1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * fs/f2fs/gc.c 4 * 5 * Copyright (c) 2012 Samsung Electronics Co., Ltd. 6 * http://www.samsung.com/ 7 */ 8 #include <linux/fs.h> 9 #include <linux/module.h> 10 #include <linux/backing-dev.h> 11 #include <linux/init.h> 12 #include <linux/f2fs_fs.h> 13 #include <linux/kthread.h> 14 #include <linux/delay.h> 15 #include <linux/freezer.h> 16 17 #include "f2fs.h" 18 #include "node.h" 19 #include "segment.h" 20 #include "gc.h" 21 #include <trace/events/f2fs.h> 22 23 static int gc_thread_func(void *data) 24 { 25 struct f2fs_sb_info *sbi = data; 26 struct f2fs_gc_kthread *gc_th = sbi->gc_thread; 27 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head; 28 unsigned int wait_ms; 29 30 wait_ms = gc_th->min_sleep_time; 31 32 set_freezable(); 33 do { 34 wait_event_interruptible_timeout(*wq, 35 kthread_should_stop() || freezing(current) || 36 gc_th->gc_wake, 37 msecs_to_jiffies(wait_ms)); 38 39 /* give it a try one time */ 40 if (gc_th->gc_wake) 41 gc_th->gc_wake = 0; 42 43 if (try_to_freeze()) { 44 stat_other_skip_bggc_count(sbi); 45 continue; 46 } 47 if (kthread_should_stop()) 48 break; 49 50 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) { 51 increase_sleep_time(gc_th, &wait_ms); 52 stat_other_skip_bggc_count(sbi); 53 continue; 54 } 55 56 if (time_to_inject(sbi, FAULT_CHECKPOINT)) { 57 f2fs_show_injection_info(FAULT_CHECKPOINT); 58 f2fs_stop_checkpoint(sbi, false); 59 } 60 61 if (!sb_start_write_trylock(sbi->sb)) { 62 stat_other_skip_bggc_count(sbi); 63 continue; 64 } 65 66 /* 67 * [GC triggering condition] 68 * 0. GC is not conducted currently. 69 * 1. There are enough dirty segments. 70 * 2. IO subsystem is idle by checking the # of writeback pages. 71 * 3. IO subsystem is idle by checking the # of requests in 72 * bdev's request list. 73 * 74 * Note) We have to avoid triggering GCs frequently. 75 * Because it is possible that some segments can be 76 * invalidated soon after by user update or deletion. 77 * So, I'd like to wait some time to collect dirty segments. 78 */ 79 if (sbi->gc_mode == GC_URGENT) { 80 wait_ms = gc_th->urgent_sleep_time; 81 mutex_lock(&sbi->gc_mutex); 82 goto do_gc; 83 } 84 85 if (!mutex_trylock(&sbi->gc_mutex)) { 86 stat_other_skip_bggc_count(sbi); 87 goto next; 88 } 89 90 if (!is_idle(sbi, GC_TIME)) { 91 increase_sleep_time(gc_th, &wait_ms); 92 mutex_unlock(&sbi->gc_mutex); 93 stat_io_skip_bggc_count(sbi); 94 goto next; 95 } 96 97 if (has_enough_invalid_blocks(sbi)) 98 decrease_sleep_time(gc_th, &wait_ms); 99 else 100 increase_sleep_time(gc_th, &wait_ms); 101 do_gc: 102 stat_inc_bggc_count(sbi); 103 104 /* if return value is not zero, no victim was selected */ 105 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true, NULL_SEGNO)) 106 wait_ms = gc_th->no_gc_sleep_time; 107 108 trace_f2fs_background_gc(sbi->sb, wait_ms, 109 prefree_segments(sbi), free_segments(sbi)); 110 111 /* balancing f2fs's metadata periodically */ 112 f2fs_balance_fs_bg(sbi); 113 next: 114 sb_end_write(sbi->sb); 115 116 } while (!kthread_should_stop()); 117 return 0; 118 } 119 120 int f2fs_start_gc_thread(struct f2fs_sb_info *sbi) 121 { 122 struct f2fs_gc_kthread *gc_th; 123 dev_t dev = sbi->sb->s_bdev->bd_dev; 124 int err = 0; 125 126 gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL); 127 if (!gc_th) { 128 err = -ENOMEM; 129 goto out; 130 } 131 132 gc_th->urgent_sleep_time = DEF_GC_THREAD_URGENT_SLEEP_TIME; 133 gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME; 134 gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME; 135 gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME; 136 137 gc_th->gc_wake= 0; 138 139 sbi->gc_thread = gc_th; 140 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head); 141 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi, 142 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev)); 143 if (IS_ERR(gc_th->f2fs_gc_task)) { 144 err = PTR_ERR(gc_th->f2fs_gc_task); 145 kfree(gc_th); 146 sbi->gc_thread = NULL; 147 } 148 out: 149 return err; 150 } 151 152 void f2fs_stop_gc_thread(struct f2fs_sb_info *sbi) 153 { 154 struct f2fs_gc_kthread *gc_th = sbi->gc_thread; 155 if (!gc_th) 156 return; 157 kthread_stop(gc_th->f2fs_gc_task); 158 kfree(gc_th); 159 sbi->gc_thread = NULL; 160 } 161 162 static int select_gc_type(struct f2fs_sb_info *sbi, int gc_type) 163 { 164 