry->cur_valid_map); 1190 offset = __find_rev_next_bit(map, size, offset); 1191 f2fs_bug_on(sbi, offset != size); 1192 blk = START_BLOCK(sbi, segno + 1); 1193 } 1194 #endif 1195 } 1196 1197 static void __init_discard_policy(struct f2fs_sb_info *sbi, 1198 struct discard_policy *dpolicy, 1199 int discard_type, unsigned int granularity) 1200 { 1201 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1202 1203 /* common policy */ 1204 dpolicy->type = discard_type; 1205 dpolicy->sync = true; 1206 dpolicy->ordered = false; 1207 dpolicy->granularity = granularity; 1208 1209 dpolicy->max_requests = dcc->max_discard_request; 1210 dpolicy->io_aware_gran = dcc->discard_io_aware_gran; 1211 dpolicy->timeout = false; 1212 1213 if (discard_type == DPOLICY_BG) { 1214 dpolicy->min_interval = dcc->min_discard_issue_time; 1215 dpolicy->mid_interval = dcc->mid_discard_issue_time; 1216 dpolicy->max_interval = dcc->max_discard_issue_time; 1217 if (dcc->discard_io_aware == DPOLICY_IO_AWARE_ENABLE) 1218 dpolicy->io_aware = true; 1219 else if (dcc->discard_io_aware == DPOLICY_IO_AWARE_DISABLE) 1220 dpolicy->io_aware = false; 1221 dpolicy->sync = false; 1222 dpolicy->ordered = true; 1223 if (utilization(sbi) > dcc->discard_urgent_util) { 1224 dpolicy->granularity = MIN_DISCARD_GRANULARITY; 1225 if (atomic_read(&dcc->discard_cmd_cnt)) 1226 dpolicy->max_interval = 1227 dcc->min_discard_issue_time; 1228 } 1229 } else if (discard_type == DPOLICY_FORCE) { 1230 dpolicy->min_interval = dcc->min_discard_issue_time; 1231 dpolicy->mid_interval = dcc->mid_discard_issue_time; 1232 dpolicy->max_interval = dcc->max_discard_issue_time; 1233 dpolicy->io_aware = false; 1234 } else if (discard_type == DPOLICY_FSTRIM) { 1235 dpolicy->io_aware = false; 1236 } else if (discard_type == DPOLICY_UMOUNT) { 1237 dpolicy->io_aware = false; 1238 /* we need to issue all to keep CP_TRIMMED_FLAG */ 1239 dpolicy->granularity = MIN_DISCARD_GRANULARITY; 1240 dpolicy->timeout = true; 1241 } 1242 } 1243 1244 static void __update_discard_tree_range(struct f2fs_sb_info *sbi, 1245 struct block_device *bdev, block_t lstart, 1246 block_t start, block_t len); 1247 1248 #ifdef CONFIG_BLK_DEV_ZONED 1249 static void __submit_zone_reset_cmd(struct f2fs_sb_info *sbi, 1250 struct discard_cmd *dc, blk_opf_t flag, 1251 struct list_head *wait_list, 1252 unsigned int *issued) 1253 { 1254 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1255 struct block_device *bdev = dc->bdev; 1256 struct bio *bio = bio_alloc(bdev, 0, REQ_OP_ZONE_RESET | flag, GFP_NOFS); 1257 unsigned long flags; 1258 1259 trace_f2fs_issue_reset_zone(bdev, dc->di.start); 1260 1261 spin_lock_irqsave(&dc->lock, flags); 1262 dc->state = D_SUBMIT; 1263 dc->bio_ref++; 1264 spin_unlock_irqrestore(&dc->lock, flags); 1265 1266 if (issued) 1267 (*issued)++; 1268 1269 atomic_inc(&dcc->queued_discard); 1270 dc->queued++; 1271 list_move_tail(&dc->list, wait_list); 1272 1273 /* sanity check on discard range */ 1274 __check_sit_bitmap(sbi, dc->di.lstart, dc->di.lstart + dc->di.len); 1275 1276 bio->bi_iter.bi_sector = SECTOR_FROM_BLOCK(dc->di.start); 1277 bio->bi_private = dc; 1278 bio->bi_end_io = f2fs_submit_discard_endio; 1279 submit_bio(bio); 1280 1281 atomic_inc(&dcc->issued_discard); 1282 f2fs_update_iostat(sbi, NULL, FS_ZONE_RESET_IO, dc->di.len * F2FS_BLKSIZE); 1283 } 1284 #endif 1285 1286 /* this function is copied from blkdev_issue_discard from block/blk-lib.c */ 1287 static int __submit_discard_cmd(struct f2fs_sb_info *sbi, 1288 struct discard_policy *dpolicy, 1289 struct discard_cmd *dc, int *issued) 1290 { 1291 struct block_device *bdev = dc->bdev; 1292 unsigned int max_discard_blocks = 1293 SECTOR_TO_BLOCK(bdev_max_discard_sectors(bdev)); 1294 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1295 struct list_head *wait_list = (dpolicy->type == DPOLICY_FSTRIM) ? 1296 &(dcc->fstrim_list) : &(dcc->wait_list); 1297 blk_opf_t flag = dpolicy->sync ? REQ_SYNC : 0; 1298 block_t lstart, start, len, total_len; 1299 1300 if (dc->state != D_PREP) 1301 return 0; 1302 1303 if (is_sbi_flag_set(sbi, SBI_NEED_FSCK)) 1304 return 0; 1305 1306 #ifdef CONFIG_BLK_DEV_ZONED 1307 if (f2fs_sb_has_blkzoned(sbi) && bdev_is_zoned(bdev)) { 1308 int devi = f2fs_bdev_index(sbi, bdev); 1309 1310 if (devi < 0) 1311 return -EINVAL; 1312 1313 if (f2fs_blkz_is_seq(sbi, devi, dc->di.start)) { 1314 __submit_zone_reset_cmd(sbi, dc, flag, 1315 wait_list, issued); 1316 return 0; 1317 } 1318 } 1319 #endif 1320 1321 /* 1322 * stop issuing discard for any of below cases: 1323 * 1. device is conventional zone, but it doesn't support discard. 1324 * 2. device is regulare device, after snapshot it doesn't support 1325 * discard. 1326 */ 1327 if (!bdev_max_discard_sectors(bdev)) 1328 return -EOPNOTSUPP; 1329 1330 trace_f2fs_issue_discard(bdev, dc->di.start, dc->di.len); 1331 1332 lstart = dc->di.lstart; 1333 start = dc->di.start; 1334 len = dc->di.len; 1335 total_len = len; 1336 1337 dc->di.len = 0; 1338 1339 while (total_len && *issued < dpolicy->max_requests) { 1340 struct bio *bio = NULL; 1341 unsigned long flags; 1342 bool last = true; 1343 1344 if (len > max_discard_blocks) { 1345 len = max_discard_blocks; 1346 last = false; 1347 } 1348 1349 (*issued)++; 1350 if (*issued == dpolicy->max_requests) 1351 last = true; 1352 1353 dc->di.len += len; 1354 1355 __blkdev_issue_discard(bdev, SECTOR_FROM_BLOCK(start), 1356 SECTOR_FROM_BLOCK(len), GFP_NOFS, &bio); 1357 f2fs_bug_on(sbi, !bio); 1358 1359 /* 1360 * should keep before submission to avoid D_DONE 1361 * right away 1362 */ 1363 spin_lock_irqsave(&dc->lock, flags); 1364 if (last) 1365 dc->state = D_SUBMIT; 1366 else 1367 dc->state = D_PARTIAL; 1368 dc->bio_ref++; 1369 spin_unlock_irqrestore(&dc->lock, flags); 1370 1371 atomic_inc(&dcc->queued_discard); 1372 dc->queued++; 1373 list_move_tail(&dc->list, wait_list); 1374 1375 /* sanity check on discard range */ 1376 __check_sit_bitmap(sbi, lstart, lstart + len); 1377 1378 bio->bi_private = dc; 1379 bio->bi_end_io = f2fs_submit_discard_endio; 1380 bio->bi_opf |= flag; 1381 submit_bio(bio); 1382 1383 atomic_inc(&dcc->issued_discard); 1384 1385 f2fs_update_iostat(sbi, NULL, FS_DISCARD_IO, len * F2FS_BLKSIZE); 1386 1387 lstart += len; 1388 start += len; 1389 total_len -= len; 1390 len = total_len; 1391 } 1392 1393 if (len) { 1394 dcc->undiscard_blks -= len; 1395 __update_discard_tree_range(sbi, bdev, lstart, start, len); 1396 } 1397 return 0; 1398 } 1399 1400 static void __insert_discard_cmd(struct f2fs_sb_info *sbi, 1401 struct block_device *bdev, block_t lstart, 1402 block_t start, block_t len) 1403 { 1404 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1405 struct rb_node **p = &dcc->root.rb_root.rb_node; 1406 struct rb_node *parent = NULL; 1407 struct discard_cmd *dc; 1408 bool leftmost = true; 1409 1410 /* look up rb tree to find parent node */ 1411 while (*p) { 1412 parent = *p; 1413 dc = rb_entry(parent, struct discard_cmd, rb_node); 1414 1415 if (lstart < dc->di.lstart) { 1416 p = &(*p)->rb_left; 1417 } else if (lstart >= dc->di.lstart + dc->di.len) { 1418 p = &(*p)->rb_right; 1419 leftmost = false; 1420 } else { 1421 /* Let's skip to add, if exists */ 1422 return; 1423 } 1424 } 1425 1426 dc = __create_discard_cmd(sbi, bdev, lstart, start, len); 1427 1428 rb_link_node(&dc->rb_node, parent, p); 1429 rb_insert_color_cached(&dc->rb_node, &dcc->root, leftmost); 1430 } 1431 1432 static void __relocate_discard_cmd(struct discard_cmd_control *dcc, 1433 struct discard_cmd *dc) 1434 { 1435 list_move_tail(&dc->list, &dcc->pend_list[plist_idx(dc->di.len)]); 1436 } 1437 1438 static void __punch_discard_cmd(struct f2fs_sb_info *sbi, 1439 struct discard_cmd *dc, block_t blkaddr) 1440 { 1441 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1442 struct discard_info di = dc->di; 1443 bool modified = false; 1444 1445 if (dc->state == D_DONE || dc->di.len == 1) { 1446 __remove_discard_cmd(sbi, dc); 1447 return; 1448 } 1449 1450 dcc->undiscard_blks -= di.len; 1451 1452 if (blkaddr > di.lstart) { 1453 dc->di.len = blkaddr - dc->di.lstart; 1454 dcc->undiscard_blks += dc->di.len; 1455 __relocate_discard_cmd(dcc, dc); 1456 modified = true; 1457 } 1458 1459 if (blkaddr < di.lstart + di.len - 1) { 1460 if (modified) { 1461 __insert_discard_cmd(sbi, dc->bdev, blkaddr + 1, 1462 di.start + blkaddr + 1 - di.lstart, 1463 di.lstart + di.len - 1 - blkaddr); 1464 } else { 1465 dc->di.lstart++; 1466 dc->di.len--; 1467 dc->di.start++; 1468 dcc->undiscard_blks += dc->di.len; 1469 __relocate_discard_cmd(dcc, dc); 1470 } 1471 } 1472 } 1473 1474 static void __update_discard_tree_range(struct f2fs_sb_info *sbi, 1475 struct block_device *bdev, block_t lstart, 1476 block_t start, block_t len) 1477 { 1478 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1479 struct discard_cmd *prev_dc = NULL, *next_dc = NULL; 1480 struct discard_cmd *dc; 1481 struct discard_info di = {0}; 1482 struct rb_node **insert_p = NULL, *insert_parent = NULL; 1483 unsigned int max_discard_blocks = 1484 SECTOR_TO_BLOCK(bdev_max_discard_sectors(bdev)); 1485 block_t end = lstart + len; 1486 1487 dc = __lookup_discard_cmd_ret(&dcc->root, lstart, 1488 &prev_dc, &next_dc, &insert_p, &insert_parent); 1489 if (dc) 1490 prev_dc = dc; 1491 1492 if (!prev_dc) { 1493 di.lstart = lstart; 1494 di.len = next_dc ? next_dc->di.lstart - lstart : len; 1495 di.len = min(di.len, len); 1496 di.start = start; 1497 } 1498 1499 while (1) { 1500 struct rb_node *node; 1501 bool merged = false; 1502 struct discard_cmd *tdc = NULL; 1503 1504 if (prev_dc) { 1505 di.lstart = prev_dc->di.lstart + prev_dc->di.len; 1506 if (di.lstart < lstart) 1507 di.lstart = lstart; 1508 if (di.lstart >= end) 1509 break; 1510 1511 if (!next_dc || next_dc->di.lstart > end) 1512 di.len = end - di.lstart; 1513 else 1514 di.len = next_dc->di.lstart - di.lstart; 1515 di.start = start + di.lstart - lstart; 1516 } 1517 1518 if (!di.len) 1519 goto next; 1520 1521 if (prev_dc && prev_dc->state == D_PREP && 1522 prev_dc->bdev == bdev && 1523 __is_discard_back_mergeable(&di, &prev_dc->di, 1524 max_discard_blocks)) { 1525 prev_dc->di.len += di.len; 1526 dcc->undiscard_blks += di.len; 1527 __relocate_discard_cmd(dcc, prev_dc); 1528 di = prev_dc->di; 1529 tdc = prev_dc; 1530 merged = true; 1531 } 1532 1533 if (next_dc && next_dc->state == D_PREP && 1534 next_dc->bdev == bdev && 1535 __is_discard_front_mergeable(&di, &next_dc->di, 1536 max_discard_blocks)) { 1537 next_dc->di.lstart = di.lstart; 1538 next_dc->di.len += di.len; 1539 next_dc->di.start = di.start; 1540 dcc->undiscard_blks += di.len; 1541 __relocate_discard_cmd(dcc, next_dc); 1542 if (tdc) 1543 __remove_discard_cmd(sbi, tdc); 1544 merged = true; 1545 } 1546 1547 if (!merged) 1548 __insert_discard_cmd(sbi, bdev, 1549 di.lstart, di.start, di.len); 1550 next: 1551 prev_dc = next_dc; 1552 if (!prev_dc) 1553 break; 1554 1555 node = rb_next(&prev_dc->rb_node); 1556 next_dc = rb_entry_safe(node, struct discard_cmd, rb_node); 1557 } 1558 } 1559 1560 #ifdef CONFIG_BLK_DEV_ZONED 1561 static void __queue_zone_reset_cmd(struct f2fs_sb_info *sbi, 1562 struct block_device *bdev, block_t blkstart, block_t lblkstart, 1563 block_t blklen) 1564 { 1565 trace_f2fs_queue_reset_zone(bdev, blkstart); 1566 1567 mutex_lock(&SM_I(sbi)->dcc_info->cmd_lock); 1568 __insert_discard_cmd(sbi, bdev, lblkstart, blkstart, blklen); 1569 mutex_unlock(&SM_I(sbi)->dcc_info->cmd_lock); 1570 } 1571 #endif 1572 1573 static void __queue_discard_cmd(struct f2fs_sb_info *sbi, 1574 struct block_device *bdev, block_t blkstart, block_t blklen) 1575 { 1576 block_t lblkstart = blkstart; 1577 1578 if (!f2fs_bdev_support_discard(bdev)) 1579 return; 1580 1581 trace_f2fs_queue_discard(bdev, blkstart, blklen); 1582 1583 if (f2fs_is_multi_device(sbi)) { 1584 int devi = f2fs_target_device_index(sbi, blkstart); 1585 1586 blkstart -= FDEV(devi).start_blk; 1587 } 1588 mutex_lock(&SM_I(sbi)->dcc_info->cmd_lock); 1589 __update_discard_tree_range(sbi, bdev, lblkstart, blkstart, blklen); 1590 mutex_unlock(&SM_I(sbi)->dcc_info->cmd_lock); 1591 } 1592 1593 static void __issue_discard_cmd_orderly(struct f2fs_sb_info *sbi, 1594 struct discard_policy *dpolicy, int *issued) 1595 { 1596 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1597 struct discard_cmd *prev_dc = NULL, *next_dc = NULL; 1598 struct rb_node **insert_p = NULL, *insert_parent = NULL; 1599 struct discard_cmd *dc; 1600 struct blk_plug plug; 1601 bool io_interrupted = false; 1602 1603 mutex_lock(&dcc->cmd_lock); 1604 dc = __lookup_discard_cmd_ret(&dcc->root, dcc->next_pos, 1605 &prev_dc, &next_dc, &insert_p, &insert_parent); 1606 if (!dc) 1607 dc = next_dc; 1608 1609 blk_start_plug(&plug); 1610 1611 while (dc) { 1612 struct rb_node *node; 1613 int err = 0; 1614 1615 if (dc->state != D_PREP) 1616 goto next; 1617 1618 if (*issued > 0 && unlikely(freezing(current))) 1619 break; 1620 1621 if (dpolicy->io_aware && !is_idle(sbi, DISCARD_TIME)) { 1622 io_interrupted = true; 1623 break; 1624 } 1625 1626 dcc->next_pos = dc->di.lstart + dc->di.len; 1627 err = __submit_discard_cmd(sbi, dpolicy, dc, issued); 1628 1629 if (*issued >= dpolicy->max_requests) 1630 break; 1631 next: 1632 node = rb_next(&dc->rb_node); 1633 if (err) 1634 __remove_discard_cmd(sbi, dc); 1635 dc = rb_entry_safe(node, struct discard_cmd, rb_node); 1636 } 1637 1638 blk_finish_plug(&plug); 1639 1640 if (!dc) 1641 dcc->next_pos = 0; 1642 1643 mutex_unlock(&dcc->cmd_lock); 1644 1645 if (!(*issued) && io_interrupted) 1646 *issued = -1; 1647 } 1648 static unsigned int __wait_all_discard_cmd(struct f2fs_sb_info *sbi, 1649 struct discard_policy *dpolicy); 1650 1651 static int __issue_discard_cmd(struct f2fs_sb_info *sbi, 1652 struct discard_policy *dpolicy) 1653 { 1654 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1655 struct list_head *pend_list; 1656 struct discard_cmd *dc, *tmp; 1657 struct blk_plug plug; 1658 int i, issued; 1659 bool io_interrupted = false; 1660 bool suspended = false; 1661 1662 if (dpolicy->timeout) 1663 f2fs_update_time(sbi, UMOUNT_DISCARD_TIMEOUT); 1664 1665 retry: 1666 issued = 0; 1667 for (i = MAX_PLIST_NUM - 1; i >= 0; i--) { 1668 if (dpolicy->timeout && 1669 f2fs_time_over(sbi, UMOUNT_DISCARD_TIMEOUT)) 1670 break; 1671 1672 if (i + 1 < dpolicy->granularity) 1673 break; 1674 1675 if (i + 1 < dcc->max_ordered_discard && dpolicy->ordered) { 1676 __issue_discard_cmd_orderly(sbi, dpolicy, &issued); 1677 return issued; 1678 } 1679 1680 pend_list = &dcc->pend_list[i]; 1681 1682 mutex_lock(&dcc->cmd_lock); 1683 if (list_empty(pend_list)) 1684 goto next; 1685 if (unlikely(dcc->rbtree_check)) 1686 f2fs_bug_on(sbi, !f2fs_check_discard_tree(sbi)); 1687 blk_start_plug(&plug); 1688 list_for_each_entry_safe(dc, tmp, pend_list, list) { 1689 f2fs_bug_on(sbi, dc->state != D_PREP); 1690 1691 if (issued > 0 && unlikely(freezing(current))) { 1692 suspended = true; 1693 break; 1694 } 1695 1696 if (dpolicy->timeout && 1697 f2fs_time_over(sbi, UMOUNT_DISCARD_TIMEOUT)) 1698 break; 1699 1700 if (dpolicy->io_aware && i < dpolicy->io_aware_gran && 1701 !is_idle(sbi, DISCARD_TIME)) { 1702 io_interrupted = true; 1703 break; 1704 } 1705 1706 __submit_discard_cmd(sbi, dpolicy, dc, &issued); 1707 1708 if (issued >= dpolicy->max_requests) 1709 break; 1710 } 1711 blk_finish_plug(&plug); 1712 next: 1713 mutex_unlock(&dcc->cmd_lock); 1714 1715 if (issued >= dpolicy->max_requests || io_interrupted || 1716 suspended) 1717 break; 1718 } 1719 1720 if (dpolicy->type == DPOLICY_UMOUNT && issued) { 1721 __wait_all_discard_cmd(sbi, dpolicy); 1722 goto retry; 1723 } 1724 1725 if (!issued && io_interrupted) 1726 issued = -1; 1727 1728 return issued; 1729 } 1730 1731 static bool __drop_discard_cmd(struct f2fs_sb_info *sbi) 1732 { 1733 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1734 struct list_head *pend_list; 1735 struct discard_cmd *dc, *tmp; 1736 int i; 1737 bool dropped = false; 1738 1739 mutex_lock(&dcc->cmd_lock); 1740 for (i = MAX_PLIST_NUM - 1; i >= 0; i--) { 1741 pend_list = &dcc->pend_list[i]; 1742 list_for_each_entry_safe(dc, tmp, pend_list, list) { 1743 f2fs_bug_on(sbi, dc->state != D_PREP); 1744 __remove_discard_cmd(sbi, dc); 1745 dropped = true; 1746 } 1747 } 1748 mutex_unlock(&dcc->cmd_lock); 1749 1750 return dropped; 1751 } 1752 1753 void f2fs_drop_discard_cmd(struct f2fs_sb_info *sbi) 1754 { 1755 __drop_discard_cmd(sbi); 1756 } 1757 1758 static unsigned int __wait_one_discard_bio(struct f2fs_sb_info *sbi, 1759 struct discard_cmd *dc) 1760 { 1761 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1762 unsigned int len = 0; 1763 1764 wait_for_completion_io(&dc->wait); 1765 mutex_lock(&dcc->cmd_lock); 1766 f2fs_bug_on(sbi, dc->state != D_DONE); 1767 dc->ref--; 1768 if (!dc->ref) { 1769 if (!dc->error) 1770 len = dc->di.len; 1771 __remove_discard_cmd(sbi, dc); 1772 } 1773 mutex_unlock(&dcc->cmd_lock); 1774 1775 return len; 1776 } 1777 1778 static unsigned int __wait_discard_cmd_range(struct f2fs_sb_info *sbi, 1779 struct discard_policy *dpolicy, 1780 block_t start, block_t end) 1781 { 1782 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1783 struct list_head *wait_list = (dpolicy->type == DPOLICY_FSTRIM) ? 