1 /* 2 * fs/f2fs/gc.c 3 * 4 * Copyright (c) 2012 Samsung Electronics Co., Ltd. 5 * http://www.samsung.com/ 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License version 2 as 9 * published by the Free Software Foundation. 10 */ 11 #include <linux/fs.h> 12 #include <linux/module.h> 13 #include <linux/backing-dev.h> 14 #include <linux/init.h> 15 #include <linux/f2fs_fs.h> 16 #include <linux/kthread.h> 17 #include <linux/delay.h> 18 #include <linux/freezer.h> 19 20 #include "f2fs.h" 21 #include "node.h" 22 #include "segment.h" 23 #include "gc.h" 24 #include <trace/events/f2fs.h> 25 26 static int gc_thread_func(void *data) 27 { 28 struct f2fs_sb_info *sbi = data; 29 struct f2fs_gc_kthread *gc_th = sbi->gc_thread; 30 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head; 31 long wait_ms; 32 33 wait_ms = gc_th->min_sleep_time; 34 35 do { 36 if (try_to_freeze()) 37 continue; 38 else 39 wait_event_interruptible_timeout(*wq, 40 kthread_should_stop(), 41 msecs_to_jiffies(wait_ms)); 42 if (kthread_should_stop()) 43 break; 44 45 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) { 46 increase_sleep_time(gc_th, &wait_ms); 47 continue; 48 } 49 50 #ifdef CONFIG_F2FS_FAULT_INJECTION 51 if (time_to_inject(sbi, FAULT_CHECKPOINT)) { 52 f2fs_show_injection_info(FAULT_CHECKPOINT); 53 f2fs_stop_checkpoint(sbi, false); 54 } 55 #endif 56 57 /* 58 * [GC triggering condition] 59 * 0. GC is not conducted currently. 60 * 1. There are enough dirty segments. 61 * 2. IO subsystem is idle by checking the # of writeback pages. 62 * 3. IO subsystem is idle by checking the # of requests in 63 * bdev's request list. 64 * 65 * Note) We have to avoid triggering GCs frequently. 66 * Because it is possible that some segments can be 67 * invalidated soon after by user update or deletion. 68 * So, I'd like to wait some time to collect dirty segments. 69 */ 70 if (!mutex_trylock(&sbi->gc_mutex)) 71 continue; 72 73 if (!is_idle(sbi)) { 74 increase_sleep_time(gc_th, &wait_ms); 75 mutex_unlock(&sbi->gc_mutex); 76 continue; 77 } 78 79 if (has_enough_invalid_blocks(sbi)) 80 decrease_sleep_time(gc_th, &wait_ms); 81 else 82 increase_sleep_time(gc_th, &wait_ms); 83 84 stat_inc_bggc_count(sbi); 85 86 /* if return value is not zero, no victim was selected */ 87 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true)) 88 wait_ms = gc_th->no_gc_sleep_time; 89 90 trace_f2fs_background_gc(sbi->sb, wait_ms, 91 prefree_segments(sbi), free_segments(sbi)); 92 93 /* balancing f2fs's metadata periodically */ 94 f2fs_balance_fs_bg(sbi); 95 96 } while (!kthread_should_stop()); 97 return 0; 98 } 99 100 int start_gc_thread(struct f2fs_sb_info *sbi) 101 { 102 struct f2fs_gc_kthread *gc_th; 103 dev_t dev = sbi->sb->s_bdev->bd_dev; 104 int err = 0; 105 106 gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL); 107 if (!gc_th) { 108 err = -ENOMEM; 109 goto out; 110 } 111 112 gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME; 113 gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME; 114 gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME; 115 116 gc_th->gc_idle = 0; 117 118 sbi->gc_thread = gc_th; 119 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head); 120 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi, 121 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev)); 122 if (IS_ERR(gc_th->f2fs_gc_task)) { 123 err = PTR_ERR(gc_th->f2fs_gc_task); 124 kfree(gc_th); 125 sbi->gc_thread = NULL; 126 } 127 