int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY; 165 166 switch (sbi->gc_mode) { 167 case GC_IDLE_CB: 168 gc_mode = GC_CB; 169 break; 170 case GC_IDLE_GREEDY: 171 case GC_URGENT: 172 gc_mode = GC_GREEDY; 173 break; 174 } 175 return gc_mode; 176 } 177 178 static void select_policy(struct f2fs_sb_info *sbi, int gc_type, 179 int type, struct victim_sel_policy *p) 180 { 181 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 182 183 if (p->alloc_mode == SSR) { 184 p->gc_mode = GC_GREEDY; 185 p->dirty_segmap = dirty_i->dirty_segmap[type]; 186 p->max_search = dirty_i->nr_dirty[type]; 187 p->ofs_unit = 1; 188 } else { 189 p->gc_mode = select_gc_type(sbi, gc_type); 190 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY]; 191 p->max_search = dirty_i->nr_dirty[DIRTY]; 192 p->ofs_unit = sbi->segs_per_sec; 193 } 194 195 /* we need to check every dirty segments in the FG_GC case */ 196 if (gc_type != FG_GC && 197 (sbi->gc_mode != GC_URGENT) && 198 p->max_search > sbi->max_victim_search) 199 p->max_search = sbi->max_victim_search; 200 201 /* let's select beginning hot/small space first in no_heap mode*/ 202 if (test_opt(sbi, NOHEAP) && 203 (type == CURSEG_HOT_DATA || IS_NODESEG(type))) 204 p->offset = 0; 205 else 206 p->offset = SIT_I(sbi)->last_victim[p->gc_mode]; 207 } 208 209 static unsigned int get_max_cost(struct f2fs_sb_info *sbi, 210 struct victim_sel_policy *p) 211 { 212 /* SSR allocates in a segment unit */ 213 if (p->alloc_mode == SSR) 214 return sbi->blocks_per_seg; 215 if (p->gc_mode == GC_GREEDY) 216 return 2 * sbi->blocks_per_seg * p->ofs_unit; 217 else if (p->gc_mode == GC_CB) 218 return UINT_MAX; 219 else /* No other gc_mode */ 220 return 0; 221 } 222 223 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi) 224 { 225 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 226 unsigned int secno; 227 228 /* 229 * If the gc_type is FG_GC, we can select victim segments 230 * selected by background GC before. 231 * Those segments guarantee they have small valid blocks. 232 */ 233 for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) { 234 if (sec_usage_check(sbi, secno)) 235 continue; 236 clear_bit(secno, dirty_i->victim_secmap); 237 return GET_SEG_FROM_SEC(sbi, secno); 238 } 239 return NULL_SEGNO; 240 } 241 242 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno) 243 { 244 struct sit_info *sit_i = SIT_I(sbi); 245 unsigned int secno = GET_SEC_FROM_SEG(sbi, segno); 246 unsigned int start = GET_SEG_FROM_SEC(sbi, secno); 247 unsigned long long mtime = 0; 248 unsigned int vblocks; 249 unsigned char age = 0; 250 unsigned char u; 251 unsigned int i; 252 253 for (i = 0; i < sbi->segs_per_sec; i++) 254 mtime += get_seg_entry(sbi, start + i)->mtime; 255 vblocks = get_valid_blocks(sbi, segno, true); 256 257 mtime = div_u64(mtime, sbi->segs_per_sec); 258 vblocks = div_u64(vblocks, sbi->segs_per_sec); 259 260 u = (vblocks * 100) >> sbi->log_blocks_per_seg; 261 262 /* Handle if the system time has changed by the user */ 263 if (mtime < sit_i->min_mtime) 264 sit_i->min_mtime = mtime; 265 if (mtime > sit_i->max_mtime) 266 sit_i->max_mtime = mtime; 267 if (sit_i->max_mtime != sit_i->min_mtime) 268 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime), 269 sit_i->max_mtime - sit_i->min_mtime); 270 271 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u)); 272 } 273 274 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi, 275 unsigned int segno, struct victim_sel_policy *p) 276 { 277 if (p->alloc_mode == SSR) 278 return get_seg_entry(sbi, segno)->ckpt_valid_blocks; 279 280 /* alloc_mode == LFS */ 281 if (p->gc_mode == GC_GREEDY) 282 return get_valid_blocks(sbi, segno, true); 283 else 284 return get_cb_cost(sbi, segno); 285 } 286 287 static unsigned int count_bits(const unsigned long *addr, 288 unsigned int offset, unsigned int len) 289 { 290 unsigned int end = offset + len, sum = 0; 291 292 while (offset < end) { 293 if (test_bit(offset++, addr)) 294 ++sum; 295 } 296 return sum; 297 } 298 299 /* 300 * This function is called from two paths. 301 * One is garbage collection and the other is SSR segment selection. 302 * When it is called during GC, it just gets a victim segment 303 * and it does not remove it from dirty seglist. 304 * When it is called from SSR segment selection, it finds a segment 305 * which has minimum valid blocks and removes it from dirty seglist. 