1784 &(dcc->fstrim_list) : &(dcc->wait_list); 1785 struct discard_cmd *dc = NULL, *iter, *tmp; 1786 unsigned int trimmed = 0; 1787 1788 next: 1789 dc = NULL; 1790 1791 mutex_lock(&dcc->cmd_lock); 1792 list_for_each_entry_safe(iter, tmp, wait_list, list) { 1793 if (iter->di.lstart + iter->di.len <= start || 1794 end <= iter->di.lstart) 1795 continue; 1796 if (iter->di.len < dpolicy->granularity) 1797 continue; 1798 if (iter->state == D_DONE && !iter->ref) { 1799 wait_for_completion_io(&iter->wait); 1800 if (!iter->error) 1801 trimmed += iter->di.len; 1802 __remove_discard_cmd(sbi, iter); 1803 } else { 1804 iter->ref++; 1805 dc = iter; 1806 break; 1807 } 1808 } 1809 mutex_unlock(&dcc->cmd_lock); 1810 1811 if (dc) { 1812 trimmed += __wait_one_discard_bio(sbi, dc); 1813 goto next; 1814 } 1815 1816 return trimmed; 1817 } 1818 1819 static unsigned int __wait_all_discard_cmd(struct f2fs_sb_info *sbi, 1820 struct discard_policy *dpolicy) 1821 { 1822 struct discard_policy dp; 1823 unsigned int discard_blks; 1824 1825 if (dpolicy) 1826 return __wait_discard_cmd_range(sbi, dpolicy, 0, UINT_MAX); 1827 1828 /* wait all */ 1829 __init_discard_policy(sbi, &dp, DPOLICY_FSTRIM, MIN_DISCARD_GRANULARITY); 1830 discard_blks = __wait_discard_cmd_range(sbi, &dp, 0, UINT_MAX); 1831 __init_discard_policy(sbi, &dp, DPOLICY_UMOUNT, MIN_DISCARD_GRANULARITY); 1832 discard_blks += __wait_discard_cmd_range(sbi, &dp, 0, UINT_MAX); 1833 1834 return discard_blks; 1835 } 1836 1837 /* This should be covered by global mutex, &sit_i->sentry_lock */ 1838 static void f2fs_wait_discard_bio(struct f2fs_sb_info *sbi, block_t blkaddr) 1839 { 1840 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1841 struct discard_cmd *dc; 1842 bool need_wait = false; 1843 1844 mutex_lock(&dcc->cmd_lock); 1845 dc = __lookup_discard_cmd(sbi, blkaddr); 1846 #ifdef CONFIG_BLK_DEV_ZONED 1847 if (dc && f2fs_sb_has_blkzoned(sbi) && bdev_is_zoned(dc->bdev)) { 1848 int devi = f2fs_bdev_index(sbi, dc->bdev); 1849 1850 if (devi < 0) { 1851 mutex_unlock(&dcc->cmd_lock); 1852 return; 1853 } 1854 1855 if (f2fs_blkz_is_seq(sbi, devi, dc->di.start)) { 1856 /* force submit zone reset */ 1857 if (dc->state == D_PREP) 1858 __submit_zone_reset_cmd(sbi, dc, REQ_SYNC, 1859 &dcc->wait_list, NULL); 1860 dc->ref++; 1861 mutex_unlock(&dcc->cmd_lock); 1862 /* wait zone reset */ 1863 __wait_one_discard_bio(sbi, dc); 1864 return; 1865 } 1866 } 1867 #endif 1868 if (dc) { 1869 if (dc->state == D_PREP) { 1870 __punch_discard_cmd(sbi, dc, blkaddr); 1871 } else { 1872 dc->ref++; 1873 need_wait = true; 1874 } 1875 } 1876 mutex_unlock(&dcc->cmd_lock); 1877 1878 if (need_wait) 1879 __wait_one_discard_bio(sbi, dc); 1880 } 1881 1882 void f2fs_stop_discard_thread(struct f2fs_sb_info *sbi) 1883 { 1884 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1885 1886 if (dcc && dcc->f2fs_issue_discard) { 1887 struct task_struct *discard_thread = dcc->f2fs_issue_discard; 1888 1889 dcc->f2fs_issue_discard = NULL; 1890 kthread_stop(discard_thread); 1891 } 1892 } 1893 1894 /** 1895 * f2fs_issue_discard_timeout() - Issue all discard cmd within UMOUNT_DISCARD_TIMEOUT 1896 * @sbi: the f2fs_sb_info data for discard cmd to issue 1897 * 1898 * When UMOUNT_DISCARD_TIMEOUT is exceeded, all remaining discard commands will be dropped 1899 * 1900 * Return true if issued all discard cmd or no discard cmd need issue, otherwise return false. 1901 */ 1902 bool f2fs_issue_discard_timeout(struct f2fs_sb_info *sbi, bool need_check) 1903 { 1904 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1905 struct discard_policy dpolicy; 1906 bool dropped; 1907 1908 if (!atomic_read(&dcc->discard_cmd_cnt)) 1909 return true; 1910 1911 __init_discard_policy(sbi, &dpolicy, DPOLICY_UMOUNT, 1912 dcc->discard_granularity); 1913 __issue_discard_cmd(sbi, &dpolicy); 1914 dropped = __drop_discard_cmd(sbi); 1915 1916 /* just to make sure there is no pending discard commands */ 1917 __wait_all_discard_cmd(sbi, NULL); 1918 1919 f2fs_bug_on(sbi, need_check && atomic_read(&dcc->discard_cmd_cnt)); 1920 return !dropped; 1921 } 1922 1923 static int issue_discard_thread(void *data) 1924 { 1925 struct f2fs_sb_info *sbi = data; 1926 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 1927 wait_queue_head_t *q = &dcc->discard_wait_queue; 1928 struct discard_policy dpolicy; 1929 unsigned int wait_ms = dcc->min_discard_issue_time; 1930 int issued; 1931 1932 set_freezable(); 1933 1934 do { 1935 wait_event_freezable_timeout(*q, 1936 kthread_should_stop() || dcc->discard_wake, 1937 msecs_to_jiffies(wait_ms)); 1938 1939 if (sbi->gc_mode == GC_URGENT_HIGH || 1940 !f2fs_available_free_memory(sbi, DISCARD_CACHE)) 1941 __init_discard_policy(sbi, &dpolicy, DPOLICY_FORCE, 1942 MIN_DISCARD_GRANULARITY); 1943 else 1944 __init_discard_policy(sbi, &dpolicy, DPOLICY_BG, 1945 dcc->discard_granularity); 1946 1947 if (dcc->discard_wake) 1948 dcc->discard_wake = false; 1949 1950 /* clean up pending candidates before going to sleep */ 1951 if (atomic_read(&dcc->queued_discard)) 1952 __wait_all_discard_cmd(sbi, NULL); 1953 1954 if (f2fs_readonly(sbi->sb)) 1955 continue; 1956 if (kthread_should_stop()) 1957 return 0; 1958 if (is_sbi_flag_set(sbi, SBI_NEED_FSCK) || 1959 !atomic_read(&dcc->discard_cmd_cnt)) { 1960 wait_ms = dpolicy.max_interval; 1961 continue; 1962 } 1963 1964 sb_start_intwrite(sbi->sb); 1965 1966 issued = __issue_discard_cmd(sbi, &dpolicy); 1967 if (issued > 0) { 1968 __wait_all_discard_cmd(sbi, &dpolicy); 1969 wait_ms = dpolicy.min_interval; 1970 } else if (issued == -1) { 1971 wait_ms = f2fs_time_to_wait(sbi, DISCARD_TIME); 1972 if (!wait_ms) 1973 wait_ms = dpolicy.mid_interval; 1974 } else { 1975 wait_ms = dpolicy.max_interval; 1976 } 1977 if (!atomic_read(&dcc->discard_cmd_cnt)) 1978 wait_ms = dpolicy.max_interval; 1979 1980 sb_end_intwrite(sbi->sb); 1981 1982 } while (!kthread_should_stop()); 1983 return 0; 1984 } 1985 1986 #ifdef CONFIG_BLK_DEV_ZONED 1987 static int __f2fs_issue_discard_zone(struct f2fs_sb_info *sbi, 1988 struct block_device *bdev, block_t blkstart, block_t blklen) 1989 { 1990 sector_t sector, nr_sects; 1991 block_t lblkstart = blkstart; 1992 int devi = 0; 1993 u64 remainder = 0; 1994 1995 if (f2fs_is_multi_device(sbi)) { 1996 devi = f2fs_target_device_index(sbi, blkstart); 1997 if (blkstart < FDEV(devi).start_blk || 1998 blkstart > FDEV(devi).end_blk) { 1999 f2fs_err(sbi, "Invalid block %x", blkstart); 2000 return -EIO; 2001 } 2002 blkstart -= FDEV(devi).start_blk; 2003 } 2004 2005 /* For sequential zones, reset the zone write pointer */ 2006 if (f2fs_blkz_is_seq(sbi, devi, blkstart)) { 2007 sector = SECTOR_FROM_BLOCK(blkstart); 2008 nr_sects = SECTOR_FROM_BLOCK(blklen); 2009 div64_u64_rem(sector, bdev_zone_sectors(bdev), &remainder); 2010 2011 if (remainder || nr_sects != bdev_zone_sectors(bdev)) { 2012 f2fs_err(sbi, "(%d) %s: Unaligned zone reset attempted (block %x + %x)", 2013 devi, sbi->s_ndevs ? FDEV(devi).path : "", 2014 blkstart, blklen); 2015 return -EIO; 2016 } 2017 2018 if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING))) { 2019 unsigned int nofs_flags; 2020 int ret; 2021 2022 trace_f2fs_issue_reset_zone(bdev, blkstart); 2023 nofs_flags = memalloc_nofs_save(); 2024 ret = blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET, 2025 sector, nr_sects); 2026 memalloc_nofs_restore(nofs_flags); 2027 return ret; 2028 } 2029 2030 __queue_zone_reset_cmd(sbi, bdev, blkstart, lblkstart, blklen); 2031 return 0; 2032 } 2033 2034 /* For conventional zones, use regular discard if supported */ 2035 __queue_discard_cmd(sbi, bdev, lblkstart, blklen); 2036 return 0; 2037 } 2038 #endif 2039 2040 static int __issue_discard_async(struct f2fs_sb_info *sbi, 2041 struct block_device *bdev, block_t blkstart, block_t blklen) 2042 { 2043 #ifdef CONFIG_BLK_DEV_ZONED 2044 if (f2fs_sb_has_blkzoned(sbi) && bdev_is_zoned(bdev)) 2045 return __f2fs_issue_discard_zone(sbi, bdev, blkstart, blklen); 2046 #endif 2047 __queue_discard_cmd(sbi, bdev, blkstart, blklen); 2048 return 0; 2049 } 2050 2051 static int f2fs_issue_discard(struct f2fs_sb_info *sbi, 2052 block_t blkstart, block_t blklen) 2053 { 2054 sector_t start = blkstart, len = 0; 2055 struct block_device *bdev; 2056 struct seg_entry *se; 2057 unsigned int offset; 2058 block_t i; 2059 int err = 0; 2060 2061 bdev = f2fs_target_device(sbi, blkstart, NULL); 2062 2063 for (i = blkstart; i < blkstart + blklen; i++, len++) { 2064 if (i != start) { 2065 struct block_device *bdev2 = 2066 f2fs_target_device(sbi, i, NULL); 2067 2068 if (bdev2 != bdev) { 2069 err = __issue_discard_async(sbi, bdev, 2070 start, len); 2071 if (err) 2072 return err; 2073 bdev = bdev2; 2074 start = i; 2075 len = 0; 2076 } 2077 } 2078 2079 se = get_seg_entry(sbi, GET_SEGNO(sbi, i)); 2080 offset = GET_BLKOFF_FROM_SEG0(sbi, i); 2081 2082 if (f2fs_block_unit_discard(sbi) && 2083 !f2fs_test_and_set_bit(offset, se->discard_map)) 2084 sbi->discard_blks--; 2085 } 2086 2087 if (len) 2088 err = __issue_discard_async(sbi, bdev, start, len); 2089 return err; 2090 } 2091 2092 static bool add_discard_addrs(struct f2fs_sb_info *sbi, struct cp_control *cpc, 2093 bool check_only) 2094 { 2095 int entries = SIT_VBLOCK_MAP_SIZE / sizeof(unsigned long); 2096 struct seg_entry *se = get_seg_entry(sbi, cpc->trim_start); 2097 unsigned long *cur_map = (unsigned long *)se->cur_valid_map; 2098 unsigned long *ckpt_map = (unsigned long *)se->ckpt_valid_map; 2099 unsigned long *discard_map = (unsigned long *)se->discard_map; 2100 unsigned long *dmap = SIT_I(sbi)->tmp_map; 2101 unsigned int start = 0, end = -1; 2102 bool force = (cpc->reason & CP_DISCARD); 2103 struct discard_entry *de = NULL; 2104 struct list_head *head = &SM_I(sbi)->dcc_info->entry_list; 2105 int i; 2106 2107 if (se->valid_blocks == BLKS_PER_SEG(sbi) || 2108 !f2fs_hw_support_discard(sbi) || 2109 !f2fs_block_unit_discard(sbi)) 2110 return false; 2111 2112 if (!force) { 2113 if (!f2fs_realtime_discard_enable(sbi) || 2114 (!se->valid_blocks && 2115 !is_curseg(sbi, cpc->trim_start)) || 2116 SM_I(sbi)->dcc_info->nr_discards >= 2117 SM_I(sbi)->dcc_info->max_discards) 2118 return false; 2119 } 2120 2121 /* SIT_VBLOCK_MAP_SIZE should be multiple of sizeof(unsigned long) */ 2122 for (i = 0; i < entries; i++) 2123 dmap[i] = force ? ~ckpt_map[i] & ~discard_map[i] : 2124 (cur_map[i] ^ ckpt_map[i]) & ckpt_map[i]; 2125 2126 while (force || SM_I(sbi)->dcc_info->nr_discards <= 2127 SM_I(sbi)->dcc_info->max_discards) { 2128 start = __find_rev_next_bit(dmap, BLKS_PER_SEG(sbi), end + 1); 2129 if (start >= BLKS_PER_SEG(sbi)) 2130 break; 2131 2132 end = __find_rev_next_zero_bit(dmap, 2133 BLKS_PER_SEG(sbi), start + 1); 2134 if (force && start && end != BLKS_PER_SEG(sbi) && 2135 (end - start) < cpc->trim_minlen) 2136 continue; 2137 2138 if (check_only) 2139 return true; 2140 2141 if (!de) { 2142 de = f2fs_kmem_cache_alloc(discard_entry_slab, 2143 GFP_F2FS_ZERO, true, NULL); 2144 de->start_blkaddr = START_BLOCK(sbi, cpc->trim_start); 2145 list_add_tail(&de->list, head); 2146 } 2147 2148 for (i = start; i < end; i++) 2149 __set_bit_le(i, (void *)de->discard_map); 2150 2151 SM_I(sbi)->dcc_info->nr_discards += end - start; 2152 } 2153 return false; 2154 } 2155 2156 static void release_discard_addr(struct discard_entry *entry) 2157 { 2158 list_del(&entry->list); 2159 kmem_cache_free(discard_entry_slab, entry); 2160 } 2161 2162 void f2fs_release_discard_addrs(struct f2fs_sb_info *sbi) 2163 { 2164 struct list_head *head = &(SM_I(sbi)->dcc_info->entry_list); 2165 struct discard_entry *entry, *this; 2166 2167 /* drop caches */ 2168 list_for_each_entry_safe(entry, this, head, list) 2169 release_discard_addr(entry); 2170 } 2171 2172 /* 2173 * Should call f2fs_clear_prefree_segments after checkpoint is done. 2174 */ 2175 static void set_prefree_as_free_segments(struct f2fs_sb_info *sbi) 2176 { 2177 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 2178 unsigned int segno; 2179 2180 mutex_lock(&dirty_i->seglist_lock); 2181 for_each_set_bit(segno, dirty_i->dirty_segmap[PRE], MAIN_SEGS(sbi)) 2182 __set_test_and_free(sbi, segno, false); 2183 mutex_unlock(&dirty_i->seglist_lock); 2184 } 2185 2186 void f2fs_clear_prefree_segments(struct f2fs_sb_info *sbi, 2187 struct cp_control *cpc) 2188 { 2189 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 2190 struct list_head *head = &dcc->entry_list; 2191 struct discard_entry *entry, *this; 2192 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 2193 unsigned long *prefree_map = dirty_i->dirty_segmap[PRE]; 2194 unsigned int start = 0, end = -1; 2195 unsigned int secno, start_segno; 2196 bool force = (cpc->reason & CP_DISCARD); 2197 bool section_alignment = F2FS_OPTION(sbi).discard_unit == 2198 DISCARD_UNIT_SECTION; 2199 2200 if (f2fs_lfs_mode(sbi) && __is_large_section(sbi)) 2201 section_alignment = true; 2202 2203 mutex_lock(&dirty_i->seglist_lock); 2204 2205 while (1) { 2206 int i; 2207 2208 if (section_alignment && end != -1) 2209 end--; 2210 start = find_next_bit(prefree_map, MAIN_SEGS(sbi), end + 1); 2211 if (start >= MAIN_SEGS(sbi)) 2212 break; 2213 end = find_next_zero_bit(prefree_map, MAIN_SEGS(sbi), 2214 start + 1); 2215 2216 if (section_alignment) { 2217 start = rounddown(start, SEGS_PER_SEC(sbi)); 2218 end = roundup(end, SEGS_PER_SEC(sbi)); 2219 } 2220 2221 for (i = start; i < end; i++) { 2222 if (test_and_clear_bit(i, prefree_map)) 2223 dirty_i->nr_dirty[PRE]--; 2224 } 2225 2226 if (!f2fs_realtime_discard_enable(sbi)) 2227 continue; 2228 2229 if (force && start >= cpc->trim_start && 2230 (end - 1) <= cpc->trim_end) 2231 continue; 2232 2233 /* Should cover 2MB zoned device for zone-based reset */ 2234 if (!f2fs_sb_has_blkzoned(sbi) && 2235 (!f2fs_lfs_mode(sbi) || !__is_large_section(sbi))) { 2236 f2fs_issue_discard(sbi, START_BLOCK(sbi, start), 2237 SEGS_TO_BLKS(sbi, end - start)); 2238 continue; 2239 } 2240 next: 2241 secno = GET_SEC_FROM_SEG(sbi, start); 2242 start_segno = GET_SEG_FROM_SEC(sbi, secno); 2243 if (!is_cursec(sbi, secno) && 2244 !get_valid_blocks(sbi, start, true)) 2245 f2fs_issue_discard(sbi, START_BLOCK(sbi, start_segno), 2246 BLKS_PER_SEC(sbi)); 2247 2248 start = start_segno + SEGS_PER_SEC(sbi); 2249 if (start < end) 2250 goto next; 2251 else 2252 end = start - 1; 2253 } 2254 mutex_unlock(&dirty_i->seglist_lock); 2255 2256 if (!f2fs_block_unit_discard(sbi)) 2257 goto wakeup; 2258 2259 /* send small discards */ 2260 list_for_each_entry_safe(entry, this, head, list) { 2261 unsigned int cur_pos = 0, next_pos, len, total_len = 0; 2262 bool is_valid = test_bit_le(0, entry->discard_map); 2263 2264 find_next: 2265 if (is_valid) { 2266 next_pos = find_next_zero_bit_le(entry->discard_map, 2267 BLKS_PER_SEG(sbi), cur_pos); 2268 len = next_pos - cur_pos; 2269 2270 if (f2fs_sb_has_blkzoned(sbi) || 2271 (force && len < cpc->trim_minlen)) 2272 goto skip; 2273 2274 f2fs_issue_discard(sbi, entry->start_blkaddr + cur_pos, 2275 len); 2276 total_len += len; 2277 } else { 2278 next_pos = find_next_bit_le(entry->discard_map, 2279 BLKS_PER_SEG(sbi), cur_pos); 2280 } 2281 skip: 2282 cur_pos = next_pos; 2283 is_valid = !is_valid; 2284 2285 if (cur_pos < BLKS_PER_SEG(sbi)) 2286 goto find_next; 2287 2288 release_discard_addr(entry); 2289 dcc->nr_discards -= total_len; 2290 } 2291 2292 wakeup: 2293 wake_up_discard_thread(sbi, false); 2294 } 2295 2296 int f2fs_start_discard_thread(struct f2fs_sb_info *sbi) 2297 { 2298 dev_t dev = sbi->sb->s_bdev->bd_dev; 2299 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 2300 int err = 0; 2301 2302 if (f2fs_sb_has_readonly(sbi)) { 2303 f2fs_info(sbi, 2304 "Skip to start discard thread for readonly image"); 2305 return 0; 2306 } 2307 2308 if (!f2fs_realtime_discard_enable(sbi)) 2309 return 0; 2310 2311 dcc->f2fs_issue_discard = kthread_run(issue_discard_thread, sbi, 2312 "f2fs_discard-%u:%u", MAJOR(dev), MINOR(dev)); 2313 if (IS_ERR(dcc->f2fs_issue_discard)) { 2314 err = PTR_ERR(dcc->f2fs_issue_discard); 2315 dcc->f2fs_issue_discard = NULL; 2316 } 2317 2318 return err; 2319 } 2320 2321 static int create_discard_cmd_control(struct f2fs_sb_info *sbi) 2322 { 2323 struct discard_cmd_control *dcc; 2324 int err = 0, i; 2325 2326 if (SM_I(sbi)->dcc_info) { 2327 dcc = SM_I(sbi)->dcc_info; 2328 goto init_thread; 2329 } 2330 2331 dcc = f2fs_kzalloc(sbi, sizeof(struct discard_cmd_control), GFP_KERNEL); 2332 if (!dcc) 2333 return -ENOMEM; 2334 2335 dcc->discard_io_aware_gran = MAX_PLIST_NUM; 2336 dcc->discard_granularity = DEFAULT_DISCARD_GRANULARITY; 2337 dcc->max_ordered_discard = DEFAULT_MAX_ORDERED_DISCARD_GRANULARITY; 2338 dcc->discard_io_aware = DPOLICY_IO_AWARE_ENABLE; 2339 if (F2FS_OPTION(sbi).discard_unit == DISCARD_UNIT_SEGMENT || 2340 F2FS_OPTION(sbi).discard_unit == DISCARD_UNIT_SECTION) 2341 dcc->discard_granularity = BLKS_PER_SEG(sbi); 2342 2343 INIT_LIST_HEAD(&dcc->entry_list); 2344 for (i = 0; i < MAX_PLIST_NUM; i++) 2345 INIT_LIST_HEAD(&dcc->pend_list[i]); 2346 INIT_LIST_HEAD(&dcc->wait_list); 2347 INIT_LIST_HEAD(&dcc->fstrim_list); 2348 mutex_init(&dcc->cmd_lock); 2349 atomic_set(&dcc->issued_discard, 0); 2350 atomic_set(&dcc->queued_discard, 0); 2351 atomic_set(&dcc->discard_cmd_cnt, 0); 2352 dcc->nr_discards = 0; 2353 dcc->max_discards = SEGS_TO_BLKS(sbi, MAIN_SEGS(sbi)); 2354 dcc->max_discard_request = DEF_MAX_DISCARD_REQUEST; 2355 dcc->min_discard_issue_time = DEF_MIN_DISCARD_ISSUE_TIME; 2356 dcc->mid_discard_issue_time = DEF_MID_DISCARD_ISSUE_TIME; 2357 dcc->max_discard_issue_time = DEF_MAX_DISCARD_ISSUE_TIME; 2358 dcc->discard_urgent_util = DEF_DISCARD_URGENT_UTIL; 2359 dcc->undiscard_blks = 0; 2360 dcc->next_pos = 0; 2361 dcc->root = RB_ROOT_CACHED; 2362 dcc->rbtree_check = false; 2363 2364 init_waitqueue_head(&dcc->discard_wait_queue); 2365 SM_I(sbi)->dcc_info = dcc; 2366 init_thread: 2367 err = f2fs_start_discard_thread(sbi); 2368 if (err) { 2369 kfree(dcc); 2370 SM_I(sbi)->dcc_info = NULL; 2371 } 2372 2373 return err; 2374 } 2375 2376 static void destroy_discard_cmd_control(struct f2fs_sb_info *sbi) 2377 { 2378 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 2379 2380 if (!dcc) 2381 return; 2382 2383 f2fs_stop_discard_thread(sbi); 2384 2385 /* 2386 * Recovery can cache discard commands, so in error path of 2387 * fill_super(), it needs to give a chance to handle them. 2388 */ 2389 f2fs_issue_discard_timeout(sbi, true); 2390 2391 kfree(dcc); 2392 SM_I(sbi)->dcc_info = NULL; 2393 } 2394 2395 static bool __mark_sit_entry_dirty(struct f2fs_sb_info *sbi, unsigned int segno) 2396 { 2397 struct sit_info *sit_i = SIT_I(sbi); 2398 2399 if (!