out: 128 return err; 129 } 130 131 void stop_gc_thread(struct f2fs_sb_info *sbi) 132 { 133 struct f2fs_gc_kthread *gc_th = sbi->gc_thread; 134 if (!gc_th) 135 return; 136 kthread_stop(gc_th->f2fs_gc_task); 137 kfree(gc_th); 138 sbi->gc_thread = NULL; 139 } 140 141 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type) 142 { 143 int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY; 144 145 if (gc_th && gc_th->gc_idle) { 146 if (gc_th->gc_idle == 1) 147 gc_mode = GC_CB; 148 else if (gc_th->gc_idle == 2) 149 gc_mode = GC_GREEDY; 150 } 151 return gc_mode; 152 } 153 154 static void select_policy(struct f2fs_sb_info *sbi, int gc_type, 155 int type, struct victim_sel_policy *p) 156 { 157 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 158 159 if (p->alloc_mode == SSR) { 160 p->gc_mode = GC_GREEDY; 161 p->dirty_segmap = dirty_i->dirty_segmap[type]; 162 p->max_search = dirty_i->nr_dirty[type]; 163 p->ofs_unit = 1; 164 } else { 165 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type); 166 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY]; 167 p->max_search = dirty_i->nr_dirty[DIRTY]; 168 p->ofs_unit = sbi->segs_per_sec; 169 } 170 171 /* we need to check every dirty segments in the FG_GC case */ 172 if (gc_type != FG_GC && p->max_search > sbi->max_victim_search) 173 p->max_search = sbi->max_victim_search; 174 175 p->offset = sbi->last_victim[p->gc_mode]; 176 } 177 178 static unsigned int get_max_cost(struct f2fs_sb_info *sbi, 179 struct victim_sel_policy *p) 180 { 181 /* SSR allocates in a segment unit */ 182 if (p->alloc_mode == SSR) 183 return sbi->blocks_per_seg; 184 if (p->gc_mode == GC_GREEDY) 185 return sbi->blocks_per_seg * p->ofs_unit; 186 else if (p->gc_mode == GC_CB) 187 return UINT_MAX; 188 else /* No other gc_mode */ 189 return 0; 190 } 191 192 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi) 193 { 194 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 195 unsigned int secno; 196 197 /* 198 * If the gc_type is FG_GC, we can select victim segments 199 * selected by background GC before. 200 * Those segments guarantee they have small valid blocks. 201 */ 202 for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) { 203 if (sec_usage_check(sbi, secno)) 204 continue; 205 206 if (no_fggc_candidate(sbi, secno)) 207 continue; 208 209 clear_bit(secno, dirty_i->victim_secmap); 210 return secno * sbi->segs_per_sec; 211 } 212 return NULL_SEGNO; 213 } 214 215 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno) 216 { 217 struct sit_info *sit_i = SIT_I(sbi); 218 unsigned int secno = GET_SECNO(sbi, segno); 219 unsigned int start = secno * sbi->segs_per_sec; 220 unsigned long long mtime = 0; 221 unsigned int vblocks; 222 unsigned char age = 0; 223 unsigned char u; 224 unsigned int i; 225 226 for (i = 0; i < sbi->segs_per_sec; i++) 227 mtime += get_seg_entry(sbi, start + i)->mtime; 228 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec); 229 230 mtime = div_u64(mtime, sbi->segs_per_sec); 231 vblocks = div_u64(vblocks, sbi->segs_per_sec); 232 233 u = (vblocks * 100) >> sbi->log_blocks_per_seg; 234 235 /* Handle if the system time has changed by the user */ 236 if (mtime < sit_i->min_mtime) 237 sit_i->min_mtime = mtime; 238 if (mtime > sit_i->max_mtime) 239 sit_i->max_mtime = mtime; 240 if (sit_i->max_mtime != sit_i->min_mtime) 241 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime), 242 sit_i->max_mtime - sit_i->min_mtime); 243 244 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u)); 245 } 246 247 static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi, 248 unsigned int segno) 249 { 250 unsigned int valid_blocks = 251 get_valid_blocks(sbi, segno, sbi->segs_per_sec); 252 253 return IS_DATASEG(get_seg_entry(sbi, segno)->type) ? 