306 */ 307 static int get_victim_by_default(struct f2fs_sb_info *sbi, 308 unsigned int *result, int gc_type, int type, char alloc_mode) 309 { 310 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 311 struct sit_info *sm = SIT_I(sbi); 312 struct victim_sel_policy p; 313 unsigned int secno, last_victim; 314 unsigned int last_segment = MAIN_SEGS(sbi); 315 unsigned int nsearched = 0; 316 317 mutex_lock(&dirty_i->seglist_lock); 318 319 p.alloc_mode = alloc_mode; 320 select_policy(sbi, gc_type, type, &p); 321 322 p.min_segno = NULL_SEGNO; 323 p.min_cost = get_max_cost(sbi, &p); 324 325 if (*result != NULL_SEGNO) { 326 if (IS_DATASEG(get_seg_entry(sbi, *result)->type) && 327 get_valid_blocks(sbi, *result, false) && 328 !sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result))) 329 p.min_segno = *result; 330 goto out; 331 } 332 333 if (p.max_search == 0) 334 goto out; 335 336 last_victim = sm->last_victim[p.gc_mode]; 337 if (p.alloc_mode == LFS && gc_type == FG_GC) { 338 p.min_segno = check_bg_victims(sbi); 339 if (p.min_segno != NULL_SEGNO) 340 goto got_it; 341 } 342 343 while (1) { 344 unsigned long cost; 345 unsigned int segno; 346 347 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset); 348 if (segno >= last_segment) { 349 if (sm->last_victim[p.gc_mode]) { 350 last_segment = 351 sm->last_victim[p.gc_mode]; 352 sm->last_victim[p.gc_mode] = 0; 353 p.offset = 0; 354 continue; 355 } 356 break; 357 } 358 359 p.offset = segno + p.ofs_unit; 360 if (p.ofs_unit > 1) { 361 p.offset -= segno % p.ofs_unit; 362 nsearched += count_bits(p.dirty_segmap, 363 p.offset - p.ofs_unit, 364 p.ofs_unit); 365 } else { 366 nsearched++; 367 } 368 369 secno = GET_SEC_FROM_SEG(sbi, segno); 370 371 if (sec_usage_check(sbi, secno)) 372 goto next; 373 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap)) 374 goto next; 375 376 cost = get_gc_cost(sbi, segno, &p); 377 378 if (p.min_cost > cost) { 379 p.min_segno = segno; 380 p.min_cost = cost; 381 } 382 next: 383 if (nsearched >= p.max_search) { 384 if (!sm->last_victim[p.gc_mode] && segno <= last_victim) 385 sm->last_victim[p.gc_mode] = last_victim + 1; 386 else 387 sm->last_victim[p.gc_mode] = segno + 1; 388 sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi); 389 break; 390 } 391 } 392 if (p.min_segno != NULL_SEGNO) { 393 got_it: 394 if (p.alloc_mode == LFS) { 395 secno = GET_SEC_FROM_SEG(sbi, p.min_segno); 396 if (gc_type == FG_GC) 397 sbi->cur_victim_sec = secno; 398 else 399 set_bit(secno, dirty_i->victim_secmap); 400 } 401 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit; 402 403 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p, 404 sbi->cur_victim_sec, 405 prefree_segments(sbi), free_segments(sbi)); 406 } 407 out: 408 mutex_unlock(&dirty_i->seglist_lock); 409 410 return (p.min_segno == NULL_SEGNO) ? 0 : 1; 411 } 412 413 static const struct victim_selection default_v_ops = { 414 .get_victim = get_victim_by_default, 415 }; 416 417 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino) 418 { 419 struct inode_entry *ie; 420 421 ie = radix_tree_lookup(&gc_list->iroot, ino); 422 if (ie) 423 return ie->inode; 424 return NULL; 425 } 426 427 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode) 428 { 429 struct inode_entry *new_ie; 430 431 if (inode == find_gc_inode(gc_list, inode->i_ino)) { 432 iput(inode); 433 return; 434 } 435 new_ie = f2fs_kmem_cache_alloc(f2fs_inode_entry_slab, GFP_NOFS); 436 new_ie->inode = inode; 437 438 f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie); 439 list_add_tail(&new_ie->list, &gc_list->ilist); 440 } 441 442 static void put_gc_inode(struct gc_inode_list *gc_list) 443 { 444 struct inode_entry *ie, *next_ie; 445 list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) { 446 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino); 447 iput(ie->inode); 448 list_del(&ie->list); 449 kmem_cache_free(f2fs_inode_entry_slab, ie); 450 } 451 } 452 453 static int check_valid_map(struct f2fs_sb_info *sbi, 454 unsigned int segno, int offset) 455 { 456 struct sit_info *sit_i = SIT_I(sbi); 457 struct seg_entry *sentry; 458 int ret; 459 460 down_read(&sit_i->sentry_lock); 461 sentry = get_seg_entry(sbi, segno); 462 ret = f2fs_test_bit(offset, sentry->cur_valid_map); 463 up_read(&sit_i->sentry_lock); 464 return ret; 465 } 466 467 /* 468 * This function compares node