__test_and_set_bit(segno, sit_i->dirty_sentries_bitmap)) { 2400 sit_i->dirty_sentries++; 2401 return false; 2402 } 2403 2404 return true; 2405 } 2406 2407 static void __set_sit_entry_type(struct f2fs_sb_info *sbi, int type, 2408 unsigned int segno, int modified) 2409 { 2410 struct seg_entry *se = get_seg_entry(sbi, segno); 2411 2412 se->type = type; 2413 if (modified) 2414 __mark_sit_entry_dirty(sbi, segno); 2415 } 2416 2417 static inline unsigned long long get_segment_mtime(struct f2fs_sb_info *sbi, 2418 block_t blkaddr) 2419 { 2420 unsigned int segno = GET_SEGNO(sbi, blkaddr); 2421 2422 if (segno == NULL_SEGNO) 2423 return 0; 2424 return get_seg_entry(sbi, segno)->mtime; 2425 } 2426 2427 static void update_segment_mtime(struct f2fs_sb_info *sbi, block_t blkaddr, 2428 unsigned long long old_mtime) 2429 { 2430 struct seg_entry *se; 2431 unsigned int segno = GET_SEGNO(sbi, blkaddr); 2432 unsigned long long ctime = get_mtime(sbi, false); 2433 unsigned long long mtime = old_mtime ? old_mtime : ctime; 2434 2435 if (segno == NULL_SEGNO) 2436 return; 2437 2438 se = get_seg_entry(sbi, segno); 2439 2440 if (!se->mtime) 2441 se->mtime = mtime; 2442 else 2443 se->mtime = div_u64(se->mtime * se->valid_blocks + mtime, 2444 se->valid_blocks + 1); 2445 2446 if (ctime > SIT_I(sbi)->max_mtime) 2447 SIT_I(sbi)->max_mtime = ctime; 2448 } 2449 2450 /* 2451 * NOTE: when updating multiple blocks at the same time, please ensure 2452 * that the consecutive input blocks belong to the same segment. 2453 */ 2454 static int update_sit_entry_for_release(struct f2fs_sb_info *sbi, struct seg_entry *se, 2455 unsigned int segno, block_t blkaddr, unsigned int offset, int del) 2456 { 2457 bool exist; 2458 int i; 2459 int del_count = -del; 2460 2461 f2fs_bug_on(sbi, GET_SEGNO(sbi, blkaddr) != GET_SEGNO(sbi, blkaddr + del_count - 1)); 2462 2463 for (i = 0; i < del_count; i++) { 2464 exist = f2fs_test_and_clear_bit(offset + i, se->cur_valid_map); 2465 if (unlikely(!exist)) { 2466 f2fs_err(sbi, "Bitmap was wrongly cleared, blk:%u", blkaddr + i); 2467 f2fs_bug_on(sbi, 1); 2468 se->valid_blocks++; 2469 del += 1; 2470 } else if (unlikely(is_sbi_flag_set(sbi, SBI_CP_DISABLED))) { 2471 /* 2472 * If checkpoints are off, we must not reuse data that 2473 * was used in the previous checkpoint. If it was used 2474 * before, we must track that to know how much space we 2475 * really have. 2476 */ 2477 if (f2fs_test_bit(offset + i, se->ckpt_valid_map)) { 2478 spin_lock(&sbi->stat_lock); 2479 sbi->unusable_block_count++; 2480 spin_unlock(&sbi->stat_lock); 2481 } 2482 } 2483 2484 if (f2fs_block_unit_discard(sbi) && 2485 f2fs_test_and_clear_bit(offset + i, se->discard_map)) 2486 sbi->discard_blks++; 2487 2488 if (!f2fs_test_bit(offset + i, se->ckpt_valid_map)) { 2489 se->ckpt_valid_blocks -= 1; 2490 if (__is_large_section(sbi)) 2491 get_sec_entry(sbi, segno)->ckpt_valid_blocks -= 1; 2492 } 2493 } 2494 2495 if (__is_large_section(sbi)) 2496 sanity_check_valid_blocks(sbi, segno); 2497 2498 return del; 2499 } 2500 2501 static int update_sit_entry_for_alloc(struct f2fs_sb_info *sbi, struct seg_entry *se, 2502 unsigned int segno, block_t blkaddr, unsigned int offset, int del) 2503 { 2504 bool exist; 2505 2506 exist = f2fs_test_and_set_bit(offset, se->cur_valid_map); 2507 if (unlikely(exist)) { 2508 f2fs_err(sbi, "Bitmap was wrongly set, blk:%u", blkaddr); 2509 f2fs_bug_on(sbi, 1); 2510 se->valid_blocks--; 2511 del = 0; 2512 } 2513 2514 if (f2fs_block_unit_discard(sbi) && 2515 !f2fs_test_and_set_bit(offset, se->discard_map)) 2516 sbi->discard_blks--; 2517 2518 /* 2519 * SSR should never reuse block which is checkpointed 2520 * or newly invalidated. 2521 */ 2522 if (!is_sbi_flag_set(sbi, SBI_CP_DISABLED)) { 2523 if (!f2fs_test_and_set_bit(offset, se->ckpt_valid_map)) { 2524 se->ckpt_valid_blocks++; 2525 if (__is_large_section(sbi)) 2526 get_sec_entry(sbi, segno)->ckpt_valid_blocks++; 2527 } 2528 } 2529 2530 if (!f2fs_test_bit(offset, se->ckpt_valid_map)) { 2531 se->ckpt_valid_blocks += del; 2532 if (__is_large_section(sbi)) 2533 get_sec_entry(sbi, segno)->ckpt_valid_blocks += del; 2534 } 2535 2536 if (__is_large_section(sbi)) 2537 sanity_check_valid_blocks(sbi, segno); 2538 2539 return del; 2540 } 2541 2542 /* 2543 * If releasing blocks, this function supports updating multiple consecutive blocks 2544 * at one time, but please note that these consecutive blocks need to belong to the 2545 * same segment. 2546 */ 2547 static void update_sit_entry(struct f2fs_sb_info *sbi, block_t blkaddr, int del) 2548 { 2549 struct seg_entry *se; 2550 unsigned int segno, offset; 2551 long int new_vblocks; 2552 2553 segno = GET_SEGNO(sbi, blkaddr); 2554 if (segno == NULL_SEGNO) 2555 return; 2556 2557 se = get_seg_entry(sbi, segno); 2558 new_vblocks = se->valid_blocks + del; 2559 offset = GET_BLKOFF_FROM_SEG0(sbi, blkaddr); 2560 2561 f2fs_bug_on(sbi, (new_vblocks < 0 || 2562 (new_vblocks > f2fs_usable_blks_in_seg(sbi, segno)))); 2563 2564 se->valid_blocks = new_vblocks; 2565 2566 /* Update valid block bitmap */ 2567 if (del > 0) { 2568 del = update_sit_entry_for_alloc(sbi, se, segno, blkaddr, offset, del); 2569 } else { 2570 del = update_sit_entry_for_release(sbi, se, segno, blkaddr, offset, del); 2571 } 2572 2573 __mark_sit_entry_dirty(sbi, segno); 2574 2575 /* update total number of valid blocks to be written in ckpt area */ 2576 SIT_I(sbi)->written_valid_blocks += del; 2577 2578 if (__is_large_section(sbi)) 2579 get_sec_entry(sbi, segno)->valid_blocks += del; 2580 } 2581 2582 void f2fs_invalidate_blocks(struct f2fs_sb_info *sbi, block_t addr, 2583 unsigned int len) 2584 { 2585 unsigned int segno = GET_SEGNO(sbi, addr); 2586 struct sit_info *sit_i = SIT_I(sbi); 2587 block_t addr_start = addr, addr_end = addr + len - 1; 2588 unsigned int seg_num = GET_SEGNO(sbi, addr_end) - segno + 1; 2589 unsigned int i = 1, max_blocks = sbi->blocks_per_seg, cnt; 2590 2591 f2fs_bug_on(sbi, addr == NULL_ADDR); 2592 if (addr == NEW_ADDR || addr == COMPRESS_ADDR) 2593 return; 2594 2595 f2fs_invalidate_internal_cache(sbi, addr, len); 2596 2597 /* add it into sit main buffer */ 2598 down_write(&sit_i->sentry_lock); 2599 2600 if (seg_num == 1) 2601 cnt = len; 2602 else 2603 cnt = max_blocks - GET_BLKOFF_FROM_SEG0(sbi, addr); 2604 2605 do { 2606 update_segment_mtime(sbi, addr_start, 0); 2607 update_sit_entry(sbi, addr_start, -cnt); 2608 2609 /* add it into dirty seglist */ 2610 locate_dirty_segment(sbi, segno); 2611 2612 /* update @addr_start and @cnt and @segno */ 2613 addr_start = START_BLOCK(sbi, ++segno); 2614 if (++i == seg_num) 2615 cnt = GET_BLKOFF_FROM_SEG0(sbi, addr_end) + 1; 2616 else 2617 cnt = max_blocks; 2618 } while (i <= seg_num); 2619 2620 up_write(&sit_i->sentry_lock); 2621 } 2622 2623 bool f2fs_is_checkpointed_data(struct f2fs_sb_info *sbi, block_t blkaddr) 2624 { 2625 struct sit_info *sit_i = SIT_I(sbi); 2626 unsigned int segno, offset; 2627 struct seg_entry *se; 2628 bool is_cp = false; 2629 2630 if (!__is_valid_data_blkaddr(blkaddr)) 2631 return true; 2632 2633 down_read(&sit_i->sentry_lock); 2634 2635 segno = GET_SEGNO(sbi, blkaddr); 2636 se = get_seg_entry(sbi, segno); 2637 offset = GET_BLKOFF_FROM_SEG0(sbi, blkaddr); 2638 2639 if (f2fs_test_bit(offset, se->ckpt_valid_map)) 2640 is_cp = true; 2641 2642 up_read(&sit_i->sentry_lock); 2643 2644 return is_cp; 2645 } 2646 2647 static unsigned short f2fs_curseg_valid_blocks(struct f2fs_sb_info *sbi, int type) 2648 { 2649 struct curseg_info *curseg = CURSEG_I(sbi, type); 2650 2651 if (sbi->ckpt->alloc_type[type] == SSR) 2652 return BLKS_PER_SEG(sbi); 2653 return curseg->next_blkoff; 2654 } 2655 2656 /* 2657 * Calculate the number of current summary pages for writing 2658 */ 2659 int f2fs_npages_for_summary_flush(struct f2fs_sb_info *sbi, bool for_ra) 2660 { 2661 int valid_sum_count = 0; 2662 int i, sum_in_page; 2663 2664 for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) { 2665 if (sbi->ckpt->alloc_type[i] != SSR && for_ra) 2666 valid_sum_count += 2667 le16_to_cpu(F2FS_CKPT(sbi)->cur_data_blkoff[i]); 2668 else 2669 valid_sum_count += f2fs_curseg_valid_blocks(sbi, i); 2670 } 2671 2672 sum_in_page = (sbi->blocksize - 2 * sbi->sum_journal_size - 2673 SUM_FOOTER_SIZE) / SUMMARY_SIZE; 2674 if (valid_sum_count <= sum_in_page) 2675 return 1; 2676 else if ((valid_sum_count - sum_in_page) <= 2677 (sbi->blocksize - SUM_FOOTER_SIZE) / SUMMARY_SIZE) 2678 return 2; 2679 return 3; 2680 } 2681 2682 /* 2683 * Caller should put this summary folio 2684 */ 2685 struct folio *f2fs_get_sum_folio(struct f2fs_sb_info *sbi, unsigned int segno) 2686 { 2687 if (unlikely(f2fs_cp_error(sbi))) 2688 return ERR_PTR(-EIO); 2689 return f2fs_get_meta_folio_retry(sbi, GET_SUM_BLOCK(sbi, segno)); 2690 } 2691 2692 void f2fs_update_meta_page(struct f2fs_sb_info *sbi, 2693 void *src, block_t blk_addr) 2694 { 2695 struct folio *folio; 2696 2697 if (!f2fs_sb_has_packed_ssa(sbi)) 2698 folio = f2fs_grab_meta_folio(sbi, blk_addr); 2699 else 2700 folio = f2fs_get_meta_folio_retry(sbi, blk_addr); 2701 2702 if (IS_ERR(folio)) 2703 return; 2704 2705 memcpy(folio_address(folio), src, PAGE_SIZE); 2706 folio_mark_dirty(folio); 2707 f2fs_folio_put(folio, true); 2708 } 2709 2710 static void write_sum_page(struct f2fs_sb_info *sbi, 2711 struct f2fs_summary_block *sum_blk, unsigned int segno) 2712 { 2713 struct folio *folio; 2714 2715 if (!f2fs_sb_has_packed_ssa(sbi)) 2716 return f2fs_update_meta_page(sbi, (void *)sum_blk, 2717 GET_SUM_BLOCK(sbi, segno)); 2718 2719 folio = f2fs_get_sum_folio(sbi, segno); 2720 if (IS_ERR(folio)) 2721 return; 2722 2723 memcpy(SUM_BLK_PAGE_ADDR(sbi, folio, segno), sum_blk, 2724 sbi->sum_blocksize); 2725 folio_mark_dirty(folio); 2726 f2fs_folio_put(folio, true); 2727 } 2728 2729 static void write_current_sum_page(struct f2fs_sb_info *sbi, 2730 int type, block_t blk_addr) 2731 { 2732 struct curseg_info *curseg = CURSEG_I(sbi, type); 2733 struct folio *folio = f2fs_grab_meta_folio(sbi, blk_addr); 2734 struct f2fs_summary_block *src = curseg->sum_blk; 2735 struct f2fs_summary_block *dst; 2736 2737 dst = folio_address(folio); 2738 memset(dst, 0, PAGE_SIZE); 2739 2740 mutex_lock(&curseg->curseg_mutex); 2741 2742 down_read(&curseg->journal_rwsem); 2743 memcpy(sum_journal(sbi, dst), curseg->journal, sbi->sum_journal_size); 2744 up_read(&curseg->journal_rwsem); 2745 2746 memcpy(sum_entries(dst), sum_entries(src), sbi->sum_entry_size); 2747 memcpy(sum_footer(sbi, dst), sum_footer(sbi, src), SUM_FOOTER_SIZE); 2748 2749 mutex_unlock(&curseg->curseg_mutex); 2750 2751 folio_mark_dirty(folio); 2752 f2fs_folio_put(folio, true); 2753 } 2754 2755 static int is_next_segment_free(struct f2fs_sb_info *sbi, 2756 struct curseg_info *curseg) 2757 { 2758 unsigned int segno = curseg->segno + 1; 2759 struct free_segmap_info *free_i = FREE_I(sbi); 2760 2761 if (segno < MAIN_SEGS(sbi) && segno % SEGS_PER_SEC(sbi)) 2762 return !test_bit(segno, free_i->free_segmap); 2763 return 0; 2764 } 2765 2766 /* 2767 * Find a new segment from the free segments bitmap to right order 2768 * This function should be returned with success, otherwise BUG 2769 */ 2770 static int get_new_segment(struct f2fs_sb_info *sbi, 2771 unsigned int *newseg, bool new_sec, bool pinning) 2772 { 2773 struct free_segmap_info *free_i = FREE_I(sbi); 2774 unsigned int segno, secno, zoneno; 2775 unsigned int total_zones = MAIN_SECS(sbi) / sbi->secs_per_zone; 2776 unsigned int hint = GET_SEC_FROM_SEG(sbi, *newseg); 2777 unsigned int old_zoneno = GET_ZONE_FROM_SEG(sbi, *newseg); 2778 unsigned int alloc_policy = sbi->allocate_section_policy; 2779 unsigned int alloc_hint = sbi->allocate_section_hint; 2780 bool init = true; 2781 int i; 2782 int ret = 0; 2783 2784 spin_lock(&free_i->segmap_lock); 2785 2786 if (time_to_inject(sbi, FAULT_NO_SEGMENT)) { 2787 ret = -ENOSPC; 2788 goto out_unlock; 2789 } 2790 2791 if (!new_sec && ((*newseg + 1) % SEGS_PER_SEC(sbi))) { 2792 segno = find_next_zero_bit(free_i->free_segmap, 2793 GET_SEG_FROM_SEC(sbi, hint + 1), *newseg + 1); 2794 if (segno < GET_SEG_FROM_SEC(sbi, hint + 1)) 2795 goto got_it; 2796 } 2797 2798 #ifdef CONFIG_BLK_DEV_ZONED 2799 /* 2800 * If we format f2fs on zoned storage, let's try to get pinned sections 2801 * from beginning of the storage, which should be a conventional one. 2802 */ 2803 if (f2fs_sb_has_blkzoned(sbi)) { 2804 /* Prioritize writing to conventional zones */ 2805 if (sbi->blkzone_alloc_policy == BLKZONE_ALLOC_PRIOR_CONV || pinning) 2806 segno = 0; 2807 else 2808 segno = max(sbi->first_seq_zone_segno, *newseg); 2809 hint = GET_SEC_FROM_SEG(sbi, segno); 2810 } 2811 #endif 2812 2813 /* 2814 * Prevent allocate_section_hint from exceeding MAIN_SECS() 2815 * due to desynchronization. 2816 */ 2817 if (alloc_policy != ALLOCATE_FORWARD_NOHINT && 2818 alloc_hint > MAIN_SECS(sbi)) 2819 alloc_hint = MAIN_SECS(sbi); 2820 2821 if (alloc_policy == ALLOCATE_FORWARD_FROM_HINT && 2822 hint < alloc_hint) 2823 hint = alloc_hint; 2824 else if (alloc_policy == ALLOCATE_FORWARD_WITHIN_HINT && 2825 hint >= alloc_hint) 2826 hint = 0; 2827 2828 find_other_zone: 2829 secno = find_next_zero_bit(free_i->free_secmap, MAIN_SECS(sbi), hint); 2830 2831 #ifdef CONFIG_BLK_DEV_ZONED 2832 if (secno >= MAIN_SECS(sbi) && f2fs_sb_has_blkzoned(sbi)) { 2833 /* Write only to sequential zones */ 2834 if (sbi->blkzone_alloc_policy == BLKZONE_ALLOC_ONLY_SEQ) { 2835 hint = GET_SEC_FROM_SEG(sbi, sbi->first_seq_zone_segno); 2836 secno = find_next_zero_bit(free_i->free_secmap, MAIN_SECS(sbi), hint); 2837 } else 2838 secno = find_first_zero_bit(free_i->free_secmap, 2839 MAIN_SECS(sbi)); 2840 if (secno >= MAIN_SECS(sbi)) { 2841 ret = -ENOSPC; 2842 f2fs_bug_on(sbi, 1); 2843 goto out_unlock; 2844 } 2845 } 2846 #endif 2847 2848 if (secno >= MAIN_SECS(sbi)) { 2849 secno = find_first_zero_bit(free_i->free_secmap, 2850 MAIN_SECS(sbi)); 2851 if (secno >= MAIN_SECS(sbi)) { 2852 ret = -ENOSPC; 2853 f2fs_bug_on(sbi, !pinning); 2854 goto out_unlock; 2855 } 2856 } 2857 segno = GET_SEG_FROM_SEC(sbi, secno); 2858 zoneno = GET_ZONE_FROM_SEC(sbi, secno); 2859 2860 /* give up on finding another zone */ 2861 if (!init) 2862 goto got_it; 2863 if (sbi->secs_per_zone == 1) 2864 goto got_it; 2865 if (zoneno == old_zoneno) 2866 goto got_it; 2867 for (i = 0; i < NR_CURSEG_TYPE; i++) 2868 if (CURSEG_I(sbi, i)->zone == zoneno) 2869 break; 2870 2871 if (i < NR_CURSEG_TYPE) { 2872 /* zone is in user, try another */ 2873 if (zoneno + 1 >= total_zones) 2874 hint = 0; 2875 else 2876 hint = (zoneno + 1) * sbi->secs_per_zone; 2877 init = false; 2878 goto find_other_zone; 2879 } 2880 got_it: 2881 /* set it as dirty segment in free segmap */ 2882 if (test_bit(segno, free_i->free_segmap)) { 2883 ret = -EFSCORRUPTED; 2884 fserror_report_metadata(sbi->sb, -EFSCORRUPTED, GFP_NOFS); 2885 f2fs_stop_checkpoint(sbi, false, STOP_CP_REASON_CORRUPTED_FREE_BITMAP); 2886 goto out_unlock; 2887 } 2888 2889 /* no free section in conventional device or conventional zone */ 2890 if (new_sec && pinning && 2891 f2fs_is_sequential_zone_area(sbi, START_BLOCK(sbi, segno))) { 2892 ret = -EAGAIN; 2893 goto out_unlock; 2894 } 2895 __set_inuse(sbi, segno); 2896 *newseg = segno; 2897 out_unlock: 2898 spin_unlock(&free_i->segmap_lock); 2899 2900 if (ret == -ENOSPC && !pinning) 2901 f2fs_stop_checkpoint(sbi, false, STOP_CP_REASON_NO_SEGMENT); 2902 return ret; 2903 } 2904 2905 static void reset_curseg(struct f2fs_sb_info *sbi, int type, int modified) 2906 { 2907 struct curseg_info *curseg = CURSEG_I(sbi, type); 2908 struct summary_footer *sum_footer; 2909 unsigned short seg_type = curseg->seg_type; 2910 2911 /* only happen when get_new_segment() fails */ 2912 if (curseg->next_segno == NULL_SEGNO) 2913 return; 2914 2915 curseg->inited = true; 2916 curseg->segno = curseg->next_segno; 2917 curseg->zone = GET_ZONE_FROM_SEG(sbi, curseg->segno); 2918 curseg->next_blkoff = 0; 2919 curseg->next_segno = NULL_SEGNO; 2920 2921 sum_footer = sum_footer(sbi, curseg->sum_blk); 2922 memset(sum_footer, 0, sizeof(struct summary_footer)); 2923 2924 sanity_check_seg_type(sbi, seg_type); 2925 2926 if (IS_DATASEG(seg_type)) 2927 SET_SUM_TYPE(sum_footer, SUM_TYPE_DATA); 2928 if (IS_NODESEG(seg_type)) 2929 SET_SUM_TYPE(sum_footer, SUM_TYPE_NODE); 2930 __set_sit_entry_type(sbi, seg_type, curseg->segno, modified); 2931 } 2932 2933 static unsigned int __get_next_segno(struct f2fs_sb_info *sbi, int type) 2934 { 2935 struct curseg_info *curseg = CURSEG_I(sbi, type); 2936 unsigned short seg_type = curseg->seg_type; 2937 2938 sanity_check_seg_type(sbi, seg_type); 2939 if (__is_large_section(sbi)) { 2940 if (f2fs_need_rand_seg(sbi)) { 2941 unsigned int hint = GET_SEC_FROM_SEG(sbi, curseg->segno); 2942 2943 if (GET_SEC_FROM_SEG(sbi, curseg->segno + 1) != hint) 2944 return curseg->segno; 2945 return get_random_u32_inclusive(curseg->segno + 1, 2946 GET_SEG_FROM_SEC(sbi, hint + 1) - 1); 2947 } 2948 return curseg->segno; 2949 } else if (f2fs_need_rand_seg(sbi)) { 2950 return get_random_u32_below(MAIN_SECS(sbi) * SEGS_PER_SEC(sbi)); 2951 } 2952 2953 /* inmem log may not locate on any segment after mount */ 2954 if (!curseg->inited) 2955 return 0; 2956 2957 if (unlikely(is_sbi_flag_set(sbi, SBI_CP_DISABLED))) 2958 return 0; 2959 2960 if (seg_type == CURSEG_HOT_DATA || IS_NODESEG(seg_type)) 2961 return 0; 2962 2963 if (SIT_I(sbi)->last_victim[ALLOC_NEXT]) 2964 return SIT_I(sbi)->last_victim[ALLOC_NEXT]; 2965 2966 /* find segments from 0 to reuse freed segments */ 2967 if (F2FS_OPTION(sbi).alloc_mode == ALLOC_MODE_REUSE) 2968 return 0; 2969 2970 return curseg->segno; 2971 } 2972 2973 static void reset_curseg_fields(struct curseg_info *curseg) 2974 { 2975 curseg->inited = false; 2976 curseg->segno = NULL_SEGNO; 2977 curseg->next_segno = 0; 2978 } 2979 2980 /* 2981 * Allocate a current working segment. 