254 valid_blocks * 2 : valid_blocks; 255 } 256 257 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi, 258 unsigned int segno, struct victim_sel_policy *p) 259 { 260 if (p->alloc_mode == SSR) 261 return get_seg_entry(sbi, segno)->ckpt_valid_blocks; 262 263 /* alloc_mode == LFS */ 264 if (p->gc_mode == GC_GREEDY) 265 return get_greedy_cost(sbi, segno); 266 else 267 return get_cb_cost(sbi, segno); 268 } 269 270 static unsigned int count_bits(const unsigned long *addr, 271 unsigned int offset, unsigned int len) 272 { 273 unsigned int end = offset + len, sum = 0; 274 275 while (offset < end) { 276 if (test_bit(offset++, addr)) 277 ++sum; 278 } 279 return sum; 280 } 281 282 /* 283 * This function is called from two paths. 284 * One is garbage collection and the other is SSR segment selection. 285 * When it is called during GC, it just gets a victim segment 286 * and it does not remove it from dirty seglist. 287 * When it is called from SSR segment selection, it finds a segment 288 * which has minimum valid blocks and removes it from dirty seglist. 289 */ 290 static int get_victim_by_default(struct f2fs_sb_info *sbi, 291 unsigned int *result, int gc_type, int type, char alloc_mode) 292 { 293 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 294 struct victim_sel_policy p; 295 unsigned int secno, last_victim; 296 unsigned int last_segment = MAIN_SEGS(sbi); 297 unsigned int nsearched = 0; 298 299 mutex_lock(&dirty_i->seglist_lock); 300 301 p.alloc_mode = alloc_mode; 302 select_policy(sbi, gc_type, type, &p); 303 304 p.min_segno = NULL_SEGNO; 305 p.min_cost = get_max_cost(sbi, &p); 306 307 if (p.max_search == 0) 308 goto out; 309 310 last_victim = sbi->last_victim[p.gc_mode]; 311 if (p.alloc_mode == LFS && gc_type == FG_GC) { 312 p.min_segno = check_bg_victims(sbi); 313 if (p.min_segno != NULL_SEGNO) 314 goto got_it; 315 } 316 317 while (1) { 318 unsigned long cost; 319 unsigned int segno; 320 321 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset); 322 if (segno >= last_segment) { 323 if (sbi->last_victim[p.gc_mode]) { 324 last_segment = sbi->last_victim[p.gc_mode]; 325 sbi->last_victim[p.gc_mode] = 0; 326 p.offset = 0; 327 continue; 328 } 329 break; 330 } 331 332 p.offset = segno + p.ofs_unit; 333 if (p.ofs_unit > 1) { 334 p.offset -= segno % p.ofs_unit; 335 nsearched += count_bits(p.dirty_segmap, 336 p.offset - p.ofs_unit, 337 p.ofs_unit); 338 } else { 339 nsearched++; 340 } 341 342 secno = GET_SECNO(sbi, segno); 343 344 if (sec_usage_check(sbi, secno)) 345 goto next; 346 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap)) 347 goto next; 348 if (gc_type == FG_GC && p.alloc_mode == LFS && 349 no_fggc_candidate(sbi, secno)) 350 goto next; 351 352 cost = get_gc_cost(sbi, segno, &p); 353 354 if (p.min_cost > cost) { 355 p.min_segno = segno; 356 p.min_cost = cost; 357 } 358 next: 359 if (nsearched >= p.max_search) { 360 if (!sbi->last_victim[p.gc_mode] && segno <= last_victim) 361 sbi->last_victim[p.gc_mode] = last_victim + 1; 362 else 363 sbi->last_victim[p.gc_mode] = segno + 1; 364 break; 365 } 366 } 367 if (p.min_segno != NULL_SEGNO) { 368 got_it: 369 if (p.alloc_mode == LFS) { 370 secno = GET_SECNO(sbi, p.min_segno); 371 if (gc_type == FG_GC) 372 sbi->cur_victim_sec = secno; 373 else 374 set_bit(secno, dirty_i->victim_secmap); 375 } 376 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit; 377 378 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p, 379 sbi->cur_victim_sec, 380 prefree_segments(sbi), free_segments(sbi)); 381 } 382 out: 383 mutex_unlock(&dirty_i->seglist_lock); 384 385 return (p.min_segno == NULL_SEGNO) ? 