address got in summary with that in NAT. 469 * On validity, copy that node with cold status, otherwise (invalid node) 470 * ignore that. 471 */ 472 static void gc_node_segment(struct f2fs_sb_info *sbi, 473 struct f2fs_summary *sum, unsigned int segno, int gc_type) 474 { 475 struct f2fs_summary *entry; 476 block_t start_addr; 477 int off; 478 int phase = 0; 479 bool fggc = (gc_type == FG_GC); 480 481 start_addr = START_BLOCK(sbi, segno); 482 483 next_step: 484 entry = sum; 485 486 if (fggc && phase == 2) 487 atomic_inc(&sbi->wb_sync_req[NODE]); 488 489 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) { 490 nid_t nid = le32_to_cpu(entry->nid); 491 struct page *node_page; 492 struct node_info ni; 493 494 /* stop BG_GC if there is not enough free sections. */ 495 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) 496 return; 497 498 if (check_valid_map(sbi, segno, off) == 0) 499 continue; 500 501 if (phase == 0) { 502 f2fs_ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1, 503 META_NAT, true); 504 continue; 505 } 506 507 if (phase == 1) { 508 f2fs_ra_node_page(sbi, nid); 509 continue; 510 } 511 512 /* phase == 2 */ 513 node_page = f2fs_get_node_page(sbi, nid); 514 if (IS_ERR(node_page)) 515 continue; 516 517 /* block may become invalid during f2fs_get_node_page */ 518 if (check_valid_map(sbi, segno, off) == 0) { 519 f2fs_put_page(node_page, 1); 520 continue; 521 } 522 523 if (f2fs_get_node_info(sbi, nid, &ni)) { 524 f2fs_put_page(node_page, 1); 525 continue; 526 } 527 528 if (ni.blk_addr != start_addr + off) { 529 f2fs_put_page(node_page, 1); 530 continue; 531 } 532 533 f2fs_move_node_page(node_page, gc_type); 534 stat_inc_node_blk_count(sbi, 1, gc_type); 535 } 536 537 if (++phase < 3) 538 goto next_step; 539 540 if (fggc) 541 atomic_dec(&sbi->wb_sync_req[NODE]); 542 } 543 544 /* 545 * Calculate start block index indicating the given node offset. 546 * Be careful, caller should give this node offset only indicating direct node 547 * blocks. If any node offsets, which point the other types of node blocks such 548 * as indirect or double indirect node blocks, are given, it must be a caller's 549 * bug. 550 */ 551 block_t f2fs_start_bidx_of_node(unsigned int node_ofs, struct inode *inode) 552 { 553 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4; 554 unsigned int bidx; 555 556 if (node_ofs == 0) 557 return 0; 558 559 if (node_ofs <= 2) { 560 bidx = node_ofs - 1; 561 } else if (node_ofs <= indirect_blks) { 562 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1); 563 bidx = node_ofs - 2 - dec; 564 } else { 565 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1); 566 bidx = node_ofs - 5 - dec; 567 } 568 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode); 569 } 570 571 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 572 struct node_info *dni, block_t blkaddr, unsigned int *nofs) 573 { 574 struct page *node_page; 575 nid_t nid; 576 unsigned int ofs_in_node; 577 block_t source_blkaddr; 578 579 nid = le32_to_cpu(sum->nid); 580 ofs_in_node = le16_to_cpu(sum->ofs_in_node); 581 582 node_page = f2fs_get_node_page(sbi, nid); 583 if (IS_ERR(node_page)) 584 return false; 585 586 if (f2fs_get_node_info(sbi, nid, dni)) { 587 f2fs_put_page(node_page, 1); 588 return false; 589 } 590 591 if (sum->version != dni->version) { 592 f2fs_msg(sbi->sb, KERN_WARNING, 593 "%s: valid data with mismatched node version.", 594 __func__); 595 set_sbi_flag(sbi, SBI_NEED_FSCK); 596 } 597 598 *nofs = ofs_of_node(node_page); 599 source_blkaddr = datablock_addr(NULL, node_page, ofs_in_node); 600 f2fs_put_page(node_page, 1); 601 602 if (source_blkaddr != blkaddr) 603 return false; 604 return true; 605 } 606 607 static int ra_data_block(struct inode *inode, pgoff_t index) 608 { 609 struct f2fs_sb_info *sbi = F2FS_I_SB(inode); 610 struct address_space *mapping = inode->i_mapping; 611 struct dnode_of_data dn; 612 struct page *page; 613 struct extent_info ei = {0, 0, 0}; 614 struct f2fs_io_info fio = { 615 .sbi = sbi, 616 .ino = inode->i_ino, 617 .type = DATA, 618 .temp = COLD, 619 .op = REQ_OP_READ, 620 .op_flags = 0, 621 .encrypted_page = NULL, 622 .in_list = false, 623 .retry = false, 624 }; 625 int err; 626 627 page = f2fs_grab_cache_page(mapping, index, true); 628 if (!page) 629 return -ENOMEM; 630 631 if (f2fs_lookup_extent_cache(inode, index, &ei)) { 632 dn.data_blkaddr = ei.blk + index - ei.fofs; 633 goto got_it; 634 } 635 636 set_new_dnode(&dn, inode, NULL, NULL, 0); 637 err = f2fs_get_dnode_of_data(&dn, index, LOOKUP_NODE); 638 if (err) 639 goto put_page; 640 f2fs_put_dnode(&dn); 641 642 if (unlikely(!f2fs_is_valid_blkaddr(sbi, dn.data_blkaddr, 643 DATA_GENERIC))) { 644 err = -EFAULT; 645 goto put_page; 646 } 647 got_it: 648 /* read page */ 649 fio.page = page; 650 fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr; 651 652 fio.encrypted_page = f2fs_pagecache_get_page(META_MAPPING(sbi), 653 dn.data_blkaddr, 654 FGP_LOCK | FGP_CREAT, GFP_NOFS); 655 if (!fio.encrypted_page) { 656 err = -ENOMEM; 657 goto put_page; 658 } 659 660 err = f2fs_submit_page_bio(&fio); 661 if (err) 662 goto put_encrypted_page; 663 f2fs_put_page(fio.encrypted_page, 0); 664 f2fs_put_page(page, 1); 665 return 0; 666 put_encrypted_page: 667 f2fs_put_page(fio.encrypted_page, 1); 668 put_page: 669 f2fs_put_page(page, 1); 670 return err; 671 } 672 673 /* 674 * Move data block via META_MAPPING while keeping locked data page. 675 * This can be used to move blocks, aka LBAs, directly on disk. 676 */ 677 static void move_data_block(struct inode *inode, block_t bidx, 678 int gc_type, unsigned int segno, int off) 679 { 680 struct f2fs_io_info fio = { 681 .sbi = F2FS_I_SB(inode), 682 .ino = inode->i_ino, 683 .type = DATA, 684 .temp = COLD, 685 .op = REQ_OP_READ, 686 .op_flags = 0, 687 .encrypted_page = NULL, 688 .in_list = false, 689 .retry = false, 690 }; 691 struct dnode_of_data dn; 692 struct f2fs_summary sum; 693 struct node_info ni; 694 struct page *page, *mpage; 695 block_t newaddr; 696 int err; 697 bool lfs_mode = test_opt(fio.sbi, LFS); 698 699 /* do not read out */ 700 page = f2fs_grab_cache_page(inode->i_mapping, bidx, false); 701 if (!page) 702 return; 703 704 if (!check_valid_map(F2FS_I_SB(inode), segno, off)) 705 goto out; 706 707 if (f2fs_is_atomic_file(inode)) { 708 F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++; 709 F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++; 710 goto out; 711 } 712 713 if (f2fs_is_pinned_file(inode)) { 714 f2fs_pin_file_control(inode, true); 715 goto out; 716 } 717 718 set_new_dnode(&dn, inode, NULL, NULL, 0); 719 err = f2fs_get_dnode_of_data(&dn, bidx, LOOKUP_NODE); 720 if (err) 721 goto out; 722 723 if (unlikely(dn.data_blkaddr == NULL_ADDR)) { 724 ClearPageUptodate(page); 725 goto put_out; 726 } 727 728 /* 729 * don't cache encrypted data into meta inode until previous dirty 730 * data were writebacked to avoid racing between GC and flush. 731 */ 732 f2fs_wait_on_page_writeback(page, DATA, true); 733 734 err = f2fs_get_node_info(fio.sbi, dn.nid, &ni); 735 if (err) 736 goto put_out; 737 738 set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version); 739 740 /* read page */ 741 fio.page = page; 742 fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr; 743 744 if (lfs_mode) 745 down_write(&fio.sbi->io_order_lock); 746 747 f2fs_allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr, 748 &sum, CURSEG_COLD_DATA, NULL, false); 749 750 fio.encrypted_page = f2fs_pagecache_get_page(META_MAPPING(fio.sbi), 751 newaddr, FGP_LOCK | FGP_CREAT, GFP_NOFS); 752 if (!fio.encrypted_page) { 753 err = -ENOMEM; 754 goto recover_block; 755 } 756 757 mpage = f2fs_pagecache_get_page(META_MAPPING(fio.sbi), 758 fio.old_blkaddr, FGP_LOCK, GFP_NOFS); 759 if (mpage) { 760 bool updated = false; 761 762 if (PageUptodate(mpage)) { 763 memcpy(page_address(fio.encrypted_page), 764 page_address(mpage), PAGE_SIZE); 765 updated = true; 766 } 767 f2fs_put_page(mpage, 1); 768 invalidate_mapping_pages(META_MAPPING(fio.sbi), 769 fio.old_blkaddr, fio.old_blkaddr); 770 if (updated) 771 goto write_page; 772 } 773 774 err = f2fs_submit_page_bio(&fio); 775 if (err) 776 goto put_page_out; 777 778 /* write page */ 779 lock_page(fio.encrypted_page); 780 781 if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) { 782 err = -EIO; 783 goto put_page_out; 784 } 785 if (unlikely(!PageUptodate(fio.encrypted_page))) { 786 err = -EIO; 787 goto put_page_out; 788 } 789 790 write_page: 791 set_page_dirty(fio.encrypted_page); 792 f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true); 793 if (clear_page_dirty_for_io(fio.encrypted_page)) 794 dec_page_count(fio.sbi, F2FS_DIRTY_META); 795 796 set_page_writeback(fio.encrypted_page); 797 