2982 * This function always allocates a free segment in LFS manner. 2983 */ 2984 static int new_curseg(struct f2fs_sb_info *sbi, int type, bool new_sec) 2985 { 2986 struct curseg_info *curseg = CURSEG_I(sbi, type); 2987 unsigned int segno = curseg->segno; 2988 bool pinning = type == CURSEG_COLD_DATA_PINNED; 2989 int ret; 2990 2991 if (curseg->inited) 2992 write_sum_page(sbi, curseg->sum_blk, segno); 2993 2994 segno = __get_next_segno(sbi, type); 2995 ret = get_new_segment(sbi, &segno, new_sec, pinning); 2996 if (ret) { 2997 if (ret == -ENOSPC) 2998 reset_curseg_fields(curseg); 2999 return ret; 3000 } 3001 3002 curseg->next_segno = segno; 3003 reset_curseg(sbi, type, 1); 3004 curseg->alloc_type = LFS; 3005 if (F2FS_OPTION(sbi).fs_mode == FS_MODE_FRAGMENT_BLK) 3006 curseg->fragment_remained_chunk = 3007 get_random_u32_inclusive(1, sbi->max_fragment_chunk); 3008 return 0; 3009 } 3010 3011 static int __next_free_blkoff(struct f2fs_sb_info *sbi, 3012 int segno, block_t start) 3013 { 3014 struct seg_entry *se = get_seg_entry(sbi, segno); 3015 int entries = SIT_VBLOCK_MAP_SIZE / sizeof(unsigned long); 3016 unsigned long *target_map = SIT_I(sbi)->tmp_map; 3017 unsigned long *ckpt_map = (unsigned long *)se->ckpt_valid_map; 3018 unsigned long *cur_map = (unsigned long *)se->cur_valid_map; 3019 int i; 3020 3021 for (i = 0; i < entries; i++) 3022 target_map[i] = ckpt_map[i] | cur_map[i]; 3023 3024 return __find_rev_next_zero_bit(target_map, BLKS_PER_SEG(sbi), start); 3025 } 3026 3027 static int f2fs_find_next_ssr_block(struct f2fs_sb_info *sbi, 3028 struct curseg_info *seg) 3029 { 3030 return __next_free_blkoff(sbi, seg->segno, seg->next_blkoff + 1); 3031 } 3032 3033 bool f2fs_segment_has_free_slot(struct f2fs_sb_info *sbi, int segno) 3034 { 3035 return __next_free_blkoff(sbi, segno, 0) < BLKS_PER_SEG(sbi); 3036 } 3037 3038 /* 3039 * This function always allocates a used segment(from dirty seglist) by SSR 3040 * manner, so it should recover the existing segment information of valid blocks 3041 */ 3042 static int change_curseg(struct f2fs_sb_info *sbi, int type) 3043 { 3044 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 3045 struct curseg_info *curseg = CURSEG_I(sbi, type); 3046 unsigned int new_segno = curseg->next_segno; 3047 struct f2fs_summary_block *sum_node; 3048 struct folio *sum_folio; 3049 3050 if (curseg->inited) 3051 write_sum_page(sbi, curseg->sum_blk, curseg->segno); 3052 3053 __set_test_and_inuse(sbi, new_segno); 3054 3055 mutex_lock(&dirty_i->seglist_lock); 3056 __remove_dirty_segment(sbi, new_segno, PRE); 3057 __remove_dirty_segment(sbi, new_segno, DIRTY); 3058 mutex_unlock(&dirty_i->seglist_lock); 3059 3060 reset_curseg(sbi, type, 1); 3061 curseg->alloc_type = SSR; 3062 curseg->next_blkoff = __next_free_blkoff(sbi, curseg->segno, 0); 3063 3064 sum_folio = f2fs_get_sum_folio(sbi, new_segno); 3065 if (IS_ERR(sum_folio)) { 3066 /* GC won't be able to use stale summary pages by cp_error */ 3067 memset(curseg->sum_blk, 0, sbi->sum_entry_size); 3068 return PTR_ERR(sum_folio); 3069 } 3070 sum_node = SUM_BLK_PAGE_ADDR(sbi, sum_folio, new_segno); 3071 memcpy(curseg->sum_blk, sum_node, sbi->sum_entry_size); 3072 f2fs_folio_put(sum_folio, true); 3073 return 0; 3074 } 3075 3076 static int get_ssr_segment(struct f2fs_sb_info *sbi, int type, 3077 int alloc_mode, unsigned long long age); 3078 3079 static int get_atssr_segment(struct f2fs_sb_info *sbi, int type, 3080 int target_type, int alloc_mode, 3081 unsigned long long age) 3082 { 3083 struct curseg_info *curseg = CURSEG_I(sbi, type); 3084 int ret = 0; 3085 3086 curseg->seg_type = target_type; 3087 3088 if (get_ssr_segment(sbi, type, alloc_mode, age)) { 3089 struct seg_entry *se = get_seg_entry(sbi, curseg->next_segno); 3090 3091 curseg->seg_type = se->type; 3092 ret = change_curseg(sbi, type); 3093 } else { 3094 /* allocate cold segment by default */ 3095 curseg->seg_type = CURSEG_COLD_DATA; 3096 ret = new_curseg(sbi, type, true); 3097 } 3098 stat_inc_seg_type(sbi, curseg); 3099 return ret; 3100 } 3101 3102 static int __f2fs_init_atgc_curseg(struct f2fs_sb_info *sbi, bool force) 3103 { 3104 struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_ALL_DATA_ATGC); 3105 int ret = 0; 3106 3107 if (!sbi->am.atgc_enabled && !force) 3108 return 0; 3109 3110 f2fs_down_read(&SM_I(sbi)->curseg_lock); 3111 3112 mutex_lock(&curseg->curseg_mutex); 3113 down_write(&SIT_I(sbi)->sentry_lock); 3114 3115 ret = get_atssr_segment(sbi, CURSEG_ALL_DATA_ATGC, 3116 CURSEG_COLD_DATA, SSR, 0); 3117 3118 up_write(&SIT_I(sbi)->sentry_lock); 3119 mutex_unlock(&curseg->curseg_mutex); 3120 3121 f2fs_up_read(&SM_I(sbi)->curseg_lock); 3122 return ret; 3123 } 3124 3125 int f2fs_init_inmem_curseg(struct f2fs_sb_info *sbi) 3126 { 3127 return __f2fs_init_atgc_curseg(sbi, false); 3128 } 3129 3130 int f2fs_reinit_atgc_curseg(struct f2fs_sb_info *sbi) 3131 { 3132 int ret; 3133 3134 if (!test_opt(sbi, ATGC)) 3135 return 0; 3136 if (sbi->am.atgc_enabled) 3137 return 0; 3138 if (le64_to_cpu(F2FS_CKPT(sbi)->elapsed_time) < 3139 sbi->am.age_threshold) 3140 return 0; 3141 3142 ret = __f2fs_init_atgc_curseg(sbi, true); 3143 if (!ret) { 3144 sbi->am.atgc_enabled = true; 3145 f2fs_info(sbi, "reenabled age threshold GC"); 3146 } 3147 return ret; 3148 } 3149 3150 static void __f2fs_save_inmem_curseg(struct f2fs_sb_info *sbi, int type) 3151 { 3152 struct curseg_info *curseg = CURSEG_I(sbi, type); 3153 3154 mutex_lock(&curseg->curseg_mutex); 3155 if (!curseg->inited) 3156 goto out; 3157 3158 if (get_valid_blocks(sbi, curseg->segno, false)) { 3159 write_sum_page(sbi, curseg->sum_blk, curseg->segno); 3160 } else { 3161 mutex_lock(&DIRTY_I(sbi)->seglist_lock); 3162 __set_test_and_free(sbi, curseg->segno, true); 3163 mutex_unlock(&DIRTY_I(sbi)->seglist_lock); 3164 } 3165 out: 3166 mutex_unlock(&curseg->curseg_mutex); 3167 } 3168 3169 void f2fs_save_inmem_curseg(struct f2fs_sb_info *sbi) 3170 { 3171 __f2fs_save_inmem_curseg(sbi, CURSEG_COLD_DATA_PINNED); 3172 3173 if (sbi->am.atgc_enabled) 3174 __f2fs_save_inmem_curseg(sbi, CURSEG_ALL_DATA_ATGC); 3175 } 3176 3177 static void __f2fs_restore_inmem_curseg(struct f2fs_sb_info *sbi, int type) 3178 { 3179 struct curseg_info *curseg = CURSEG_I(sbi, type); 3180 3181 mutex_lock(&curseg->curseg_mutex); 3182 if (!curseg->inited) 3183 goto out; 3184 if (get_valid_blocks(sbi, curseg->segno, false)) 3185 goto out; 3186 3187 mutex_lock(&DIRTY_I(sbi)->seglist_lock); 3188 __set_test_and_inuse(sbi, curseg->segno); 3189 mutex_unlock(&DIRTY_I(sbi)->seglist_lock); 3190 out: 3191 mutex_unlock(&curseg->curseg_mutex); 3192 } 3193 3194 void f2fs_restore_inmem_curseg(struct f2fs_sb_info *sbi) 3195 { 3196 __f2fs_restore_inmem_curseg(sbi, CURSEG_COLD_DATA_PINNED); 3197 3198 if (sbi->am.atgc_enabled) 3199 __f2fs_restore_inmem_curseg(sbi, CURSEG_ALL_DATA_ATGC); 3200 } 3201 3202 static int get_ssr_segment(struct f2fs_sb_info *sbi, int type, 3203 int alloc_mode, unsigned long long age) 3204 { 3205 struct curseg_info *curseg = CURSEG_I(sbi, type); 3206 unsigned segno = NULL_SEGNO; 3207 unsigned short seg_type = curseg->seg_type; 3208 int i, cnt; 3209 bool reversed = false; 3210 3211 sanity_check_seg_type(sbi, seg_type); 3212 3213 /* f2fs_need_SSR() already forces to do this */ 3214 if (!f2fs_get_victim(sbi, &segno, BG_GC, seg_type, 3215 alloc_mode, age, false)) { 3216 curseg->next_segno = segno; 3217 return 1; 3218 } 3219 3220 /* For node segments, let's do SSR more intensively */ 3221 if (IS_NODESEG(seg_type)) { 3222 if (seg_type >= CURSEG_WARM_NODE) { 3223 reversed = true; 3224 i = CURSEG_COLD_NODE; 3225 } else { 3226 i = CURSEG_HOT_NODE; 3227 } 3228 cnt = NR_CURSEG_NODE_TYPE; 3229 } else { 3230 if (seg_type >= CURSEG_WARM_DATA) { 3231 reversed = true; 3232 i = CURSEG_COLD_DATA; 3233 } else { 3234 i = CURSEG_HOT_DATA; 3235 } 3236 cnt = NR_CURSEG_DATA_TYPE; 3237 } 3238 3239 for (; cnt-- > 0; reversed ? i-- : i++) { 3240 if (i == seg_type) 3241 continue; 3242 if (!f2fs_get_victim(sbi, &segno, BG_GC, i, 3243 alloc_mode, age, false)) { 3244 curseg->next_segno = segno; 3245 return 1; 3246 } 3247 } 3248 3249 /* find valid_blocks=0 in dirty list */ 3250 if (unlikely(is_sbi_flag_set(sbi, SBI_CP_DISABLED))) { 3251 segno = get_free_segment(sbi); 3252 if (segno != NULL_SEGNO) { 3253 curseg->next_segno = segno; 3254 return 1; 3255 } 3256 } 3257 return 0; 3258 } 3259 3260 static bool need_new_seg(struct f2fs_sb_info *sbi, int type) 3261 { 3262 struct curseg_info *curseg = CURSEG_I(sbi, type); 3263 3264 if (!is_set_ckpt_flags(sbi, CP_CRC_RECOVERY_FLAG) && 3265 curseg->seg_type == CURSEG_WARM_NODE) 3266 return true; 3267 if (curseg->alloc_type == LFS && is_next_segment_free(sbi, curseg) && 3268 likely(!is_sbi_flag_set(sbi, SBI_CP_DISABLED))) 3269 return true; 3270 if (!f2fs_need_SSR(sbi) || !get_ssr_segment(sbi, type, SSR, 0)) 3271 return true; 3272 return false; 3273 } 3274 3275 int f2fs_allocate_segment_for_resize(struct f2fs_sb_info *sbi, int type, 3276 unsigned int start, unsigned int end) 3277 { 3278 struct curseg_info *curseg = CURSEG_I(sbi, type); 3279 unsigned int segno; 3280 int ret = 0; 3281 3282 f2fs_down_read(&SM_I(sbi)->curseg_lock); 3283 mutex_lock(&curseg->curseg_mutex); 3284 down_write(&SIT_I(sbi)->sentry_lock); 3285 3286 segno = CURSEG_I(sbi, type)->segno; 3287 if (segno < start || segno > end) 3288 goto unlock; 3289 3290 if (f2fs_need_SSR(sbi) && get_ssr_segment(sbi, type, SSR, 0)) 3291 ret = change_curseg(sbi, type); 3292 else 3293 ret = new_curseg(sbi, type, true); 3294 3295 stat_inc_seg_type(sbi, curseg); 3296 3297 locate_dirty_segment(sbi, segno); 3298 unlock: 3299 up_write(&SIT_I(sbi)->sentry_lock); 3300 3301 if (segno != curseg->segno) 3302 f2fs_notice(sbi, "For resize: curseg of type %d: %u ==> %u", 3303 type, segno, curseg->segno); 3304 3305 mutex_unlock(&curseg->curseg_mutex); 3306 f2fs_up_read(&SM_I(sbi)->curseg_lock); 3307 return ret; 3308 } 3309 3310 static int __allocate_new_segment(struct f2fs_sb_info *sbi, int type, 3311 bool new_sec, bool force) 3312 { 3313 struct curseg_info *curseg = CURSEG_I(sbi, type); 3314 unsigned int old_segno; 3315 int err = 0; 3316 3317 if (type == CURSEG_COLD_DATA_PINNED && !curseg->inited) 3318 goto allocate; 3319 3320 if (!force && curseg->inited && 3321 !curseg->next_blkoff && 3322 !get_valid_blocks(sbi, curseg->segno, new_sec) && 3323 !get_ckpt_valid_blocks(sbi, curseg->segno, new_sec)) 3324 return 0; 3325 3326 allocate: 3327 old_segno = curseg->segno; 3328 err = new_curseg(sbi, type, true); 3329 if (err) 3330 return err; 3331 stat_inc_seg_type(sbi, curseg); 3332 locate_dirty_segment(sbi, old_segno); 3333 return 0; 3334 } 3335 3336 int f2fs_allocate_new_section(struct f2fs_sb_info *sbi, int type, bool force) 3337 { 3338 int ret; 3339 3340 f2fs_down_read(&SM_I(sbi)->curseg_lock); 3341 down_write(&SIT_I(sbi)->sentry_lock); 3342 ret = __allocate_new_segment(sbi, type, true, force); 3343 up_write(&SIT_I(sbi)->sentry_lock); 3344 f2fs_up_read(&SM_I(sbi)->curseg_lock); 3345 3346 return ret; 3347 } 3348 3349 int f2fs_allocate_pinning_section(struct f2fs_sb_info *sbi) 3350 { 3351 struct f2fs_lock_context lc; 3352 int err; 3353 bool gc_required = true; 3354 3355 retry: 3356 f2fs_lock_op(sbi, &lc); 3357 err = f2fs_allocate_new_section(sbi, CURSEG_COLD_DATA_PINNED, false); 3358 f2fs_unlock_op(sbi, &lc); 3359 3360 if (f2fs_sb_has_blkzoned(sbi) && err == -EAGAIN && gc_required) { 3361 f2fs_down_write_trace(&sbi->gc_lock, &lc); 3362 err = f2fs_gc_range(sbi, 0, sbi->first_seq_zone_segno - 1, 3363 true, ZONED_PIN_SEC_REQUIRED_COUNT); 3364 f2fs_up_write_trace(&sbi->gc_lock, &lc); 3365 3366 gc_required = false; 3367 if (!err) 3368 goto retry; 3369 } 3370 3371 return err; 3372 } 3373 3374 int f2fs_allocate_new_segments(struct f2fs_sb_info *sbi) 3375 { 3376 int i; 3377 int err = 0; 3378 3379 f2fs_down_read(&SM_I(sbi)->curseg_lock); 3380 down_write(&SIT_I(sbi)->sentry_lock); 3381 for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) 3382 err += __allocate_new_segment(sbi, i, false, false); 3383 up_write(&SIT_I(sbi)->sentry_lock); 3384 f2fs_up_read(&SM_I(sbi)->curseg_lock); 3385 3386 return err; 3387 } 3388 3389 bool f2fs_exist_trim_candidates(struct f2fs_sb_info *sbi, 3390 struct cp_control *cpc) 3391 { 3392 __u64 trim_start = cpc->trim_start; 3393 bool has_candidate = false; 3394 3395 down_write(&SIT_I(sbi)->sentry_lock); 3396 for (; cpc->trim_start <= cpc->trim_end; cpc->trim_start++) { 3397 if (add_discard_addrs(sbi, cpc, true)) { 3398 has_candidate = true; 3399 break; 3400 } 3401 } 3402 up_write(&SIT_I(sbi)->sentry_lock); 3403 3404 cpc->trim_start = trim_start; 3405 return has_candidate; 3406 } 3407 3408 static unsigned int __issue_discard_cmd_range(struct f2fs_sb_info *sbi, 3409 struct discard_policy *dpolicy, 3410 unsigned int start, unsigned int end) 3411 { 3412 struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info; 3413 struct discard_cmd *prev_dc = NULL, *next_dc = NULL; 3414 struct rb_node **insert_p = NULL, *insert_parent = NULL; 3415 struct discard_cmd *dc; 3416 struct blk_plug plug; 3417 int issued; 3418 unsigned int trimmed = 0; 3419 3420 next: 3421 issued = 0; 3422 3423 mutex_lock(&dcc->cmd_lock); 3424 if (unlikely(dcc->rbtree_check)) 3425 f2fs_bug_on(sbi, !f2fs_check_discard_tree(sbi)); 3426 3427 dc = __lookup_discard_cmd_ret(&dcc->root, start, 3428 &prev_dc, &next_dc, &insert_p, &insert_parent); 3429 if (!dc) 3430 dc = next_dc; 3431 3432 blk_start_plug(&plug); 3433 3434 while (dc && dc->di.lstart <= end) { 3435 struct rb_node *node; 3436 int err = 0; 3437 3438 if (dc->di.len < dpolicy->granularity) 3439 goto skip; 3440 3441 if (dc->state != D_PREP) { 3442 list_move_tail(&dc->list, &dcc->fstrim_list); 3443 goto skip; 3444 } 3445 3446 err = __submit_discard_cmd(sbi, dpolicy, dc, &issued); 3447 3448 if (issued >= dpolicy->max_requests) { 3449 start = dc->di.lstart + dc->di.len; 3450 3451 if (err) 3452 __remove_discard_cmd(sbi, dc); 3453 3454 blk_finish_plug(&plug); 3455 mutex_unlock(&dcc->cmd_lock); 3456 trimmed += __wait_all_discard_cmd(sbi, NULL); 3457 f2fs_schedule_timeout(DEFAULT_DISCARD_INTERVAL); 3458 goto next; 3459 } 3460 skip: 3461 node = rb_next(&dc->rb_node); 3462 if (err) 3463 __remove_discard_cmd(sbi, dc); 3464 dc = rb_entry_safe(node, struct discard_cmd, rb_node); 3465 3466 if (fatal_signal_pending(current)) 3467 break; 3468 } 3469 3470 blk_finish_plug(&plug); 3471 mutex_unlock(&dcc->cmd_lock); 3472 3473 return trimmed; 3474 } 3475 3476 int f2fs_trim_fs(struct f2fs_sb_info *sbi, struct fstrim_range *range) 3477 { 3478 __u64 start = F2FS_BYTES_TO_BLK(range->start); 3479 __u64 end = start + F2FS_BYTES_TO_BLK(range->len) - 1; 3480 unsigned int start_segno, end_segno; 3481 block_t start_block, end_block; 3482 struct cp_control cpc; 3483 struct discard_policy dpolicy; 3484 struct f2fs_lock_context lc; 3485 unsigned long long trimmed = 0; 3486 int err = 0; 3487 bool need_align = f2fs_lfs_mode(sbi) && __is_large_section(sbi); 3488 3489 if (start >= MAX_BLKADDR(sbi) || range->len < sbi->blocksize) 3490 return -EINVAL; 3491 3492 if (end < MAIN_BLKADDR(sbi)) 3493 goto out; 3494 3495 if (is_sbi_flag_set(sbi, SBI_NEED_FSCK)) { 3496 f2fs_warn(sbi, "Found FS corruption, run fsck to fix."); 3497 return -EFSCORRUPTED; 3498 } 3499 3500 /* start/end segment number in main_area */ 3501 start_segno = (start <= MAIN_BLKADDR(sbi)) ? 0 : GET_SEGNO(sbi, start); 3502 end_segno = (end >= MAX_BLKADDR(sbi)) ? MAIN_SEGS(sbi) - 1 : 3503 GET_SEGNO(sbi, end); 3504 if (need_align) { 3505 start_segno = rounddown(start_segno, SEGS_PER_SEC(sbi)); 3506 end_segno = roundup(end_segno + 1, SEGS_PER_SEC(sbi)) - 1; 3507 } 3508 3509 cpc.reason = CP_DISCARD; 3510 cpc.trim_minlen = max_t(__u64, 1, F2FS_BYTES_TO_BLK(range->minlen)); 3511 cpc.trim_start = start_segno; 3512 cpc.trim_end = end_segno; 3513 3514 if (sbi->discard_blks == 0) 3515 goto out; 3516 3517 f2fs_down_write_trace(&sbi->gc_lock, &lc); 3518 stat_inc_cp_call_count(sbi, TOTAL_CALL); 3519 err = f2fs_write_checkpoint(sbi, &cpc); 3520 f2fs_up_write_trace(&sbi->gc_lock, &lc); 3521 if (err) 3522 goto out; 3523 3524 /* 3525 * We filed discard candidates, but actually we don't need to wait for 3526 * all of them, since they'll be issued in idle time along with runtime 3527 * discard option. User configuration looks like using runtime discard 3528 * or periodic fstrim instead of it. 3529 */ 3530 if (f2fs_realtime_discard_enable(sbi)) 3531 goto out; 3532 3533 start_block = START_BLOCK(sbi, start_segno); 3534 end_block = START_BLOCK(sbi, end_segno + 1); 3535 3536 __init_discard_policy(sbi, &dpolicy, DPOLICY_FSTRIM, cpc.trim_minlen); 3537 trimmed = __issue_discard_cmd_range(sbi, &dpolicy, 3538 start_block, end_block); 3539 3540 trimmed += __wait_discard_cmd_range(sbi, &dpolicy, 3541 start_block, end_block); 3542 out: 3543 if (!err) 3544 range->len = F2FS_BLK_TO_BYTES(trimmed); 3545 return err; 3546 } 3547 3548 int f2fs_rw_hint_to_seg_type(struct f2fs_sb_info *sbi, enum rw_hint hint) 3549 { 3550 if (F2FS_OPTION(sbi).active_logs == 2) 3551 return CURSEG_HOT_DATA; 3552 else if (F2FS_OPTION(sbi).active_logs == 4) 3553 return CURSEG_COLD_DATA; 3554 3555 /* active_log == 6 */ 3556 switch (hint) { 3557 case WRITE_LIFE_SHORT: 3558 return CURSEG_HOT_DATA; 3559 case WRITE_LIFE_EXTREME: 3560 return CURSEG_COLD_DATA; 3561 default: 3562 return CURSEG_WARM_DATA; 3563 } 3564 } 3565 3566 /* 3567 * This returns write hints for each segment type. This hints will be 3568 * passed down to block layer as below by default. 