ClearPageError(page); 798 799 /* allocate block address */ 800 f2fs_wait_on_page_writeback(dn.node_page, NODE, true); 801 802 fio.op = REQ_OP_WRITE; 803 fio.op_flags = REQ_SYNC; 804 fio.new_blkaddr = newaddr; 805 f2fs_submit_page_write(&fio); 806 if (fio.retry) { 807 if (PageWriteback(fio.encrypted_page)) 808 end_page_writeback(fio.encrypted_page); 809 goto put_page_out; 810 } 811 812 f2fs_update_iostat(fio.sbi, FS_GC_DATA_IO, F2FS_BLKSIZE); 813 814 f2fs_update_data_blkaddr(&dn, newaddr); 815 set_inode_flag(inode, FI_APPEND_WRITE); 816 if (page->index == 0) 817 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN); 818 put_page_out: 819 f2fs_put_page(fio.encrypted_page, 1); 820 recover_block: 821 if (lfs_mode) 822 up_write(&fio.sbi->io_order_lock); 823 if (err) 824 f2fs_do_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr, 825 true, true); 826 put_out: 827 f2fs_put_dnode(&dn); 828 out: 829 f2fs_put_page(page, 1); 830 } 831 832 static void move_data_page(struct inode *inode, block_t bidx, int gc_type, 833 unsigned int segno, int off) 834 { 835 struct page *page; 836 837 page = f2fs_get_lock_data_page(inode, bidx, true); 838 if (IS_ERR(page)) 839 return; 840 841 if (!check_valid_map(F2FS_I_SB(inode), segno, off)) 842 goto out; 843 844 if (f2fs_is_atomic_file(inode)) { 845 F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++; 846 F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++; 847 goto out; 848 } 849 if (f2fs_is_pinned_file(inode)) { 850 if (gc_type == FG_GC) 851 f2fs_pin_file_control(inode, true); 852 goto out; 853 } 854 855 if (gc_type == BG_GC) { 856 if (PageWriteback(page)) 857 goto out; 858 set_page_dirty(page); 859 set_cold_data(page); 860 } else { 861 struct f2fs_io_info fio = { 862 .sbi = F2FS_I_SB(inode), 863 .ino = inode->i_ino, 864 .type = DATA, 865 .temp = COLD, 866 .op = REQ_OP_WRITE, 867 .op_flags = REQ_SYNC, 868 .old_blkaddr = NULL_ADDR, 869 .page = page, 870 .encrypted_page = NULL, 871 .need_lock = LOCK_REQ, 872 .io_type = FS_GC_DATA_IO, 873 }; 874 bool is_dirty = PageDirty(page); 875 int err; 876 877 retry: 878 set_page_dirty(page); 879 f2fs_wait_on_page_writeback(page, DATA, true); 880 if (clear_page_dirty_for_io(page)) { 881 inode_dec_dirty_pages(inode); 882 f2fs_remove_dirty_inode(inode); 883 } 884 885 set_cold_data(page); 886 887 err = f2fs_do_write_data_page(&fio); 888 if (err) { 889 clear_cold_data(page); 890 if (err == -ENOMEM) { 891 congestion_wait(BLK_RW_ASYNC, HZ/50); 892 goto retry; 893 } 894 if (is_dirty) 895 set_page_dirty(page); 896 } 897 } 898 out: 899 f2fs_put_page(page, 1); 900 } 901 902 /* 903 * This function tries to get parent node of victim data block, and identifies 904 * data block validity. If the block is valid, copy that with cold status and 905 * modify parent node. 906 * If the parent node is not valid or the data block address is different, 907 * the victim data block is ignored. 908 */ 909 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 910 struct gc_inode_list *gc_list, unsigned int segno, int gc_type) 911 { 912 struct super_block *sb = sbi->sb; 913 struct f2fs_summary *entry; 914 block_t start_addr; 915 int off; 916 int phase = 0; 917 918 start_addr = START_BLOCK(sbi, segno); 919 920 next_step: 921 entry = sum; 922 923 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) { 924 struct page *data_page; 925 struct inode *inode; 926 struct node_info dni; /* dnode info for the data */ 927 unsigned int ofs_in_node, nofs; 928 block_t start_bidx; 929 nid_t nid = le32_to_cpu(entry->nid); 930 931 /* stop BG_GC if there is not enough free sections. */ 932 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) 933 return; 934 935 if (check_valid_map(sbi, segno, off) == 0) 936 continue; 937 938 if (phase == 0) { 939 f2fs_ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1, 940 META_NAT, true); 941 continue; 942 } 943 944 if (phase == 1) { 945 f2fs_ra_node_page(sbi, nid); 946 continue; 947 } 948 949 /* Get an inode by ino with checking validity */ 950 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs)) 951 continue; 952 953 if (phase == 2) { 954 f2fs_ra_node_page(sbi, dni.ino); 955 continue; 956 } 957 958 ofs_in_node = le16_to_cpu(entry->ofs_in_node); 959 960 if (phase == 3) { 961 inode = f2fs_iget(sb, dni.ino); 962 if (IS_ERR(inode) || is_bad_inode(inode)) 963 continue; 964 965 if (!down_write_trylock( 