3569 * 3570 * User F2FS Block 3571 * ---- ---- ----- 3572 * META WRITE_LIFE_NONE|REQ_META 3573 * HOT_NODE WRITE_LIFE_NONE 3574 * WARM_NODE WRITE_LIFE_MEDIUM 3575 * COLD_NODE WRITE_LIFE_LONG 3576 * ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME 3577 * extension list " " 3578 * 3579 * -- buffered io 3580 * COLD_DATA WRITE_LIFE_EXTREME 3581 * HOT_DATA WRITE_LIFE_SHORT 3582 * WARM_DATA WRITE_LIFE_NOT_SET 3583 * 3584 * -- direct io 3585 * WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME 3586 * WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT 3587 * WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET 3588 * WRITE_LIFE_NONE " WRITE_LIFE_NONE 3589 * WRITE_LIFE_MEDIUM " WRITE_LIFE_MEDIUM 3590 * WRITE_LIFE_LONG " WRITE_LIFE_LONG 3591 */ 3592 enum rw_hint f2fs_io_type_to_rw_hint(struct f2fs_sb_info *sbi, 3593 enum page_type type, enum temp_type temp) 3594 { 3595 switch (type) { 3596 case DATA: 3597 switch (temp) { 3598 case WARM: 3599 return WRITE_LIFE_NOT_SET; 3600 case HOT: 3601 return WRITE_LIFE_SHORT; 3602 case COLD: 3603 return WRITE_LIFE_EXTREME; 3604 default: 3605 return WRITE_LIFE_NONE; 3606 } 3607 case NODE: 3608 switch (temp) { 3609 case WARM: 3610 return WRITE_LIFE_MEDIUM; 3611 case HOT: 3612 return WRITE_LIFE_NONE; 3613 case COLD: 3614 return WRITE_LIFE_LONG; 3615 default: 3616 return WRITE_LIFE_NONE; 3617 } 3618 case META: 3619 return WRITE_LIFE_NONE; 3620 default: 3621 return WRITE_LIFE_NONE; 3622 } 3623 } 3624 3625 u8 f2fs_io_type_to_write_stream(struct block_device *bdev, 3626 enum page_type type, enum temp_type temp) 3627 { 3628 unsigned short nr = bdev_max_write_streams(bdev); 3629 3630 if (type != DATA || !nr) 3631 return 0; 3632 if (nr < NR_TEMP_TYPE) 3633 return temp == COLD ? nr : HOT + 1; 3634 3635 return temp + 1; 3636 } 3637 3638 static int __get_segment_type_2(struct f2fs_io_info *fio) 3639 { 3640 if (fio->type == DATA) 3641 return CURSEG_HOT_DATA; 3642 else 3643 return CURSEG_HOT_NODE; 3644 } 3645 3646 static int __get_segment_type_4(struct f2fs_io_info *fio) 3647 { 3648 if (fio->type == DATA) { 3649 struct inode *inode = fio_inode(fio); 3650 3651 if (S_ISDIR(inode->i_mode)) 3652 return CURSEG_HOT_DATA; 3653 else 3654 return CURSEG_COLD_DATA; 3655 } else { 3656 if (IS_DNODE(fio->folio) && is_cold_node(fio->folio)) 3657 return CURSEG_WARM_NODE; 3658 else 3659 return CURSEG_COLD_NODE; 3660 } 3661 } 3662 3663 static int __get_age_segment_type(struct inode *inode, pgoff_t pgofs) 3664 { 3665 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 3666 struct extent_info ei = {}; 3667 3668 if (f2fs_lookup_age_extent_cache(inode, pgofs, &ei)) { 3669 if (!ei.age) 3670 return NO_CHECK_TYPE; 3671 if (ei.age <= sbi->hot_data_age_threshold) 3672 return CURSEG_HOT_DATA; 3673 if (ei.age <= sbi->warm_data_age_threshold) 3674 return CURSEG_WARM_DATA; 3675 return CURSEG_COLD_DATA; 3676 } 3677 return NO_CHECK_TYPE; 3678 } 3679 3680 static int __get_segment_type_6(struct f2fs_io_info *fio) 3681 { 3682 if (fio->type == DATA) { 3683 struct inode *inode = fio_inode(fio); 3684 int type; 3685 3686 if (is_inode_flag_set(inode, FI_ALIGNED_WRITE)) 3687 return CURSEG_COLD_DATA_PINNED; 3688 3689 if (page_private_gcing(fio->page)) { 3690 if (fio->sbi->am.atgc_enabled && 3691 (fio->io_type == FS_DATA_IO) && 3692 (fio->sbi->gc_mode != GC_URGENT_HIGH) && 3693 __is_valid_data_blkaddr(fio->old_blkaddr) && 3694 !is_inode_flag_set(inode, FI_OPU_WRITE)) 3695 return CURSEG_ALL_DATA_ATGC; 3696 else 3697 return CURSEG_COLD_DATA; 3698 } 3699 if (file_is_cold(inode) || f2fs_need_compress_data(inode)) 3700 return CURSEG_COLD_DATA; 3701 3702 type = __get_age_segment_type(inode, fio->folio->index); 3703 if (type != NO_CHECK_TYPE) 3704 return type; 3705 3706 if (file_is_hot(inode) || 3707 is_inode_flag_set(inode, FI_HOT_DATA) || 3708 f2fs_is_cow_file(inode) || 3709 is_inode_flag_set(inode, FI_NEED_IPU)) 3710 return CURSEG_HOT_DATA; 3711 return f2fs_rw_hint_to_seg_type(F2FS_I_SB(inode), 3712 inode->i_write_hint); 3713 } else { 3714 if (IS_DNODE(fio->folio)) 3715 return is_cold_node(fio->folio) ? CURSEG_WARM_NODE : 3716 CURSEG_HOT_NODE; 3717 return CURSEG_COLD_NODE; 3718 } 3719 } 3720 3721 enum temp_type f2fs_get_segment_temp(struct f2fs_sb_info *sbi, 3722 enum log_type type) 3723 { 3724 struct curseg_info *curseg = CURSEG_I(sbi, type); 3725 enum temp_type temp = COLD; 3726 3727 switch (curseg->seg_type) { 3728 case CURSEG_HOT_NODE: 3729 case CURSEG_HOT_DATA: 3730 temp = HOT; 3731 break; 3732 case CURSEG_WARM_NODE: 3733 case CURSEG_WARM_DATA: 3734 temp = WARM; 3735 break; 3736 case CURSEG_COLD_NODE: 3737 case CURSEG_COLD_DATA: 3738 temp = COLD; 3739 break; 3740 default: 3741 f2fs_bug_on(sbi, 1); 3742 } 3743 3744 return temp; 3745 } 3746 3747 static int __get_segment_type(struct f2fs_io_info *fio) 3748 { 3749 enum log_type type = CURSEG_HOT_DATA; 3750 3751 switch (F2FS_OPTION(fio->sbi).active_logs) { 3752 case 2: 3753 type = __get_segment_type_2(fio); 3754 break; 3755 case 4: 3756 type = __get_segment_type_4(fio); 3757 break; 3758 case 6: 3759 type = __get_segment_type_6(fio); 3760 break; 3761 default: 3762 f2fs_bug_on(fio->sbi, true); 3763 } 3764 3765 fio->temp = f2fs_get_segment_temp(fio->sbi, type); 3766 3767 return type; 3768 } 3769 3770 static void f2fs_randomize_chunk(struct f2fs_sb_info *sbi, 3771 struct curseg_info *seg) 3772 { 3773 /* To allocate block chunks in different sizes, use random number */ 3774 if (--seg->fragment_remained_chunk > 0) 3775 return; 3776 3777 seg->fragment_remained_chunk = 3778 get_random_u32_inclusive(1, sbi->max_fragment_chunk); 3779 seg->next_blkoff += 3780 get_random_u32_inclusive(1, sbi->max_fragment_hole); 3781 } 3782 3783 int f2fs_allocate_data_block(struct f2fs_sb_info *sbi, struct folio *folio, 3784 block_t old_blkaddr, block_t *new_blkaddr, 3785 struct f2fs_summary *sum, int type, 3786 struct f2fs_io_info *fio) 3787 { 3788 struct sit_info *sit_i = SIT_I(sbi); 3789 struct curseg_info *curseg = CURSEG_I(sbi, type); 3790 unsigned long long old_mtime; 3791 bool from_gc = (type == CURSEG_ALL_DATA_ATGC); 3792 struct seg_entry *se = NULL; 3793 bool segment_full = false; 3794 int ret = 0; 3795 3796 f2fs_down_read(&SM_I(sbi)->curseg_lock); 3797 3798 mutex_lock(&curseg->curseg_mutex); 3799 down_write(&sit_i->sentry_lock); 3800 3801 if (curseg->segno == NULL_SEGNO) { 3802 ret = -ENOSPC; 3803 goto out_err; 3804 } 3805 3806 if (from_gc) { 3807 f2fs_bug_on(sbi, GET_SEGNO(sbi, old_blkaddr) == NULL_SEGNO); 3808 se = get_seg_entry(sbi, GET_SEGNO(sbi, old_blkaddr)); 3809 sanity_check_seg_type(sbi, se->type); 3810 f2fs_bug_on(sbi, IS_NODESEG(se->type)); 3811 } 3812 *new_blkaddr = NEXT_FREE_BLKADDR(sbi, curseg); 3813 3814 f2fs_bug_on(sbi, curseg->next_blkoff >= BLKS_PER_SEG(sbi)); 3815 3816 f2fs_wait_discard_bio(sbi, *new_blkaddr); 3817 3818 sum_entries(curseg->sum_blk)[curseg->next_blkoff] = *sum; 3819 if (curseg->alloc_type == SSR) { 3820 curseg->next_blkoff = f2fs_find_next_ssr_block(sbi, curseg); 3821 } else { 3822 curseg->next_blkoff++; 3823 if (F2FS_OPTION(sbi).fs_mode == FS_MODE_FRAGMENT_BLK) 3824 f2fs_randomize_chunk(sbi, curseg); 3825 } 3826 if (curseg->next_blkoff >= f2fs_usable_blks_in_seg(sbi, curseg->segno)) 3827 segment_full = true; 3828 stat_inc_block_count(sbi, curseg); 3829 3830 if (from_gc) { 3831 old_mtime = get_segment_mtime(sbi, old_blkaddr); 3832 } else { 3833 update_segment_mtime(sbi, old_blkaddr, 0); 3834 old_mtime = 0; 3835 } 3836 update_segment_mtime(sbi, *new_blkaddr, old_mtime); 3837 3838 /* 3839 * SIT information should be updated before segment allocation, 3840 * since SSR needs latest valid block information. 3841 */ 3842 update_sit_entry(sbi, *new_blkaddr, 1); 3843 update_sit_entry(sbi, old_blkaddr, -1); 3844 3845 /* 3846 * If the current segment is full, flush it out and replace it with a 3847 * new segment. 3848 */ 3849 if (segment_full) { 3850 if (type == CURSEG_COLD_DATA_PINNED && 3851 !((curseg->segno + 1) % sbi->segs_per_sec)) { 3852 write_sum_page(sbi, curseg->sum_blk, curseg->segno); 3853 reset_curseg_fields(curseg); 3854 goto skip_new_segment; 3855 } 3856 3857 if (from_gc) { 3858 ret = get_atssr_segment(sbi, type, se->type, 3859 AT_SSR, se->mtime); 3860 } else { 3861 if (need_new_seg(sbi, type)) 3862 ret = new_curseg(sbi, type, false); 3863 else 3864 ret = change_curseg(sbi, type); 3865 stat_inc_seg_type(sbi, curseg); 3866 } 3867 3868 if (ret) 3869 goto out_err; 3870 } 3871 3872 skip_new_segment: 3873 /* 3874 * segment dirty status should be updated after segment allocation, 3875 * so we just need to update status only one time after previous 3876 * segment being closed. 3877 */ 3878 locate_dirty_segment(sbi, GET_SEGNO(sbi, old_blkaddr)); 3879 locate_dirty_segment(sbi, GET_SEGNO(sbi, *new_blkaddr)); 3880 3881 if (IS_DATASEG(curseg->seg_type)) { 3882 unsigned long long new_val; 3883 3884 new_val = atomic64_inc_return(&sbi->allocated_data_blocks); 3885 if (unlikely(new_val == ULLONG_MAX)) 3886 atomic64_set(&sbi->allocated_data_blocks, 0); 3887 } 3888 3889 up_write(&sit_i->sentry_lock); 3890 3891 if (folio && IS_NODESEG(curseg->seg_type)) { 3892 fill_node_footer_blkaddr(folio, NEXT_FREE_BLKADDR(sbi, curseg)); 3893 3894 f2fs_inode_chksum_set(sbi, folio); 3895 } 3896 3897 if (fio) { 3898 struct f2fs_bio_info *io; 3899 3900 INIT_LIST_HEAD(&fio->list); 3901 fio->in_list = 1; 3902 io = sbi->write_io[fio->type] + fio->temp; 3903 spin_lock(&io->io_lock); 3904 list_add_tail(&fio->list, &io->io_list); 3905 spin_unlock(&io->io_lock); 3906 } 3907 3908 mutex_unlock(&curseg->curseg_mutex); 3909 f2fs_up_read(&SM_I(sbi)->curseg_lock); 3910 return 0; 3911 3912 out_err: 3913 *new_blkaddr = NULL_ADDR; 3914 up_write(&sit_i->sentry_lock); 3915 mutex_unlock(&curseg->curseg_mutex); 3916 f2fs_up_read(&SM_I(sbi)->curseg_lock); 3917 return ret; 3918 } 3919 3920 void f2fs_update_device_state(struct f2fs_sb_info *sbi, nid_t ino, 3921 block_t blkaddr, unsigned int blkcnt) 3922 { 3923 if (!f2fs_is_multi_device(sbi)) 3924 return; 3925 3926 while (1) { 3927 unsigned int devidx = f2fs_target_device_index(sbi, blkaddr); 3928 unsigned int blks = FDEV(devidx).end_blk - blkaddr + 1; 3929 3930 /* update device state for fsync */ 3931 f2fs_set_dirty_device(sbi, ino, devidx, FLUSH_INO); 3932 3933 /* update device state for checkpoint */ 3934 if (!f2fs_test_bit(devidx, (char *)&sbi->dirty_device)) { 3935 spin_lock(&sbi->dev_lock); 3936 f2fs_set_bit(devidx, (char *)&sbi->dirty_device); 3937 spin_unlock(&sbi->dev_lock); 3938 } 3939 3940 if (blkcnt <= blks) 3941 break; 3942 blkcnt -= blks; 3943 blkaddr += blks; 3944 } 3945 } 3946 3947 static int log_type_to_seg_type(enum log_type type) 3948 { 3949 int seg_type = CURSEG_COLD_DATA; 3950 3951 switch (type) { 3952 case CURSEG_HOT_DATA: 3953 case CURSEG_WARM_DATA: 3954 case CURSEG_COLD_DATA: 3955 case CURSEG_HOT_NODE: 3956 case CURSEG_WARM_NODE: 3957 case CURSEG_COLD_NODE: 3958 seg_type = (int)type; 3959 break; 3960 case CURSEG_COLD_DATA_PINNED: 3961 case CURSEG_ALL_DATA_ATGC: 3962 seg_type = CURSEG_COLD_DATA; 3963 break; 3964 default: 3965 break; 3966 } 3967 return seg_type; 3968 } 3969 3970 static void do_write_page(struct f2fs_summary *sum, struct f2fs_io_info *fio) 3971 { 3972 struct folio *folio = fio->folio; 3973 enum log_type type = __get_segment_type(fio); 3974 int seg_type = log_type_to_seg_type(type); 3975 bool keep_order = (f2fs_lfs_mode(fio->sbi) && 3976 seg_type == CURSEG_COLD_DATA); 3977 int err; 3978 3979 if (keep_order) 3980 f2fs_down_read(&fio->sbi->io_order_lock); 3981 3982 err = f2fs_allocate_data_block(fio->sbi, folio, fio->old_blkaddr, 3983 &fio->new_blkaddr, sum, type, fio); 3984 if (unlikely(err)) { 3985 f2fs_err_ratelimited(fio->sbi, 3986 "%s Failed to allocate data block, ino:%u, index:%lu, type:%d, old_blkaddr:0x%x, new_blkaddr:0x%x, err:%d", 3987 __func__, fio->ino, folio->index, type, 3988 fio->old_blkaddr, fio->new_blkaddr, err); 3989 folio_end_writeback(folio); 3990 if (f2fs_in_warm_node_list(folio)) 3991 f2fs_del_fsync_node_entry(fio->sbi, folio); 3992 f2fs_bug_on(fio->sbi, !is_set_ckpt_flags(fio->sbi, 3993 CP_ERROR_FLAG)); 3994 goto out; 3995 } 3996 3997 f2fs_bug_on(fio->sbi, !f2fs_is_valid_blkaddr_raw(fio->sbi, 3998 fio->new_blkaddr, DATA_GENERIC_ENHANCE)); 3999 4000 if (GET_SEGNO(fio->sbi, fio->old_blkaddr) != NULL_SEGNO) 4001 f2fs_invalidate_internal_cache(fio->sbi, fio->old_blkaddr, 1); 4002 4003 /* writeout dirty page into bdev */ 4004 f2fs_submit_page_write(fio); 4005 4006 f2fs_update_device_state(fio->sbi, fio->ino, fio->new_blkaddr, 1); 4007 out: 4008 if (keep_order) 4009 f2fs_up_read(&fio->sbi->io_order_lock); 4010 } 4011 4012 void f2fs_do_write_meta_page(struct f2fs_sb_info *sbi, struct folio *folio, 4013 enum iostat_type io_type) 4014 { 4015 struct f2fs_io_info fio = { 4016 .sbi = sbi, 4017 .type = META, 4018 .temp = HOT, 4019 .op = REQ_OP_WRITE, 4020 .op_flags = REQ_SYNC | REQ_META | REQ_PRIO, 4021 .old_blkaddr = folio->index, 4022 .new_blkaddr = folio->index, 4023 .folio = folio, 4024 .encrypted_page = NULL, 4025 .in_list = 0, 4026 }; 4027 4028 if (unlikely(folio->index >= MAIN_BLKADDR(sbi))) 4029 fio.op_flags &= ~REQ_META; 4030 4031 folio_start_writeback(folio); 4032 f2fs_submit_page_write(&fio); 4033 4034 stat_inc_meta_count(sbi, folio->index); 4035 f2fs_update_iostat(sbi, NULL, io_type, F2FS_BLKSIZE); 4036 } 4037 4038 void f2fs_do_write_node_page(unsigned int nid, struct f2fs_io_info *fio) 4039 { 4040 struct f2fs_summary sum; 4041 4042 set_summary(&sum, nid, 0, 0); 4043 do_write_page(&sum, fio); 4044 4045 f2fs_update_iostat(fio->sbi, NULL, fio->io_type, F2FS_BLKSIZE); 4046 } 4047 4048 void f2fs_outplace_write_data(struct dnode_of_data *dn, 4049 struct f2fs_io_info *fio) 4050 { 4051 struct f2fs_sb_info *sbi = fio->sbi; 4052 struct f2fs_summary sum; 4053 4054 f2fs_bug_on(sbi, dn->data_blkaddr == NULL_ADDR); 4055 if (fio->io_type == FS_DATA_IO || fio->io_type == FS_CP_DATA_IO) 4056 f2fs_update_age_extent_cache(dn); 4057 set_summary(&sum, dn->nid, dn->ofs_in_node, fio->version); 4058 do_write_page(&sum, fio); 4059 f2fs_update_data_blkaddr(dn, fio->new_blkaddr); 4060 4061 f2fs_update_iostat(sbi, dn->inode, fio->io_type, F2FS_BLKSIZE); 4062 } 4063 4064 int f2fs_inplace_write_data(struct f2fs_io_info *fio) 4065 { 4066 int err; 4067 struct f2fs_sb_info *sbi = fio->sbi; 4068 unsigned int segno; 4069 4070 fio->new_blkaddr = fio->old_blkaddr; 4071 /* i/o temperature is needed for passing down write hints */ 4072 __get_segment_type(fio); 4073 4074 segno = GET_SEGNO(sbi, fio->new_blkaddr); 4075 4076 if (!IS_DATASEG(get_seg_entry(sbi, segno)->type)) { 4077 set_sbi_flag(sbi, SBI_NEED_FSCK); 4078 f2fs_warn(sbi, "%s: incorrect segment(%u) type, run fsck to fix.", 4079 __func__, segno); 4080 err = -EFSCORRUPTED; 4081 f2fs_handle_error(sbi, ERROR_INCONSISTENT_SUM_TYPE); 4082 goto drop_bio; 4083 } 4084 4085 if (f2fs_cp_error(sbi)) { 4086 err = -EIO; 4087 goto drop_bio; 4088 } 4089 4090 if (fio->meta_gc) 4091 f2fs_truncate_meta_inode_pages(sbi, fio->new_blkaddr, 1); 4092 4093 stat_inc_inplace_blocks(fio->sbi); 4094 4095 if (fio->bio && !IS_F2FS_IPU_NOCACHE(sbi)) 4096 err = f2fs_merge_page_bio(fio); 4097 else 4098 err = f2fs_submit_page_bio(fio); 4099 if (!err) { 4100 f2fs_update_device_state(fio->sbi, fio->ino, 4101 fio->new_blkaddr, 1); 4102 f2fs_update_iostat(fio->sbi, fio_inode(fio), 4103 fio->io_type, F2FS_BLKSIZE); 4104 } 4105 4106 return err; 4107 drop_bio: 4108 if (fio->bio && *(fio->bio)) { 4109 struct bio *bio = *(fio->bio); 4110 4111 bio->bi_status = BLK_STS_IOERR; 4112 bio_endio(bio); 4113 *(fio->bio) = NULL; 4114 } 4115 return err; 4116 } 4117 4118 static inline int __f2fs_get_curseg(struct f2fs_sb_info *sbi, 4119 unsigned int segno) 4120 { 4121 int i; 4122 4123 for (i = CURSEG_HOT_DATA; i < NO_CHECK_TYPE; i++) { 4124 if (CURSEG_I(sbi, i)->segno == segno) 4125 break; 4126 } 4127 return i; 4128 } 4129 4130 void f2fs_do_replace_block(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 4131 block_t old_blkaddr, block_t new_blkaddr, 4132 bool recover_curseg, bool recover_newaddr, 4133 bool from_gc) 4134 { 4135 struct sit_info *sit_i = SIT_I(sbi); 4136 struct curseg_info *curseg; 4137 unsigned int segno, old_cursegno; 4138 struct seg_entry *se; 4139 int type; 4140 unsigned short old_blkoff; 4141 unsigned char old_alloc_type; 4142 4143 segno = GET_SEGNO(sbi, new_blkaddr); 4144 se = get_seg_entry(sbi, segno); 4145 type = se->type; 4146 4147 f2fs_down_write(&SM_I(sbi)->curseg_lock); 4148 4149 if (!recover_curseg) { 4150 /* for recovery flow */ 4151 if (se->valid_blocks == 0 && !is_curseg(sbi, segno)) { 4152 if (old_blkaddr == NULL_ADDR) 4153 type = CURSEG_COLD_DATA; 4154 else 4155 type = CURSEG_WARM_DATA; 4156 } 4157 } else { 4158 if (is_curseg(sbi, segno)) { 4159 /* se->type is volatile as SSR allocation */ 4160 type = __f2fs_get_curseg(sbi, segno); 4161 f2fs_bug_on(sbi, type == NO_CHECK_TYPE); 4162 } else { 4163 type = CURSEG_WARM_DATA; 4164 } 4165 } 4166 4167 curseg = CURSEG_I(sbi, type); 4168 f2fs_bug_on(sbi, !IS_DATASEG(curseg->seg_type)); 4169 4170 mutex_lock(&curseg->curseg_mutex); 4171 down_write(&sit_i->sentry_lock); 4172 4173 old_cursegno = curseg->segno; 4174 old_blkoff = curseg->next_blkoff; 4175 old_alloc_type = curseg->alloc_type; 4176 4177 /* change the current segment */ 4178 if (segno != curseg->segno) { 4179 curseg->next_segno = segno; 4180 if (change_curseg(sbi, type)) 4181 goto out_unlock; 4182 } 4183 4184 curseg->next_blkoff = GET_BLKOFF_FROM_SEG0(sbi, new_blkaddr); 4185 sum_entries(curseg->sum_blk)[curseg->next_blkoff] = *sum; 4186 4187 if (!recover_curseg || recover_newaddr) { 4188 if (!from_gc) 4189 update_segment_mtime(sbi, new_blkaddr, 0); 4190 update_sit_entry(sbi, new_blkaddr, 1); 4191 } 4192 if (GET_SEGNO(sbi, old_blkaddr) != NULL_SEGNO) { 4193 f2fs_invalidate_internal_cache(sbi, old_blkaddr, 1); 4194 if (!from_gc) 4195 update_segment_mtime(sbi, old_blkaddr, 0); 4196 update_sit_entry(sbi, old_blkaddr, -1); 4197 } 4198 4199 locate_dirty_segment(sbi, GET_SEGNO(sbi, old_blkaddr)); 4200 locate_dirty_segment(sbi, GET_SEGNO(sbi, new_blkaddr)); 4201 4202 locate_dirty_segment(sbi, old_cursegno); 4203 4204 if (recover_curseg) { 4205 if (old_cursegno != curseg->segno) { 4206 curseg->next_segno = old_cursegno; 4207 if (change_curseg(sbi, type)) 4208 goto out_unlock; 4209 } 4210 curseg->next_blkoff = old_blkoff; 4211 curseg->alloc_type = old_alloc_type; 4212 } 4213 4214 out_unlock: 4215 up_write(&sit_i->sentry_lock); 4216 mutex_unlock(&curseg->curseg_mutex); 4217 f2fs_up_write(&SM_I(sbi)->curseg_lock); 4218 } 4219 4220 void f2fs_replace_block(struct f2fs_sb_info *sbi, struct dnode_of_data *dn, 4221 block_t old_addr, block_t new_addr, 4222 unsigned char version, bool recover_curseg, 4223 bool recover_newaddr) 4224 { 4225 struct f2fs_summary sum; 4226 4227 set_summary(&sum, dn->nid, dn->ofs_in_node, version); 4228 4229 f2fs_do_replace_block(sbi, &sum, old_addr, new_addr, 4230 recover_curseg, recover_newaddr, false); 4231 4232 f2fs_update_data_blkaddr(dn, new_addr); 4233 } 4234 4235 void f2fs_folio_wait_writeback(struct folio *folio, enum page_type type, 4236 bool ordered, bool locked) 4237 { 4238 if (folio_test_writeback(folio)) { 4239 struct f2fs_sb_info *sbi = F2FS_F_SB(folio); 4240 4241 /* submit cached LFS IO */ 4242 f2fs_submit_merged_write_folio(sbi, folio, type); 4243 /* submit cached IPU IO */ 4244 f2fs_submit_merged_ipu_write(sbi, NULL, folio); 4245 if (ordered) { 4246 folio_wait_writeback(folio); 4247 f2fs_bug_on(sbi, locked && folio_test_writeback(folio)); 4248 } else { 4249 folio_wait_stable(folio); 4250 } 4251 } 4252 } 4253 4254 void f2fs_wait_on_block_writeback(struct inode *inode, block_t blkaddr) 4255 { 4256 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 4257 struct folio *cfolio; 4258 4259 if (!f2fs_meta_inode_gc_required(inode)) 4260 return; 4261 4262 if (!__is_valid_data_blkaddr(blkaddr)) 4263 return; 4264 4265 cfolio = filemap_lock_folio(META_MAPPING(sbi), blkaddr); 4266 if (!IS_ERR(cfolio)) { 4267 f2fs_folio_wait_writeback(cfolio, DATA, true, true); 4268 f2fs_folio_put(cfolio, true); 4269 } 4270 } 4271 4272 void f2fs_wait_on_block_writeback_range(struct inode *inode, block_t blkaddr, 4273 block_t len) 4274 { 4275 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 4276 block_t i; 4277 4278 if (!f2fs_meta_inode_gc_required(inode)) 4279 return; 4280 4281 for (i = 0; i < len; i++) 4282 f2fs_wait_on_block_writeback(inode, blkaddr + i); 4283 4284 f2fs_truncate_meta_inode_pages(sbi, blkaddr, len); 4285 } 4286 4287 static int read_compacted_summaries(struct f2fs_sb_info *sbi) 4288 { 4289 struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi); 4290 struct curseg_info *seg_i; 4291 unsigned char *kaddr; 4292 struct folio *folio; 4293 block_t start; 4294 int i, j, offset; 4295 4296 start = start_sum_block(sbi); 4297 4298 folio = f2fs_get_meta_folio(sbi, start++); 4299 if (IS_ERR(folio)) 4300 return PTR_ERR(folio); 4301 kaddr = folio_address(folio); 4302 4303 /* Step 1: restore nat cache */ 4304 seg_i = CURSEG_I(sbi, CURSEG_HOT_DATA); 4305 memcpy(seg_i->journal, kaddr, sbi->sum_journal_size); 4306 4307 /* Step 2: restore sit cache */ 4308 seg_i = CURSEG_I(sbi, CURSEG_COLD_DATA); 4309 memcpy(seg_i->journal, kaddr + sbi->sum_journal_size, sbi->sum_journal_size); 4310 offset = 2 * sbi->sum_journal_size; 4311 4312 /* Step 3: restore summary entries */ 4313 for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) { 4314 unsigned short blk_off; 4315 unsigned int segno; 4316 4317 seg_i = CURSEG_I(sbi, i); 4318 segno = le32_to_cpu(ckpt->cur_data_segno[i]); 4319 blk_off = le16_to_cpu(ckpt->cur_data_blkoff[i]); 4320 seg_i->next_segno = segno; 4321 reset_curseg(sbi, i, 0); 4322 seg_i->alloc_type = ckpt->alloc_type[i]; 4323 seg_i->next_blkoff = blk_off; 4324 4325 if (seg_i->alloc_type == SSR) 4326 blk_off = BLKS_PER_SEG(sbi); 4327 4328 for (j = 0; j < blk_off; j++) { 4329 struct f2fs_summary *s; 4330 4331 s = (struct f2fs_summary *)(kaddr + offset); 4332 sum_entries(seg_i->sum_blk)[j] = *s; 4333 offset += SUMMARY_SIZE; 4334 if (offset + SUMMARY_SIZE <= sbi->blocksize - 4335 SUM_FOOTER_SIZE) 4336 continue; 4337 4338 f2fs_folio_put(folio, true); 4339 4340 folio = f2fs_get_meta_folio(sbi, start++); 4341 if (IS_ERR(folio)) 4342 return PTR_ERR(folio); 4343 kaddr = folio_address(folio); 4344 offset = 0; 4345 } 4346 } 4347 f2fs_folio_put(folio, true); 4348 return 0; 4349 } 4350 4351 static int read_normal_summaries(struct f2fs_sb_info *sbi, int type) 4352 { 4353 struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi); 4354 struct f2fs_summary_block *sum; 4355 struct curseg_info *curseg; 4356 struct folio *new; 4357 unsigned short blk_off; 4358 unsigned int segno = 0; 4359 block_t blk_addr = 0; 4360 int err = 0; 4361 4362 /* get segment number and block addr */ 4363 if (IS_DATASEG(type)) { 4364 segno = le32_to_cpu(ckpt->cur_data_segno[type]); 4365 blk_off = le16_to_cpu(ckpt->cur_data_blkoff[type - 4366 CURSEG_HOT_DATA]); 4367 if (__exist_node_summaries(sbi)) 4368 blk_addr = sum_blk_addr(sbi, NR_CURSEG_PERSIST_TYPE, type); 4369 else 4370 blk_addr = sum_blk_addr(sbi, NR_CURSEG_DATA_TYPE, type); 4371 } else { 4372 segno = le32_to_cpu(ckpt->cur_node_segno[type - 4373 CURSEG_HOT_NODE]); 4374 blk_off = le16_to_cpu(ckpt->cur_node_blkoff[type - 4375 CURSEG_HOT_NODE]); 4376 if (__exist_node_summaries(sbi)) 4377 blk_addr = sum_blk_addr(sbi, NR_CURSEG_NODE_TYPE, 4378 type - CURSEG_HOT_NODE); 4379 else 4380 blk_addr = GET_SUM_BLOCK(sbi, segno); 4381 } 4382 4383 new = f2fs_get_meta_folio(sbi, blk_addr); 4384 if (IS_ERR(new)) 4385 return PTR_ERR(new); 4386 sum = folio_address(new); 4387 4388 if (IS_NODESEG(type)) { 4389 if (__exist_node_summaries(sbi)) { 4390 struct f2fs_summary *ns = sum_entries(sum); 4391 int i; 4392 4393 for (i = 0; i < BLKS_PER_SEG(sbi); i++, ns++) { 4394 ns->version = 0; 4395 ns->ofs_in_node = 0; 4396 } 4397 } else { 4398 err = f2fs_restore_node_summary(sbi, segno, sum); 4399 if (err) 4400 goto out; 4401 } 4402 } 4403 4404 /* set uncompleted segment to curseg */ 4405 curseg = CURSEG_I(sbi, type); 4406 mutex_lock(&curseg->curseg_mutex); 4407 4408 /* update journal info */ 4409 down_write(&curseg->journal_rwsem); 4410 memcpy(curseg->journal, sum_journal(sbi, sum), sbi->sum_journal_size); 4411 up_write(&curseg->journal_rwsem); 4412 4413 memcpy(sum_entries(curseg->sum_blk), sum_entries(sum), 4414 sbi->sum_entry_size); 4415 memcpy(sum_footer(sbi, curseg->sum_blk), sum_footer(sbi, sum), 4416 SUM_FOOTER_SIZE); 4417 curseg->next_segno = segno; 4418 reset_curseg(sbi, type, 0); 4419 curseg->alloc_type = ckpt->alloc_type[type]; 4420 curseg->next_blkoff = blk_off; 4421 mutex_unlock(&curseg->curseg_mutex); 4422 out: 4423 f2fs_folio_put(new, true); 4424 return err; 4425 } 4426 4427 static int restore_curseg_summaries(struct f2fs_sb_info *sbi) 4428 { 4429 struct f2fs_journal *sit_j = CURSEG_I(sbi, CURSEG_COLD_DATA)->journal; 4430 struct f2fs_journal *nat_j = CURSEG_I(sbi, CURSEG_HOT_DATA)->journal; 4431 int type = CURSEG_HOT_DATA; 4432 int err; 4433 4434 if (is_set_ckpt_flags(sbi, CP_COMPACT_SUM_FLAG)) { 4435 int npages = f2fs_npages_for_summary_flush(sbi, true); 4436 4437 if (npages >= 2) 4438 f2fs_ra_meta_pages(sbi, start_sum_block(sbi), npages, 4439 META_CP, true); 4440 4441 /* restore for compacted data summary */ 4442 err = read_compacted_summaries(sbi); 4443 if (err) 4444 return err; 4445 type = CURSEG_HOT_NODE; 4446 } 4447 4448 if (__exist_node_summaries(sbi)) 4449 f2fs_ra_meta_pages(sbi, 4450 sum_blk_addr(sbi, NR_CURSEG_PERSIST_TYPE, type), 4451 NR_CURSEG_PERSIST_TYPE - type, META_CP, true); 4452 4453 for (; type <= CURSEG_COLD_NODE; type++) { 4454 err = read_normal_summaries(sbi, type); 4455 if (err) 4456 return err; 4457 } 4458 4459 /* sanity check for summary blocks */ 4460 if (nats_in_cursum(nat_j) > sbi->nat_journal_entries || 4461 sits_in_cursum(sit_j) > sbi->sit_journal_entries) { 4462 f2fs_err(sbi, "invalid journal entries nats %u sits %u", 4463 nats_in_cursum(nat_j), sits_in_cursum(sit_j)); 4464 return -EINVAL; 4465 } 4466 4467 return 0; 4468 } 4469 4470 static void write_compacted_summaries(struct f2fs_sb_info *sbi, block_t blkaddr) 4471 { 4472 struct folio *folio; 4473 unsigned char *kaddr; 4474 struct f2fs_summary *summary; 4475 struct curseg_info *seg_i; 4476 int written_size = 0; 4477 int i, j; 4478 4479 folio = f2fs_grab_meta_folio(sbi, blkaddr++); 4480 kaddr = folio_address(folio); 4481 memset(kaddr, 0, PAGE_SIZE); 4482 4483 /* Step 1: write nat cache */ 4484 seg_i = CURSEG_I(sbi, CURSEG_HOT_DATA); 4485 memcpy(kaddr, seg_i->journal, sbi->sum_journal_size); 4486 written_size += sbi->sum_journal_size; 4487 4488 /* Step 2: write sit cache */ 4489 seg_i = CURSEG_I(sbi, CURSEG_COLD_DATA); 4490 memcpy(kaddr + written_size, seg_i->journal, sbi->sum_journal_size); 4491 written_size += sbi->sum_journal_size; 4492 4493 /* Step 3: write summary entries */ 4494 for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) { 4495 seg_i = CURSEG_I(sbi, i); 4496 for (j = 0; j < f2fs_curseg_valid_blocks(sbi, i); j++) { 4497 if (!folio) { 4498 folio = f2fs_grab_meta_folio(sbi, blkaddr++); 4499 kaddr = folio_address(folio); 4500 memset(kaddr, 0, PAGE_SIZE); 4501 written_size = 0; 4502 } 4503 summary = (struct f2fs_summary *)(kaddr + written_size); 4504 *summary = sum_entries(seg_i->sum_blk)[j]; 4505 written_size += SUMMARY_SIZE; 4506 4507 if (written_size + SUMMARY_SIZE <= sbi->blocksize - 4508 SUM_FOOTER_SIZE) 4509 continue; 4510 4511 folio_mark_dirty(folio); 4512 f2fs_folio_put(folio, true); 4513 folio = NULL; 4514 } 4515 } 4516 if (folio) { 4517 folio_mark_dirty(folio); 4518 f2fs_folio_put(folio, true); 4519 } 4520 } 4521 4522 static void write_normal_summaries(struct f2fs_sb_info *sbi, 4523 block_t blkaddr, int type) 4524 { 4525 int i, end; 4526 4527 if (IS_DATASEG(type)) 4528 end = type + NR_CURSEG_DATA_TYPE; 4529 else 4530 end = type + NR_CURSEG_NODE_TYPE; 4531 4532 for (i = type; i < end; i++) 4533 write_current_sum_page(sbi, i, blkaddr + (i - type)); 4534 } 4535 4536 void f2fs_write_data_summaries(struct f2fs_sb_info *sbi, block_t start_blk) 4537 { 4538 if (is_set_ckpt_flags(sbi, CP_COMPACT_SUM_FLAG)) 4539 write_compacted_summaries(sbi, start_blk); 4540 else 4541 write_normal_summaries(sbi, start_blk, CURSEG_HOT_DATA); 4542 } 4543 4544 void f2fs_write_node_summaries(struct f2fs_sb_info *sbi, block_t start_blk) 4545 { 4546 write_normal_summaries(sbi, start_blk, CURSEG_HOT_NODE); 4547 } 4548 4549 int f2fs_lookup_journal_in_cursum(struct f2fs_sb_info *sbi, 4550 struct f2fs_journal *journal, int type, 4551 unsigned int val, int alloc) 4552 { 4553 int i; 4554 4555 if (type == NAT_JOURNAL) { 4556 for (i = 0; i < nats_in_cursum(journal); i++) { 4557 if (le32_to_cpu(nid_in_journal(journal, i)) == val) 4558 return i; 4559 } 4560 if (alloc && __has_cursum_space(sbi, journal, 1, NAT_JOURNAL)) 4561 return update_nats_in_cursum(journal, 1); 4562 } else if (type == SIT_JOURNAL) { 4563 for (i = 0; i < sits_in_cursum(journal); i++) 4564 if (le32_to_cpu(segno_in_journal(journal, i)) == val) 4565 return i; 4566 if (alloc && __has_cursum_space(sbi, journal, 1, SIT_JOURNAL)) 4567 return update_sits_in_cursum(journal, 1); 4568 } 4569 return -1; 4570 } 4571 4572 static struct folio *get_current_sit_folio(struct f2fs_sb_info *sbi, 4573 unsigned int segno) 4574 { 4575 return f2fs_get_meta_folio(sbi, current_sit_addr(sbi, segno)); 4576 } 4577 4578 static struct folio *get_next_sit_folio(struct f2fs_sb_info *sbi, 4579 unsigned int start) 4580 { 4581 struct sit_info *sit_i = SIT_I(sbi); 4582 struct folio *folio; 4583 pgoff_t src_off, dst_off; 4584 4585 src_off = current_sit_addr(sbi, start); 4586 dst_off = next_sit_addr(sbi, src_off); 4587 4588 folio = f2fs_grab_meta_folio(sbi, dst_off); 4589 seg_info_to_sit_folio(sbi, folio, start); 4590 4591 folio_mark_dirty(folio); 4592 set_to_next_sit(sit_i, start); 4593 4594 return folio; 4595 } 4596 4597 static struct sit_entry_set *grab_sit_entry_set(void) 4598 { 4599 struct sit_entry_set *ses = 4600 f2fs_kmem_cache_alloc(sit_entry_set_slab, 4601 GFP_NOFS, true, NULL); 4602 4603 ses->entry_cnt = 0; 4604 INIT_LIST_HEAD(&ses->set_list); 4605 return ses; 4606 } 4607 4608 static void release_sit_entry_set(struct sit_entry_set *ses) 4609 { 4610 list_del(&ses->set_list); 4611 kmem_cache_free(sit_entry_set_slab, ses); 4612 } 4613 4614 static void adjust_sit_entry_set(struct sit_entry_set *ses, 4615 struct list_head *head) 4616 { 4617 struct sit_entry_set *next = ses; 4618 4619 if (list_is_last(&ses->set_list, head)) 4620 return; 4621 4622 list_for_each_entry_continue(next, head, set_list) 4623 if (ses->entry_cnt <= next->entry_cnt) { 4624 list_move_tail(&ses->set_list, &next->set_list); 4625 return; 4626 } 4627 4628 list_move_tail(&ses->set_list, head); 4629 } 4630 4631 static void add_sit_entry(unsigned int segno, struct list_head *head) 4632 { 4633 struct sit_entry_set *ses; 4634 unsigned int start_segno = START_SEGNO(segno); 4635 4636 list_for_each_entry(ses, head, set_list) { 4637 if (ses->start_segno == start_segno) { 4638 ses->entry_cnt++; 4639 adjust_sit_entry_set(ses, head); 4640 return; 4641 } 4642 } 4643 4644 ses = grab_sit_entry_set(); 4645 4646 ses->start_segno = start_segno; 4647 ses->entry_cnt++; 4648 list_add(&ses->set_list, head); 4649 } 4650 4651 static void add_sits_in_set(struct f2fs_sb_info *sbi) 4652 { 4653 struct f2fs_sm_info *sm_info = SM_I(sbi); 4654 struct list_head *set_list = &sm_info->sit_entry_set; 4655 unsigned long *bitmap = SIT_I(sbi)->dirty_sentries_bitmap; 4656 unsigned int segno; 4657 4658 for_each_set_bit(segno, bitmap, MAIN_SEGS(sbi)) 4659 add_sit_entry(segno, set_list); 4660 } 4661 4662 static void remove_sits_in_journal(struct f2fs_sb_info *sbi) 4663 { 4664 struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA); 4665 struct f2fs_journal *journal = curseg->journal; 4666 int i; 4667 4668 down_write(&curseg->journal_rwsem); 4669 for (i = 0; i < sits_in_cursum(journal); i++) { 4670 unsigned int segno; 4671 bool dirtied; 4672 4673 segno = le32_to_cpu(segno_in_journal(journal, i)); 4674 dirtied = __mark_sit_entry_dirty(sbi, segno); 4675 4676 if (!dirtied) 4677 add_sit_entry(segno, &SM_I(sbi)->sit_entry_set); 4678 } 4679 update_sits_in_cursum(journal, -i); 4680 up_write(&curseg->journal_rwsem); 4681 } 4682 4683 /* 4684 * CP calls this function, which flushes SIT entries including sit_journal, 4685 * and moves prefree segs to free segs. 4686 */ 4687 void f2fs_flush_sit_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc) 4688 { 4689 struct sit_info *sit_i = SIT_I(sbi); 4690 unsigned long *bitmap = sit_i->dirty_sentries_bitmap; 4691 struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA); 4692 struct f2fs_journal *journal = curseg->journal; 4693 struct sit_entry_set *ses, *tmp; 4694 struct list_head *head = &SM_I(sbi)->sit_entry_set; 4695 bool to_journal = !is_sbi_flag_set(sbi, SBI_IS_RESIZEFS); 4696 struct seg_entry *se; 4697 4698 down_write(&sit_i->sentry_lock); 4699 4700 if (!sit_i->dirty_sentries) 4701 goto out; 4702 4703 /* 4704 * add and account sit entries of dirty bitmap in sit entry 4705 * set temporarily 4706 */ 4707 add_sits_in_set(sbi); 4708 4709 /* 4710 * if there are no enough space in journal to store dirty sit 4711 * entries, remove all entries from journal and add and account 4712 * them in sit entry set. 4713 */ 4714 if (!__has_cursum_space(sbi, journal, 4715 sit_i->dirty_sentries, SIT_JOURNAL) || !to_journal) 4716 remove_sits_in_journal(sbi); 4717 4718 /* 4719 * there are two steps to flush sit entries: 4720 * #1, flush sit entries to journal in current cold data summary block. 4721 * #2, flush sit entries to sit page. 4722 */ 4723 list_for_each_entry_safe(ses, tmp, head, set_list) { 4724 struct folio *folio = NULL; 4725 struct f2fs_sit_block *raw_sit = NULL; 4726 unsigned int start_segno = ses->start_segno; 4727 unsigned int end = min(start_segno + SIT_ENTRY_PER_BLOCK, 4728 (unsigned long)MAIN_SEGS(sbi)); 4729 unsigned int segno = start_segno; 4730 4731 if (to_journal && 4732 !__has_cursum_space(sbi, journal, ses->entry_cnt, 4733 SIT_JOURNAL)) 4734 to_journal = false; 4735 4736 if (to_journal) { 4737 down_write(&curseg->journal_rwsem); 4738 } else { 4739 folio = get_next_sit_folio(sbi, start_segno); 4740 raw_sit = folio_address(folio); 4741 } 4742 4743 /* flush dirty sit entries in region of current sit set */ 4744 for_each_set_bit_from(segno, bitmap, end) { 4745 int offset, sit_offset; 4746 4747 se = get_seg_entry(sbi, segno); 4748 4749 /* add discard candidates */ 4750 if (!