966 &F2FS_I(inode)->i_gc_rwsem[WRITE])) { 967 iput(inode); 968 sbi->skipped_gc_rwsem++; 969 continue; 970 } 971 972 start_bidx = f2fs_start_bidx_of_node(nofs, inode) + 973 ofs_in_node; 974 975 if (f2fs_post_read_required(inode)) { 976 int err = ra_data_block(inode, start_bidx); 977 978 up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]); 979 if (err) { 980 iput(inode); 981 continue; 982 } 983 add_gc_inode(gc_list, inode); 984 continue; 985 } 986 987 data_page = f2fs_get_read_data_page(inode, 988 start_bidx, REQ_RAHEAD, true); 989 up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]); 990 if (IS_ERR(data_page)) { 991 iput(inode); 992 continue; 993 } 994 995 f2fs_put_page(data_page, 0); 996 add_gc_inode(gc_list, inode); 997 continue; 998 } 999 1000 /* phase 4 */ 1001 inode = find_gc_inode(gc_list, dni.ino); 1002 if (inode) { 1003 struct f2fs_inode_info *fi = F2FS_I(inode); 1004 bool locked = false; 1005 1006 if (S_ISREG(inode->i_mode)) { 1007 if (!down_write_trylock(&fi->i_gc_rwsem[READ])) 1008 continue; 1009 if (!down_write_trylock( 1010 &fi->i_gc_rwsem[WRITE])) { 1011 sbi->skipped_gc_rwsem++; 1012 up_write(&fi->i_gc_rwsem[READ]); 1013 continue; 1014 } 1015 locked = true; 1016 1017 /* wait for all inflight aio data */ 1018 inode_dio_wait(inode); 1019 } 1020 1021 start_bidx = f2fs_start_bidx_of_node(nofs, inode) 1022 + ofs_in_node; 1023 if (f2fs_post_read_required(inode)) 1024 move_data_block(inode, start_bidx, gc_type, 1025 segno, off); 1026 else 1027 move_data_page(inode, start_bidx, gc_type, 1028 segno, off); 1029 1030 if (locked) { 1031 up_write(&fi->i_gc_rwsem[WRITE]); 1032 up_write(&fi->i_gc_rwsem[READ]); 1033 } 1034 1035 stat_inc_data_blk_count(sbi, 1, gc_type); 1036 } 1037 } 1038 1039 if (++phase < 5) 1040 goto next_step; 1041 } 1042 1043 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim, 1044 int gc_type) 1045 { 1046 struct sit_info *sit_i = SIT_I(sbi); 1047 int ret; 1048 1049 down_write(&sit_i->sentry_lock); 1050 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, 1051 NO_CHECK_TYPE, LFS); 1052 up_write(&sit_i->sentry_lock); 1053 return ret; 1054 } 1055 1056 static int do_garbage_collect(struct f2fs_sb_info *sbi, 1057 unsigned int start_segno, 1058 struct gc_inode_list *gc_list, int gc_type) 1059 { 1060 struct page *sum_page; 1061 struct f2fs_summary_block *sum; 1062 struct blk_plug plug; 1063 unsigned int segno = start_segno; 1064 unsigned int end_segno = start_segno + sbi->segs_per_sec; 1065 int seg_freed = 0; 1066 unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ? 1067 SUM_TYPE_DATA : SUM_TYPE_NODE; 1068 1069 /* readahead multi ssa blocks those have contiguous address */ 1070 if (sbi->segs_per_sec > 1) 1071 f2fs_ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), 1072 sbi->segs_per_sec, META_SSA, true); 1073 1074 /* reference all summary page */ 1075 while (segno < end_segno) { 1076 sum_page = f2fs_get_sum_page(sbi, segno++); 1077 if (IS_ERR(sum_page)) { 1078 int err = PTR_ERR(sum_page); 1079 1080 end_segno = segno - 1; 1081 for (segno = start_segno; segno < end_segno; segno++) { 1082 sum_page = find_get_page(META_MAPPING(sbi), 1083 GET_SUM_BLOCK(sbi, segno)); 1084 f2fs_put_page(sum_page, 0); 1085 f2fs_put_page(sum_page, 0); 1086 } 1087 return err; 1088 } 1089 unlock_page(sum_page); 1090 } 1091 1092 blk_start_plug(&plug); 1093 1094 for (segno = start_segno; segno < end_segno; segno++) { 1095 1096 /* find segment summary of victim */ 1097 sum_page = find_get_page(META_MAPPING(sbi), 1098 GET_SUM_BLOCK(sbi, segno)); 1099 f2fs_put_page(sum_page, 0); 1100 1101 if (get_valid_blocks(sbi, segno, false) == 0 || 1102 !PageUptodate(sum_page) || 1103 unlikely(f2fs_cp_error(sbi))) 1104 goto next; 1105 1106 sum = page_address(sum_page); 1107 if (type != GET_SUM_TYPE((&sum->footer))) { 1108 f2fs_msg(sbi->sb, KERN_ERR, "Inconsistent segment (%u) " 1109 "type [%d, %d] in SSA and SIT", 1110 segno, type, GET_SUM_TYPE((&sum->footer))); 1111 set_sbi_flag(sbi, SBI_NEED_FSCK); 1112 goto next; 1113 } 1114 1115 /* 1116 * this is to avoid deadlock: 1117 * - lock_page(sum_page) - f2fs_replace_block 1118 * - check_valid_map() - down_write(sentry_lock) 1119 * - down_read(sentry_lock) - change_curseg() 1120 * - lock_page(sum_page) 1121 */ 1122 if (type == SUM_TYPE_NODE) 1123 gc_node_segment(sbi, sum->entries, segno, gc_type); 1124 else 1125 