(cpc->reason & CP_DISCARD)) { 4751 cpc->trim_start = segno; 4752 add_discard_addrs(sbi, cpc, false); 4753 } 4754 4755 if (to_journal) { 4756 offset = f2fs_lookup_journal_in_cursum(sbi, journal, 4757 SIT_JOURNAL, segno, 1); 4758 f2fs_bug_on(sbi, offset < 0); 4759 segno_in_journal(journal, offset) = 4760 cpu_to_le32(segno); 4761 seg_info_to_raw_sit(se, 4762 &sit_in_journal(journal, offset)); 4763 check_block_count(sbi, segno, 4764 &sit_in_journal(journal, offset)); 4765 } else { 4766 sit_offset = SIT_ENTRY_OFFSET(sit_i, segno); 4767 seg_info_to_raw_sit(se, 4768 &raw_sit->entries[sit_offset]); 4769 check_block_count(sbi, segno, 4770 &raw_sit->entries[sit_offset]); 4771 } 4772 4773 /* update ckpt_valid_block */ 4774 if (__is_large_section(sbi)) 4775 set_ckpt_valid_blocks(sbi, segno); 4776 4777 __clear_bit(segno, bitmap); 4778 sit_i->dirty_sentries--; 4779 ses->entry_cnt--; 4780 } 4781 4782 if (to_journal) 4783 up_write(&curseg->journal_rwsem); 4784 else 4785 f2fs_folio_put(folio, true); 4786 4787 f2fs_bug_on(sbi, ses->entry_cnt); 4788 release_sit_entry_set(ses); 4789 } 4790 4791 f2fs_bug_on(sbi, !list_empty(head)); 4792 f2fs_bug_on(sbi, sit_i->dirty_sentries); 4793 out: 4794 if (cpc->reason & CP_DISCARD) { 4795 __u64 trim_start = cpc->trim_start; 4796 4797 for (; cpc->trim_start <= cpc->trim_end; cpc->trim_start++) 4798 add_discard_addrs(sbi, cpc, false); 4799 4800 cpc->trim_start = trim_start; 4801 } 4802 up_write(&sit_i->sentry_lock); 4803 4804 set_prefree_as_free_segments(sbi); 4805 } 4806 4807 static int build_sit_info(struct f2fs_sb_info *sbi) 4808 { 4809 struct f2fs_super_block *raw_super = F2FS_RAW_SUPER(sbi); 4810 struct sit_info *sit_i; 4811 unsigned int sit_segs, start; 4812 char *src_bitmap, *bitmap; 4813 unsigned int bitmap_size, main_bitmap_size, sit_bitmap_size; 4814 unsigned int discard_map = f2fs_block_unit_discard(sbi) ? 1 : 0; 4815 4816 /* allocate memory for SIT information */ 4817 sit_i = f2fs_kzalloc(sbi, sizeof(struct sit_info), GFP_KERNEL); 4818 if (!sit_i) 4819 return -ENOMEM; 4820 4821 SM_I(sbi)->sit_info = sit_i; 4822 4823 sit_i->sentries = 4824 f2fs_kvzalloc(sbi, array_size(sizeof(struct seg_entry), 4825 MAIN_SEGS(sbi)), 4826 GFP_KERNEL); 4827 if (!sit_i->sentries) 4828 return -ENOMEM; 4829 4830 main_bitmap_size = f2fs_bitmap_size(MAIN_SEGS(sbi)); 4831 sit_i->dirty_sentries_bitmap = f2fs_kvzalloc(sbi, main_bitmap_size, 4832 GFP_KERNEL); 4833 if (!sit_i->dirty_sentries_bitmap) 4834 return -ENOMEM; 4835 4836 bitmap_size = MAIN_SEGS(sbi) * SIT_VBLOCK_MAP_SIZE * (2 + discard_map); 4837 sit_i->bitmap = f2fs_kvzalloc(sbi, bitmap_size, GFP_KERNEL); 4838 if (!sit_i->bitmap) 4839 return -ENOMEM; 4840 4841 bitmap = sit_i->bitmap; 4842 4843 for (start = 0; start < MAIN_SEGS(sbi); start++) { 4844 sit_i->sentries[start].cur_valid_map = bitmap; 4845 bitmap += SIT_VBLOCK_MAP_SIZE; 4846 4847 sit_i->sentries[start].ckpt_valid_map = bitmap; 4848 bitmap += SIT_VBLOCK_MAP_SIZE; 4849 4850 if (discard_map) { 4851 sit_i->sentries[start].discard_map = bitmap; 4852 bitmap += SIT_VBLOCK_MAP_SIZE; 4853 } 4854 } 4855 4856 sit_i->tmp_map = f2fs_kzalloc(sbi, SIT_VBLOCK_MAP_SIZE, GFP_KERNEL); 4857 if (!sit_i->tmp_map) 4858 return -ENOMEM; 4859 4860 if (__is_large_section(sbi)) { 4861 sit_i->sec_entries = 4862 f2fs_kvzalloc(sbi, array_size(sizeof(struct sec_entry), 4863 MAIN_SECS(sbi)), 4864 GFP_KERNEL); 4865 if (!sit_i->sec_entries) 4866 return -ENOMEM; 4867 } 4868 4869 /* get information related with SIT */ 4870 sit_segs = le32_to_cpu(raw_super->segment_count_sit) >> 1; 4871 4872 /* setup SIT bitmap from ckeckpoint pack */ 4873 sit_bitmap_size = __bitmap_size(sbi, SIT_BITMAP); 4874 src_bitmap = __bitmap_ptr(sbi, SIT_BITMAP); 4875 4876 sit_i->sit_bitmap = kmemdup(src_bitmap, sit_bitmap_size, GFP_KERNEL); 4877 if (!sit_i->sit_bitmap) 4878 return -ENOMEM; 4879 4880 #ifdef CONFIG_F2FS_CHECK_FS 4881 sit_i->invalid_segmap = f2fs_kvzalloc(sbi, 4882 main_bitmap_size, GFP_KERNEL); 4883 if (!sit_i->invalid_segmap) 4884 return -ENOMEM; 4885 #endif 4886 4887 sit_i->sit_base_addr = le32_to_cpu(raw_super->sit_blkaddr); 4888 sit_i->sit_blocks = SEGS_TO_BLKS(sbi, sit_segs); 4889 sit_i->written_valid_blocks = 0; 4890 sit_i->bitmap_size = sit_bitmap_size; 4891 sit_i->dirty_sentries = 0; 4892 sit_i->sents_per_block = SIT_ENTRY_PER_BLOCK; 4893 sit_i->elapsed_time = le64_to_cpu(sbi->ckpt->elapsed_time); 4894 sit_i->mounted_time = ktime_get_boottime_seconds(); 4895 init_rwsem(&sit_i->sentry_lock); 4896 return 0; 4897 } 4898 4899 static int build_free_segmap(struct f2fs_sb_info *sbi) 4900 { 4901 struct free_segmap_info *free_i; 4902 unsigned int bitmap_size, sec_bitmap_size; 4903 4904 /* allocate memory for free segmap information */ 4905 free_i = f2fs_kzalloc(sbi, sizeof(struct free_segmap_info), GFP_KERNEL); 4906 if (!free_i) 4907 return -ENOMEM; 4908 4909 SM_I(sbi)->free_info = free_i; 4910 4911 bitmap_size = f2fs_bitmap_size(MAIN_SEGS(sbi)); 4912 free_i->free_segmap = f2fs_kvmalloc(sbi, bitmap_size, GFP_KERNEL); 4913 if (!free_i->free_segmap) 4914 return -ENOMEM; 4915 4916 sec_bitmap_size = f2fs_bitmap_size(MAIN_SECS(sbi)); 4917 free_i->free_secmap = f2fs_kvmalloc(sbi, sec_bitmap_size, GFP_KERNEL); 4918 if (!free_i->free_secmap) 4919 return -ENOMEM; 4920 4921 /* set all segments as dirty temporarily */ 4922 memset(free_i->free_segmap, 0xff, bitmap_size); 4923 memset(free_i->free_secmap, 0xff, sec_bitmap_size); 4924 4925 /* init free segmap information */ 4926 free_i->start_segno = GET_SEGNO_FROM_SEG0(sbi, MAIN_BLKADDR(sbi)); 4927 free_i->free_segments = 0; 4928 free_i->free_sections = 0; 4929 spin_lock_init(&free_i->segmap_lock); 4930 return 0; 4931 } 4932 4933 static int build_curseg(struct f2fs_sb_info *sbi) 4934 { 4935 struct curseg_info *array; 4936 int i; 4937 4938 array = f2fs_kzalloc(sbi, array_size(NR_CURSEG_TYPE, 4939 sizeof(*array)), GFP_KERNEL); 4940 if (!array) 4941 return -ENOMEM; 4942 4943 SM_I(sbi)->curseg_array = array; 4944 4945 for (i = 0; i < NO_CHECK_TYPE; i++) { 4946 mutex_init(&array[i].curseg_mutex); 4947 array[i].sum_blk = f2fs_kzalloc(sbi, sbi->sum_blocksize, 4948 GFP_KERNEL); 4949 if (!array[i].sum_blk) 4950 return -ENOMEM; 4951 init_rwsem(&array[i].journal_rwsem); 4952 array[i].journal = f2fs_kzalloc(sbi, 4953 sbi->sum_journal_size, GFP_KERNEL); 4954 if (!array[i].journal) 4955 return -ENOMEM; 4956 array[i].seg_type = log_type_to_seg_type(i); 4957 reset_curseg_fields(&array[i]); 4958 } 4959 return restore_curseg_summaries(sbi); 4960 } 4961 4962 static int build_sit_entries(struct f2fs_sb_info *sbi) 4963 { 4964 struct sit_info *sit_i = SIT_I(sbi); 4965 struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA); 4966 struct f2fs_journal *journal = curseg->journal; 4967 struct seg_entry *se; 4968 struct f2fs_sit_entry sit; 4969 int sit_blk_cnt = SIT_BLK_CNT(sbi); 4970 unsigned int i, start, end; 4971 unsigned int readed, start_blk = 0; 4972 int err = 0; 4973 block_t sit_valid_blocks[2] = {0, 0}; 4974 4975 do { 4976 readed = f2fs_ra_meta_pages(sbi, start_blk, BIO_MAX_VECS, 4977 META_SIT, true); 4978 4979 start = start_blk * sit_i->sents_per_block; 4980 end = (start_blk + readed) * sit_i->sents_per_block; 4981 4982 for (; start < end && start < MAIN_SEGS(sbi); start++) { 4983 struct f2fs_sit_block *sit_blk; 4984 struct folio *folio; 4985 4986 se = &sit_i->sentries[start]; 4987 folio = get_current_sit_folio(sbi, start); 4988 if (IS_ERR(folio)) 4989 return PTR_ERR(folio); 4990 sit_blk = folio_address(folio); 4991 sit = sit_blk->entries[SIT_ENTRY_OFFSET(sit_i, start)]; 4992 f2fs_folio_put(folio, true); 4993 4994 err = check_block_count(sbi, start, &sit); 4995 if (err) 4996 return err; 4997 seg_info_from_raw_sit(se, &sit); 4998 4999 if (se->type >= NR_PERSISTENT_LOG) { 5000 f2fs_err(sbi, "Invalid segment type: %u, segno: %u", 5001 se->type, start); 5002 f2fs_handle_error(sbi, 5003 ERROR_INCONSISTENT_SUM_TYPE); 5004 return -EFSCORRUPTED; 5005 } 5006 5007 sit_valid_blocks[SE_PAGETYPE(se)] += se->valid_blocks; 5008 5009 if (!f2fs_block_unit_discard(sbi)) 5010 goto init_discard_map_done; 5011 5012 /* build discard map only one time */ 5013 if (is_set_ckpt_flags(sbi, CP_TRIMMED_FLAG)) { 5014 memset(se->discard_map, 0xff, 5015 SIT_VBLOCK_MAP_SIZE); 5016 goto init_discard_map_done; 5017 } 5018 memcpy(se->discard_map, se->cur_valid_map, 5019 SIT_VBLOCK_MAP_SIZE); 5020 sbi->discard_blks += BLKS_PER_SEG(sbi) - 5021 se->valid_blocks; 5022 init_discard_map_done: 5023 if (__is_large_section(sbi)) 5024 get_sec_entry(sbi, start)->valid_blocks += 5025 se->valid_blocks; 5026 } 5027 start_blk += readed; 5028 } while (start_blk < sit_blk_cnt); 5029 5030 down_read(&curseg->journal_rwsem); 5031 for (i = 0; i < sits_in_cursum(journal); i++) { 5032 unsigned int old_valid_blocks; 5033 5034 start = le32_to_cpu(segno_in_journal(journal, i)); 5035 if (start >= MAIN_SEGS(sbi)) { 5036 f2fs_err(sbi, "Wrong journal entry on segno %u", 5037 start); 5038 err = -EFSCORRUPTED; 5039 f2fs_handle_error(sbi, ERROR_CORRUPTED_JOURNAL); 5040 break; 5041 } 5042 5043 se = &sit_i->sentries[start]; 5044 sit = sit_in_journal(journal, i); 5045 5046 old_valid_blocks = se->valid_blocks; 5047 5048 sit_valid_blocks[SE_PAGETYPE(se)] -= old_valid_blocks; 5049 5050 err = check_block_count(sbi, start, &sit); 5051 if (err) 5052 break; 5053 seg_info_from_raw_sit(se, &sit); 5054 5055 if (se->type >= NR_PERSISTENT_LOG) { 5056 f2fs_err(sbi, "Invalid segment type: %u, segno: %u", 5057 se->type, start); 5058 err = -EFSCORRUPTED; 5059 f2fs_handle_error(sbi, ERROR_INCONSISTENT_SUM_TYPE); 5060 break; 5061 } 5062 5063 sit_valid_blocks[SE_PAGETYPE(se)] += se->valid_blocks; 5064 5065 if (f2fs_block_unit_discard(sbi)) { 5066 if (is_set_ckpt_flags(sbi, CP_TRIMMED_FLAG)) { 5067 memset(se->discard_map, 0xff, SIT_VBLOCK_MAP_SIZE); 5068 } else { 5069 memcpy(se->discard_map, se->cur_valid_map, 5070 SIT_VBLOCK_MAP_SIZE); 5071 sbi->discard_blks += old_valid_blocks; 5072 sbi->discard_blks -= se->valid_blocks; 5073 } 5074 } 5075 5076 if (__is_large_section(sbi)) { 5077 get_sec_entry(sbi, start)->valid_blocks += 5078 se->valid_blocks; 5079 get_sec_entry(sbi, start)->valid_blocks -= 5080 old_valid_blocks; 5081 } 5082 } 5083 up_read(&curseg->journal_rwsem); 5084 5085 /* update ckpt_valid_block */ 5086 if (__is_large_section(sbi)) { 5087 unsigned int segno; 5088 5089 for (segno = 0; segno < MAIN_SEGS(sbi); segno += SEGS_PER_SEC(sbi)) 5090 set_ckpt_valid_blocks(sbi, segno); 5091 } 5092 5093 if (err) 5094 return err; 5095 5096 if (sit_valid_blocks[NODE] != valid_node_count(sbi)) { 5097 f2fs_err(sbi, "SIT is corrupted node# %u vs %u", 5098 sit_valid_blocks[NODE], valid_node_count(sbi)); 5099 f2fs_handle_error(sbi, ERROR_INCONSISTENT_NODE_COUNT); 5100 return -EFSCORRUPTED; 5101 } 5102 5103 if (sit_valid_blocks[DATA] + sit_valid_blocks[NODE] > 5104 valid_user_blocks(sbi)) { 5105 f2fs_err(sbi, "SIT is corrupted data# %u %u vs %u", 5106 sit_valid_blocks[DATA], sit_valid_blocks[NODE], 5107 valid_user_blocks(sbi)); 5108 f2fs_handle_error(sbi, ERROR_INCONSISTENT_BLOCK_COUNT); 5109 return -EFSCORRUPTED; 5110 } 5111 5112 return 0; 5113 } 5114 5115 static void init_free_segmap(struct f2fs_sb_info *sbi) 5116 { 5117 unsigned int start; 5118 int type; 5119 struct seg_entry *sentry; 5120 5121 for (start = 0; start < MAIN_SEGS(sbi); start++) { 5122 if (f2fs_usable_blks_in_seg(sbi, start) == 0) 5123 continue; 5124 sentry = get_seg_entry(sbi, start); 5125 if (!sentry->valid_blocks) 5126 __set_free(sbi, start); 5127 else 5128 SIT_I(sbi)->written_valid_blocks += 5129 sentry->valid_blocks; 5130 } 5131 5132 /* set use the current segments */ 5133 for (type = CURSEG_HOT_DATA; type <= CURSEG_COLD_NODE; type++) { 5134 struct curseg_info *curseg_t = CURSEG_I(sbi, type); 5135 5136 __set_test_and_inuse(sbi, curseg_t->segno); 5137 } 5138 } 5139 5140 static void init_dirty_segmap(struct f2fs_sb_info *sbi) 5141 { 5142 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 5143 struct free_segmap_info *free_i = FREE_I(sbi); 5144 unsigned int segno = 0, offset = 0, secno; 5145 block_t valid_blocks, usable_blks_in_seg; 5146 5147 while (1) { 5148 /* find dirty segment based on free segmap */ 5149 segno = find_next_inuse(free_i, MAIN_SEGS(sbi), offset); 5150 if (segno >= MAIN_SEGS(sbi)) 5151 break; 5152 offset = segno + 1; 5153 valid_blocks = get_valid_blocks(sbi, segno, false); 5154 usable_blks_in_seg = f2fs_usable_blks_in_seg(sbi, segno); 5155 if (valid_blocks == usable_blks_in_seg || !valid_blocks) 5156 continue; 5157 if (valid_blocks > usable_blks_in_seg) { 5158 f2fs_bug_on(sbi, 1); 5159 continue; 5160 } 5161 mutex_lock(&dirty_i->seglist_lock); 5162 __locate_dirty_segment(sbi, segno, DIRTY); 5163 mutex_unlock(&dirty_i->seglist_lock); 5164 } 5165 5166 if (!__is_large_section(sbi)) 5167 return; 5168 5169 mutex_lock(&dirty_i->seglist_lock); 5170 for (segno = 0; segno < MAIN_SEGS(sbi); segno += SEGS_PER_SEC(sbi)) { 5171 valid_blocks = get_valid_blocks(sbi, segno, true); 5172 secno = GET_SEC_FROM_SEG(sbi, segno); 5173 5174 if (!valid_blocks || valid_blocks == CAP_BLKS_PER_SEC(sbi)) 5175 continue; 5176 if (is_cursec(sbi, secno)) 5177 continue; 5178 set_bit(secno, dirty_i->dirty_secmap); 5179 } 5180 mutex_unlock(&dirty_i->seglist_lock); 5181 } 5182 5183 static int init_victim_secmap(struct f2fs_sb_info *sbi) 5184 { 5185 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 5186 unsigned int bitmap_size = f2fs_bitmap_size(MAIN_SECS(sbi)); 5187 5188 dirty_i->victim_secmap = f2fs_kvzalloc(sbi, bitmap_size, GFP_KERNEL); 5189 if (!dirty_i->victim_secmap) 5190 return -ENOMEM; 5191 5192 dirty_i->pinned_secmap = f2fs_kvzalloc(sbi, bitmap_size, GFP_KERNEL); 5193 if (!dirty_i->pinned_secmap) 5194 return -ENOMEM; 5195 5196 dirty_i->pinned_secmap_cnt = 0; 5197 dirty_i->enable_pin_section = true; 5198 return 0; 5199 } 5200 5201 static int build_dirty_segmap(struct f2fs_sb_info *sbi) 5202 { 5203 struct dirty_seglist_info *dirty_i; 5204 unsigned int bitmap_size, i; 5205 5206 /* allocate memory for dirty segments list information */ 5207 dirty_i = f2fs_kzalloc(sbi, sizeof(struct dirty_seglist_info), 5208 GFP_KERNEL); 5209 if (!dirty_i) 5210 return -ENOMEM; 5211 5212 SM_I(sbi)->dirty_info = dirty_i; 5213 mutex_init(&dirty_i->seglist_lock); 5214 5215 bitmap_size = f2fs_bitmap_size(MAIN_SEGS(sbi)); 5216 5217 for (i = 0; i < NR_DIRTY_TYPE; i++) { 5218 dirty_i->dirty_segmap[i] = f2fs_kvzalloc(sbi, bitmap_size, 5219 GFP_KERNEL); 5220 if (!dirty_i->dirty_segmap[i]) 5221 return -ENOMEM; 5222 } 5223 5224 if (__is_large_section(sbi)) { 5225 bitmap_size = f2fs_bitmap_size(MAIN_SECS(sbi)); 5226 dirty_i->dirty_secmap = f2fs_kvzalloc(sbi, 5227 bitmap_size, GFP_KERNEL); 5228 if (!dirty_i->dirty_secmap) 5229 return -ENOMEM; 5230 } 5231 5232 init_dirty_segmap(sbi); 5233 return init_victim_secmap(sbi); 5234 } 5235 5236 static int sanity_check_curseg(struct f2fs_sb_info *sbi) 5237 { 5238 int i; 5239 5240 /* 5241 * In LFS/SSR curseg, .next_blkoff should point to an unused blkaddr; 5242 * In LFS curseg, all blkaddr after .next_blkoff should be unused. 5243 */ 5244 for (i = 0; i < NR_PERSISTENT_LOG; i++) { 5245 struct curseg_info *curseg = CURSEG_I(sbi, i); 5246 struct seg_entry *se = get_seg_entry(sbi, curseg->segno); 5247 unsigned int blkofs = curseg->next_blkoff; 5248 5249 if (f2fs_sb_has_readonly(sbi) && 5250 i != CURSEG_HOT_DATA && i != CURSEG_HOT_NODE) 5251 continue; 5252 5253 sanity_check_seg_type(sbi, curseg->seg_type); 5254 5255 if (curseg->alloc_type != LFS && curseg->alloc_type != SSR) { 5256 f2fs_err(sbi, 5257 "Current segment has invalid alloc_type:%d", 5258 curseg->alloc_type); 5259 f2fs_handle_error(sbi, ERROR_INVALID_CURSEG); 5260 return -EFSCORRUPTED; 5261 } 5262 5263 if (f2fs_test_bit(blkofs, se->cur_valid_map)) 5264 goto out; 5265 5266 if (curseg->alloc_type == SSR) 5267 continue; 5268 5269 for (blkofs += 1; blkofs < BLKS_PER_SEG(sbi); blkofs++) { 5270 if (!f2fs_test_bit(blkofs, se->cur_valid_map)) 5271 continue; 5272 out: 5273 f2fs_err(sbi, 5274 "Current segment's next free block offset is inconsistent with bitmap, logtype:%u, segno:%u, type:%u, next_blkoff:%u, blkofs:%u", 5275 i, curseg->segno, curseg->alloc_type, 5276 curseg->next_blkoff, blkofs); 5277 f2fs_handle_error(sbi, ERROR_INVALID_CURSEG); 5278 return -EFSCORRUPTED; 5279 } 5280 } 5281 return 0; 5282 } 5283 5284 #ifdef CONFIG_BLK_DEV_ZONED 5285 static int check_zone_write_pointer(struct f2fs_sb_info *sbi, 5286 struct f2fs_dev_info *fdev, 5287 struct blk_zone *zone) 5288 { 5289 unsigned int zone_segno; 5290 block_t zone_block, valid_block_cnt; 5291 unsigned int log_sectors_per_block = sbi->log_blocksize - SECTOR_SHIFT; 5292 int ret; 5293 unsigned int nofs_flags; 5294 5295 if (zone->type != BLK_ZONE_TYPE_SEQWRITE_REQ) 5296 return 0; 5297 5298 zone_block = fdev->start_blk + (zone->start >> log_sectors_per_block); 5299 zone_segno = GET_SEGNO(sbi, zone_block); 5300 5301 /* 5302 * Skip check of zones cursegs point to, since 5303 * fix_curseg_write_pointer() checks them. 5304 */ 5305 if (zone_segno >= MAIN_SEGS(sbi)) 5306 return 0; 5307 5308 /* 5309 * Get # of valid block of the zone. 5310 */ 5311 valid_block_cnt = get_valid_blocks(sbi, zone_segno, true); 5312 if (is_cursec(sbi, GET_SEC_FROM_SEG(sbi, zone_segno))) { 5313 f2fs_notice(sbi, "Open zones: valid block[0x%x,0x%x] cond[%s]", 5314 zone_segno, valid_block_cnt, 5315 blk_zone_cond_str(zone->cond)); 5316 return 0; 5317 } 5318 5319 if ((!valid_block_cnt && zone->cond == BLK_ZONE_COND_EMPTY) || 5320 (valid_block_cnt && zone->cond == BLK_ZONE_COND_FULL)) 5321 return 0; 5322 5323 if (!valid_block_cnt) { 5324 f2fs_notice(sbi, "Zone without valid block has non-zero write " 5325 "pointer. Reset the write pointer: cond[%s]", 5326 blk_zone_cond_str(zone->cond)); 5327 ret = __f2fs_issue_discard_zone(sbi, fdev->bdev, zone_block, 5328 zone->len >> log_sectors_per_block); 5329 if (ret) 5330 f2fs_err(sbi, "Discard zone failed: %s (errno=%d)", 5331 fdev->path, ret); 5332 return ret; 5333 } 5334 5335 /* 5336 * If there are valid blocks and the write pointer doesn't match 5337 * with them, we need to report the inconsistency and fill 5338 * the zone till the end to close the zone. This inconsistency 5339 * does not cause write error because the zone will not be 5340 * selected for write operation until it get discarded. 