gc_data_segment(sbi, sum->entries, gc_list, segno, 1126 gc_type); 1127 1128 stat_inc_seg_count(sbi, type, gc_type); 1129 1130 if (gc_type == FG_GC && 1131 get_valid_blocks(sbi, segno, false) == 0) 1132 seg_freed++; 1133 next: 1134 f2fs_put_page(sum_page, 0); 1135 } 1136 1137 if (gc_type == FG_GC) 1138 f2fs_submit_merged_write(sbi, 1139 (type == SUM_TYPE_NODE) ? NODE : DATA); 1140 1141 blk_finish_plug(&plug); 1142 1143 stat_inc_call_count(sbi->stat_info); 1144 1145 return seg_freed; 1146 } 1147 1148 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, 1149 bool background, unsigned int segno) 1150 { 1151 int gc_type = sync ? FG_GC : BG_GC; 1152 int sec_freed = 0, seg_freed = 0, total_freed = 0; 1153 int ret = 0; 1154 struct cp_control cpc; 1155 unsigned int init_segno = segno; 1156 struct gc_inode_list gc_list = { 1157 .ilist = LIST_HEAD_INIT(gc_list.ilist), 1158 .iroot = RADIX_TREE_INIT(gc_list.iroot, GFP_NOFS), 1159 }; 1160 unsigned long long last_skipped = sbi->skipped_atomic_files[FG_GC]; 1161 unsigned long long first_skipped; 1162 unsigned int skipped_round = 0, round = 0; 1163 1164 trace_f2fs_gc_begin(sbi->sb, sync, background, 1165 get_pages(sbi, F2FS_DIRTY_NODES), 1166 get_pages(sbi, F2FS_DIRTY_DENTS), 1167 get_pages(sbi, F2FS_DIRTY_IMETA), 1168 free_sections(sbi), 1169 free_segments(sbi), 1170 reserved_segments(sbi), 1171 prefree_segments(sbi)); 1172 1173 cpc.reason = __get_cp_reason(sbi); 1174 sbi->skipped_gc_rwsem = 0; 1175 first_skipped = last_skipped; 1176 gc_more: 1177 if (unlikely(!(sbi->sb->s_flags & SB_ACTIVE))) { 1178 ret = -EINVAL; 1179 goto stop; 1180 } 1181 if (unlikely(f2fs_cp_error(sbi))) { 1182 ret = -EIO; 1183 goto stop; 1184 } 1185 1186 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) { 1187 /* 1188 * For example, if there are many prefree_segments below given 1189 * threshold, we can make them free by checkpoint. Then, we 1190 * secure free segments which doesn't need fggc any more. 1191 */ 1192 if (prefree_segments(sbi)) { 1193 ret = f2fs_write_checkpoint(sbi, &cpc); 1194 if (ret) 1195 goto stop; 1196 } 1197 if (has_not_enough_free_secs(sbi, 0, 0)) 1198 gc_type = FG_GC; 1199 } 1200 1201 /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */ 1202 if (gc_type == BG_GC && !background) { 1203 ret = -EINVAL; 1204 goto stop; 1205 } 1206 if (!__get_victim(sbi, &segno, gc_type)) { 1207 ret = -ENODATA; 1208 goto stop; 1209 } 1210 1211 seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type); 1212 if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec) 1213 sec_freed++; 1214 total_freed += seg_freed; 1215 1216 if (gc_type == FG_GC) { 1217 if (sbi->skipped_atomic_files[FG_GC] > last_skipped || 1218 sbi->skipped_gc_rwsem) 1219 skipped_round++; 1220 last_skipped = sbi->skipped_atomic_files[FG_GC]; 1221 round++; 1222 } 1223 1224 if (gc_type == FG_GC) 1225 sbi->cur_victim_sec = NULL_SEGNO; 1226 1227 if (sync) 1228 goto stop; 1229 1230 if (has_not_enough_free_secs(sbi, sec_freed, 0)) { 1231 if (skipped_round <= MAX_SKIP_GC_COUNT || 1232 skipped_round * 2 < round) { 1233 segno = NULL_SEGNO; 1234 goto gc_more; 1235 } 1236 1237 if (first_skipped < last_skipped && 1238 (last_skipped - first_skipped) > 1239 sbi->skipped_gc_rwsem) { 1240 f2fs_drop_inmem_pages_all(sbi, true); 1241 segno = NULL_SEGNO; 1242 goto gc_more; 1243 } 1244 if (gc_type == FG_GC) 1245 ret = f2fs_write_checkpoint(sbi, &cpc); 1246 } 1247 stop: 1248 SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0; 1249 SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno; 1250 1251 trace_f2fs_gc_end(sbi->sb, ret, total_freed, sec_freed, 1252 get_pages(sbi, F2FS_DIRTY_NODES), 1253 get_pages(sbi, F2FS_DIRTY_DENTS), 1254 get_pages(sbi, F2FS_DIRTY_IMETA), 1255 free_sections(sbi), 1256 free_segments(sbi), 1257 reserved_segments(sbi), 1258 prefree_segments(sbi)); 1259 1260 mutex_unlock(&sbi->gc_mutex); 1261 1262 put_gc_inode(&gc_list); 1263 1264 if (sync && !ret) 1265 ret = sec_freed ? 0 : -EAGAIN; 1266 return ret; 1267 } 1268 1269 void f2fs_build_gc_manager(struct f2fs_sb_info *sbi) 1270 { 1271 DIRTY_I(sbi)->v_ops = &default_v_ops; 1272 1273 sbi->gc_pin_file_threshold = DEF_GC_FAILED_PINNED_FILES; 1274 1275 /* give warm/cold data area from slower device */ 1276 if (sbi->s_ndevs && sbi->segs_per_sec == 1) 1277 SIT_I(sbi)->last_victim[ALLOC_NEXT] = 1278 GET_SEGNO(sbi, FDEV(0).end_blk) + 1; 1279 } 1280