5341 */ 5342 f2fs_notice(sbi, "Valid blocks are not aligned with write " 5343 "pointer: valid block[0x%x,0x%x] cond[%s]", 5344 zone_segno, valid_block_cnt, blk_zone_cond_str(zone->cond)); 5345 5346 nofs_flags = memalloc_nofs_save(); 5347 ret = blkdev_zone_mgmt(fdev->bdev, REQ_OP_ZONE_FINISH, 5348 zone->start, zone->len); 5349 memalloc_nofs_restore(nofs_flags); 5350 if (ret == -EOPNOTSUPP) { 5351 ret = blkdev_issue_zeroout(fdev->bdev, zone->wp, 5352 zone->len - (zone->wp - zone->start), 5353 GFP_NOFS, 0); 5354 if (ret) 5355 f2fs_err(sbi, "Fill up zone failed: %s (errno=%d)", 5356 fdev->path, ret); 5357 } else if (ret) { 5358 f2fs_err(sbi, "Finishing zone failed: %s (errno=%d)", 5359 fdev->path, ret); 5360 } 5361 5362 return ret; 5363 } 5364 5365 static struct f2fs_dev_info *get_target_zoned_dev(struct f2fs_sb_info *sbi, 5366 block_t zone_blkaddr) 5367 { 5368 int i; 5369 5370 for (i = 0; i < sbi->s_ndevs; i++) { 5371 if (!bdev_is_zoned(FDEV(i).bdev)) 5372 continue; 5373 if (sbi->s_ndevs == 1 || (FDEV(i).start_blk <= zone_blkaddr && 5374 zone_blkaddr <= FDEV(i).end_blk)) 5375 return &FDEV(i); 5376 } 5377 5378 return NULL; 5379 } 5380 5381 static int report_one_zone_cb(struct blk_zone *zone, unsigned int idx, 5382 void *data) 5383 { 5384 memcpy(data, zone, sizeof(struct blk_zone)); 5385 return 0; 5386 } 5387 5388 static int do_fix_curseg_write_pointer(struct f2fs_sb_info *sbi, int type) 5389 { 5390 struct curseg_info *cs = CURSEG_I(sbi, type); 5391 struct f2fs_dev_info *zbd; 5392 struct blk_zone zone; 5393 unsigned int cs_section, wp_segno, wp_blkoff, wp_sector_off; 5394 block_t cs_zone_block, wp_block; 5395 unsigned int log_sectors_per_block = sbi->log_blocksize - SECTOR_SHIFT; 5396 sector_t zone_sector; 5397 int err; 5398 5399 cs_section = GET_SEC_FROM_SEG(sbi, cs->segno); 5400 cs_zone_block = START_BLOCK(sbi, GET_SEG_FROM_SEC(sbi, cs_section)); 5401 5402 zbd = get_target_zoned_dev(sbi, cs_zone_block); 5403 if (!zbd) 5404 return 0; 5405 5406 /* report zone for the sector the curseg points to */ 5407 zone_sector = (sector_t)(cs_zone_block - zbd->start_blk) 5408 << log_sectors_per_block; 5409 err = blkdev_report_zones(zbd->bdev, zone_sector, 1, 5410 report_one_zone_cb, &zone); 5411 if (err != 1) { 5412 f2fs_err(sbi, "Report zone failed: %s errno=(%d)", 5413 zbd->path, err); 5414 return err; 5415 } 5416 5417 if (zone.type != BLK_ZONE_TYPE_SEQWRITE_REQ) 5418 return 0; 5419 5420 /* 5421 * When safely unmounted in the previous mount, we could use current 5422 * segments. Otherwise, allocate new sections. 5423 */ 5424 if (is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG)) { 5425 wp_block = zbd->start_blk + (zone.wp >> log_sectors_per_block); 5426 wp_segno = GET_SEGNO(sbi, wp_block); 5427 wp_blkoff = wp_block - START_BLOCK(sbi, wp_segno); 5428 wp_sector_off = zone.wp & GENMASK(log_sectors_per_block - 1, 0); 5429 5430 if (cs->segno == wp_segno && cs->next_blkoff == wp_blkoff && 5431 wp_sector_off == 0) 5432 return 0; 5433 5434 f2fs_notice(sbi, "Unaligned curseg[%d] with write pointer: " 5435 "curseg[0x%x,0x%x] wp[0x%x,0x%x]", type, cs->segno, 5436 cs->next_blkoff, wp_segno, wp_blkoff); 5437 } 5438 5439 /* Allocate a new section if it's not new. */ 5440 if (cs->next_blkoff || 5441 cs->segno != GET_SEG_FROM_SEC(sbi, GET_ZONE_FROM_SEC(sbi, cs_section))) { 5442 unsigned int old_segno = cs->segno, old_blkoff = cs->next_blkoff; 5443 5444 f2fs_allocate_new_section(sbi, type, true); 5445 f2fs_notice(sbi, "Assign new section to curseg[%d]: " 5446 "[0x%x,0x%x] -> [0x%x,0x%x]", 5447 type, old_segno, old_blkoff, 5448 cs->segno, cs->next_blkoff); 5449 } 5450 5451 /* check consistency of the zone curseg pointed to */ 5452 if (check_zone_write_pointer(sbi, zbd, &zone)) 5453 return -EIO; 5454 5455 /* check newly assigned zone */ 5456 cs_section = GET_SEC_FROM_SEG(sbi, cs->segno); 5457 cs_zone_block = START_BLOCK(sbi, GET_SEG_FROM_SEC(sbi, cs_section)); 5458 5459 zbd = get_target_zoned_dev(sbi, cs_zone_block); 5460 if (!zbd) 5461 return 0; 5462 5463 zone_sector = (sector_t)(cs_zone_block - zbd->start_blk) 5464 << log_sectors_per_block; 5465 err = blkdev_report_zones(zbd->bdev, zone_sector, 1, 5466 report_one_zone_cb, &zone); 5467 if (err != 1) { 5468 f2fs_err(sbi, "Report zone failed: %s errno=(%d)", 5469 zbd->path, err); 5470 return err; 5471 } 5472 5473 if (zone.type != BLK_ZONE_TYPE_SEQWRITE_REQ) 5474 return 0; 5475 5476 if (zone.wp != zone.start) { 5477 f2fs_notice(sbi, 5478 "New zone for curseg[%d] is not yet discarded. " 5479 "Reset the zone: curseg[0x%x,0x%x]", 5480 type, cs->segno, cs->next_blkoff); 5481 err = __f2fs_issue_discard_zone(sbi, zbd->bdev, cs_zone_block, 5482 zone.len >> log_sectors_per_block); 5483 if (err) { 5484 f2fs_err(sbi, "Discard zone failed: %s (errno=%d)", 5485 zbd->path, err); 5486 return err; 5487 } 5488 } 5489 5490 return 0; 5491 } 5492 5493 static int fix_curseg_write_pointer(struct f2fs_sb_info *sbi) 5494 { 5495 int i, ret; 5496 5497 for (i = 0; i < NR_PERSISTENT_LOG; i++) { 5498 ret = do_fix_curseg_write_pointer(sbi, i); 5499 if (ret) 5500 return ret; 5501 } 5502 5503 return 0; 5504 } 5505 5506 struct check_zone_write_pointer_args { 5507 struct f2fs_sb_info *sbi; 5508 struct f2fs_dev_info *fdev; 5509 }; 5510 5511 static int check_zone_write_pointer_cb(struct blk_zone *zone, unsigned int idx, 5512 void *data) 5513 { 5514 struct check_zone_write_pointer_args *args; 5515 5516 args = (struct check_zone_write_pointer_args *)data; 5517 5518 return check_zone_write_pointer(args->sbi, args->fdev, zone); 5519 } 5520 5521 static int check_write_pointer(struct f2fs_sb_info *sbi) 5522 { 5523 int i, ret; 5524 struct check_zone_write_pointer_args args; 5525 5526 for (i = 0; i < sbi->s_ndevs; i++) { 5527 if (!bdev_is_zoned(FDEV(i).bdev)) 5528 continue; 5529 5530 args.sbi = sbi; 5531 args.fdev = &FDEV(i); 5532 ret = blkdev_report_zones(FDEV(i).bdev, 0, BLK_ALL_ZONES, 5533 check_zone_write_pointer_cb, &args); 5534 if (ret < 0) 5535 return ret; 5536 } 5537 5538 return 0; 5539 } 5540 5541 int f2fs_check_and_fix_write_pointer(struct f2fs_sb_info *sbi) 5542 { 5543 int ret; 5544 5545 if (!f2fs_sb_has_blkzoned(sbi) || f2fs_readonly(sbi->sb) || 5546 f2fs_hw_is_readonly(sbi)) 5547 return 0; 5548 5549 f2fs_notice(sbi, "Checking entire write pointers"); 5550 ret = fix_curseg_write_pointer(sbi); 5551 if (!ret) 5552 ret = check_write_pointer(sbi); 5553 return ret; 5554 } 5555 5556 /* 5557 * Return the number of usable blocks in a segment. The number of blocks 5558 * returned is always equal to the number of blocks in a segment for 5559 * segments fully contained within a sequential zone capacity or a 5560 * conventional zone. For segments partially contained in a sequential 5561 * zone capacity, the number of usable blocks up to the zone capacity 5562 * is returned. 0 is returned in all other cases. 5563 */ 5564 static inline unsigned int f2fs_usable_zone_blks_in_seg( 5565 struct f2fs_sb_info *sbi, unsigned int segno) 5566 { 5567 block_t seg_start, sec_start_blkaddr, sec_cap_blkaddr; 5568 unsigned int secno; 5569 5570 if (!sbi->unusable_blocks_per_sec) 5571 return BLKS_PER_SEG(sbi); 5572 5573 secno = GET_SEC_FROM_SEG(sbi, segno); 5574 seg_start = START_BLOCK(sbi, segno); 5575 sec_start_blkaddr = START_BLOCK(sbi, GET_SEG_FROM_SEC(sbi, secno)); 5576 sec_cap_blkaddr = sec_start_blkaddr + CAP_BLKS_PER_SEC(sbi); 5577 5578 /* 5579 * If segment starts before zone capacity and spans beyond 5580 * zone capacity, then usable blocks are from seg start to 5581 * zone capacity. If the segment starts after the zone capacity, 5582 * then there are no usable blocks. 5583 */ 5584 if (seg_start >= sec_cap_blkaddr) 5585 return 0; 5586 if (seg_start + BLKS_PER_SEG(sbi) > sec_cap_blkaddr) 5587 return sec_cap_blkaddr - seg_start; 5588 5589 return BLKS_PER_SEG(sbi); 5590 } 5591 #else 5592 int f2fs_check_and_fix_write_pointer(struct f2fs_sb_info *sbi) 5593 { 5594 return 0; 5595 } 5596 5597 static inline unsigned int f2fs_usable_zone_blks_in_seg(struct f2fs_sb_info *sbi, 5598 unsigned int segno) 5599 { 5600 return 0; 5601 } 5602 5603 #endif 5604 unsigned int f2fs_usable_blks_in_seg(struct f2fs_sb_info *sbi, 5605 unsigned int segno) 5606 { 5607 if (f2fs_sb_has_blkzoned(sbi)) 5608 return f2fs_usable_zone_blks_in_seg(sbi, segno); 5609 5610 return BLKS_PER_SEG(sbi); 5611 } 5612 5613 unsigned int f2fs_usable_segs_in_sec(struct f2fs_sb_info *sbi) 5614 { 5615 if (f2fs_sb_has_blkzoned(sbi)) 5616 return CAP_SEGS_PER_SEC(sbi); 5617 5618 return SEGS_PER_SEC(sbi); 5619 } 5620 5621 unsigned long long f2fs_get_section_mtime(struct f2fs_sb_info *sbi, 5622 unsigned int segno) 5623 { 5624 unsigned int usable_segs_per_sec = f2fs_usable_segs_in_sec(sbi); 5625 unsigned int secno = 0, start = 0; 5626 unsigned int total_valid_blocks = 0; 5627 unsigned long long mtime = 0; 5628 unsigned int i = 0; 5629 5630 secno = GET_SEC_FROM_SEG(sbi, segno); 5631 start = GET_SEG_FROM_SEC(sbi, secno); 5632 5633 if (!__is_large_section(sbi)) { 5634 mtime = get_seg_entry(sbi, start + i)->mtime; 5635 goto out; 5636 } 5637 5638 for (i = 0; i < usable_segs_per_sec; i++) { 5639 /* for large section, only check the mtime of valid segments */ 5640 struct seg_entry *se = get_seg_entry(sbi, start+i); 5641 5642 mtime += se->mtime * se->valid_blocks; 5643 total_valid_blocks += se->valid_blocks; 5644 } 5645 5646 if (total_valid_blocks == 0) 5647 return INVALID_MTIME; 5648 5649 mtime = div_u64(mtime, total_valid_blocks); 5650 out: 5651 if (unlikely(mtime == INVALID_MTIME)) 5652 mtime -= 1; 5653 return mtime; 5654 } 5655 5656 /* 5657 * Update min, max modified time for cost-benefit GC algorithm 5658 */ 5659 static void init_min_max_mtime(struct f2fs_sb_info *sbi) 5660 { 5661 struct sit_info *sit_i = SIT_I(sbi); 5662 unsigned int segno; 5663 5664 down_write(&sit_i->sentry_lock); 5665 5666 sit_i->min_mtime = ULLONG_MAX; 5667 5668 for (segno = 0; segno < MAIN_SEGS(sbi); segno += SEGS_PER_SEC(sbi)) { 5669 unsigned long long mtime = 0; 5670 5671 mtime = f2fs_get_section_mtime(sbi, segno); 5672 5673 if (sit_i->min_mtime > mtime) 5674 sit_i->min_mtime = mtime; 5675 } 5676 sit_i->max_mtime = get_mtime(sbi, false); 5677 sit_i->dirty_max_mtime = 0; 5678 up_write(&sit_i->sentry_lock); 5679 } 5680 5681 int f2fs_build_segment_manager(struct f2fs_sb_info *sbi) 5682 { 5683 struct f2fs_super_block *raw_super = F2FS_RAW_SUPER(sbi); 5684 struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi); 5685 struct f2fs_sm_info *sm_info; 5686 int err; 5687 5688 sm_info = f2fs_kzalloc(sbi, sizeof(struct f2fs_sm_info), GFP_KERNEL); 5689 if (!sm_info) 5690 return -ENOMEM; 5691 5692 /* init sm info */ 5693 sbi->sm_info = sm_info; 5694 sm_info->seg0_blkaddr = le32_to_cpu(raw_super->segment0_blkaddr); 5695 sm_info->main_blkaddr = le32_to_cpu(raw_super->main_blkaddr); 5696 sm_info->segment_count = le32_to_cpu(raw_super->segment_count); 5697 sm_info->reserved_segments = le32_to_cpu(ckpt->rsvd_segment_count); 5698 sm_info->ovp_segments = le32_to_cpu(ckpt->overprov_segment_count); 5699 sm_info->main_segments = le32_to_cpu(raw_super->segment_count_main); 5700 sm_info->ssa_blkaddr = le32_to_cpu(raw_super->ssa_blkaddr); 5701 sm_info->rec_prefree_segments = sm_info->main_segments * 5702 DEF_RECLAIM_PREFREE_SEGMENTS / 100; 5703 if (sm_info->rec_prefree_segments > DEF_MAX_RECLAIM_PREFREE_SEGMENTS) 5704 sm_info->rec_prefree_segments = DEF_MAX_RECLAIM_PREFREE_SEGMENTS; 5705 5706 if (!f2fs_lfs_mode(sbi)) 5707 sm_info->ipu_policy = BIT(F2FS_IPU_FSYNC); 5708 sm_info->min_ipu_util = DEF_MIN_IPU_UTIL; 5709 sm_info->min_fsync_blocks = DEF_MIN_FSYNC_BLOCKS; 5710 sm_info->min_seq_blocks = BLKS_PER_SEG(sbi); 5711 sm_info->min_hot_blocks = DEF_MIN_HOT_BLOCKS; 5712 sm_info->min_ssr_sections = reserved_sections(sbi); 5713 5714 INIT_LIST_HEAD(&sm_info->sit_entry_set); 5715 5716 init_f2fs_rwsem(&sm_info->curseg_lock); 5717 5718 err = f2fs_create_flush_cmd_control(sbi); 5719 if (err) 5720 return err; 5721 5722 err = create_discard_cmd_control(sbi); 5723 if (err) 5724 return err; 5725 5726 err = build_sit_info(sbi); 5727 if (err) 5728 return err; 5729 err = build_free_segmap(sbi); 5730 if (err) 5731 return err; 5732 err = build_curseg(sbi); 5733 if (err) 5734 return err; 5735 5736 /* reinit free segmap based on SIT */ 5737 err = build_sit_entries(sbi); 5738 if (err) 5739 return err; 5740 5741 init_free_segmap(sbi); 5742 err = build_dirty_segmap(sbi); 5743 if (err) 5744 return err; 5745 5746 err = sanity_check_curseg(sbi); 5747 if (err) 5748 return err; 5749 5750 init_min_max_mtime(sbi); 5751 return 0; 5752 } 5753 5754 static void discard_dirty_segmap(struct f2fs_sb_info *sbi, 5755 enum dirty_type dirty_type) 5756 { 5757 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 5758 5759 mutex_lock(&dirty_i->seglist_lock); 5760 kvfree(dirty_i->dirty_segmap[dirty_type]); 5761 dirty_i->nr_dirty[dirty_type] = 0; 5762 mutex_unlock(&dirty_i->seglist_lock); 5763 } 5764 5765 static void destroy_victim_secmap(struct f2fs_sb_info *sbi) 5766 { 5767 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 5768 5769 kvfree(dirty_i->pinned_secmap); 5770 kvfree(dirty_i->victim_secmap); 5771 } 5772 5773 static void destroy_dirty_segmap(struct f2fs_sb_info *sbi) 5774 { 5775 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 5776 int i; 5777 5778 if (!dirty_i) 5779 return; 5780 5781 /* discard pre-free/dirty segments list */ 5782 for (i = 0; i < NR_DIRTY_TYPE; i++) 5783 discard_dirty_segmap(sbi, i); 5784 5785 if (__is_large_section(sbi)) { 5786 mutex_lock(&dirty_i->seglist_lock); 5787 kvfree(dirty_i->dirty_secmap); 5788 mutex_unlock(&dirty_i->seglist_lock); 5789 } 5790 5791 destroy_victim_secmap(sbi); 5792 SM_I(sbi)->dirty_info = NULL; 5793 kfree(dirty_i); 5794 } 5795 5796 static void destroy_curseg(struct f2fs_sb_info *sbi) 5797 { 5798 struct curseg_info *array = SM_I(sbi)->curseg_array; 5799 int i; 5800 5801 if (!array) 5802 return; 5803 SM_I(sbi)->curseg_array = NULL; 5804 for (i = 0; i < NR_CURSEG_TYPE; i++) { 5805 kfree(array[i].sum_blk); 5806 kfree(array[i].journal); 5807 } 5808 kfree(array); 5809 } 5810 5811 static void destroy_free_segmap(struct f2fs_sb_info *sbi) 5812 { 5813 struct free_segmap_info *free_i = SM_I(sbi)->free_info; 5814 5815 if (!free_i) 5816 return; 5817 SM_I(sbi)->free_info = NULL; 5818 kvfree(free_i->free_segmap); 5819 kvfree(free_i->free_secmap); 5820 kfree(free_i); 5821 } 5822 5823 static void destroy_sit_info(struct f2fs_sb_info *sbi) 5824 { 5825 struct sit_info *sit_i = SIT_I(sbi); 5826 5827 if (!sit_i) 5828 return; 5829 5830 if (sit_i->sentries) 5831 kvfree(sit_i->bitmap); 5832 kfree(sit_i->tmp_map); 5833 5834 kvfree(sit_i->sentries); 5835 kvfree(sit_i->sec_entries); 5836 kvfree(sit_i->dirty_sentries_bitmap); 5837 5838 SM_I(sbi)->sit_info = NULL; 5839 kfree(sit_i->sit_bitmap); 5840 #ifdef CONFIG_F2FS_CHECK_FS 5841 kvfree(sit_i->invalid_segmap); 5842 #endif 5843 kfree(sit_i); 5844 } 5845 5846 void f2fs_destroy_segment_manager(struct f2fs_sb_info *sbi) 5847 { 5848 struct f2fs_sm_info *sm_info = SM_I(sbi); 5849 5850 if (!sm_info) 5851 return; 5852 f2fs_destroy_flush_cmd_control(sbi, true); 5853 destroy_discard_cmd_control(sbi); 5854 destroy_dirty_segmap(sbi); 5855 destroy_curseg(sbi); 5856 destroy_free_segmap(sbi); 5857 destroy_sit_info(sbi); 5858 sbi->sm_info = NULL; 5859 kfree(sm_info); 5860 } 5861 5862 int __init f2fs_create_segment_manager_caches(void) 5863 { 5864 discard_entry_slab = f2fs_kmem_cache_create("f2fs_discard_entry", 5865 sizeof(struct discard_entry)); 5866 if (!discard_entry_slab) 5867 goto fail; 5868 5869 discard_cmd_slab = f2fs_kmem_cache_create("f2fs_discard_cmd", 5870 sizeof(struct discard_cmd)); 5871 if (!discard_cmd_slab) 5872 goto destroy_discard_entry; 5873 5874 sit_entry_set_slab = f2fs_kmem_cache_create("f2fs_sit_entry_set", 5875 sizeof(struct sit_entry_set)); 5876 if (!sit_entry_set_slab) 5877 goto destroy_discard_cmd; 5878 5879 revoke_entry_slab = f2fs_kmem_cache_create("f2fs_revoke_entry", 5880 sizeof(struct revoke_entry)); 5881 if (!revoke_entry_slab) 5882 goto destroy_sit_entry_set; 5883 return 0; 5884 5885 destroy_sit_entry_set: 5886 kmem_cache_destroy(sit_entry_set_slab); 5887 destroy_discard_cmd: 5888 kmem_cache_destroy(discard_cmd_slab); 5889 destroy_discard_entry: 5890 kmem_cache_destroy(discard_entry_slab); 5891 fail: 5892 return -ENOMEM; 5893 } 5894 5895 void f2fs_destroy_segment_manager_caches(void) 5896 { 5897 kmem_cache_destroy(sit_entry_set_slab); 5898 kmem_cache_destroy(discard_cmd_slab); 5899 kmem_cache_destroy(discard_entry_slab); 5900 kmem_cache_destroy(revoke_entry_slab); 5901 } 5902