0 : 1; 386 } 387 388 static const struct victim_selection default_v_ops = { 389 .get_victim = get_victim_by_default, 390 }; 391 392 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino) 393 { 394 struct inode_entry *ie; 395 396 ie = radix_tree_lookup(&gc_list->iroot, ino); 397 if (ie) 398 return ie->inode; 399 return NULL; 400 } 401 402 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode) 403 { 404 struct inode_entry *new_ie; 405 406 if (inode == find_gc_inode(gc_list, inode->i_ino)) { 407 iput(inode); 408 return; 409 } 410 new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS); 411 new_ie->inode = inode; 412 413 f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie); 414 list_add_tail(&new_ie->list, &gc_list->ilist); 415 } 416 417 static void put_gc_inode(struct gc_inode_list *gc_list) 418 { 419 struct inode_entry *ie, *next_ie; 420 list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) { 421 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino); 422 iput(ie->inode); 423 list_del(&ie->list); 424 kmem_cache_free(inode_entry_slab, ie); 425 } 426 } 427 428 static int check_valid_map(struct f2fs_sb_info *sbi, 429 unsigned int segno, int offset) 430 { 431 struct sit_info *sit_i = SIT_I(sbi); 432 struct seg_entry *sentry; 433 int ret; 434 435 mutex_lock(&sit_i->sentry_lock); 436 sentry = get_seg_entry(sbi, segno); 437 ret = f2fs_test_bit(offset, sentry->cur_valid_map); 438 mutex_unlock(&sit_i->sentry_lock); 439 return ret; 440 } 441 442 /* 443 * This function compares node address got in summary with that in NAT. 444 * On validity, copy that node with cold status, otherwise (invalid node) 445 * ignore that. 446 */ 447 static void gc_node_segment(struct f2fs_sb_info *sbi, 448 struct f2fs_summary *sum, unsigned int segno, int gc_type) 449 { 450 struct f2fs_summary *entry; 451 block_t start_addr; 452 int off; 453 int phase = 0; 454 455 start_addr = START_BLOCK(sbi, segno); 456 457 next_step: 458 entry = sum; 459 460 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) { 461 nid_t nid = le32_to_cpu(entry->nid); 462 struct page *node_page; 463 struct node_info ni; 464 465 /* stop BG_GC if there is not enough free sections. */ 466 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) 467 return; 468 469 if (check_valid_map(sbi, segno, off) == 0) 470 continue; 471 472 if (phase == 0) { 473 ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1, 474 META_NAT, true); 475 continue; 476 } 477 478 if (phase == 1) { 479 ra_node_page(sbi, nid); 480 continue; 481 } 482 483 /* phase == 2 */ 484 node_page = get_node_page(sbi, nid); 485 if (IS_ERR(node_page)) 486 continue; 487 488 /* block may become invalid during get_node_page */ 489 if (check_valid_map(sbi, segno, off) == 0) { 490 f2fs_put_page(node_page, 1); 491 continue; 492 } 493 494 get_node_info(sbi, nid, &ni); 495 if (ni.blk_addr != start_addr + off) { 496 f2fs_put_page(node_page, 1); 497 continue; 498 } 499 500 move_node_page(node_page, gc_type); 501 stat_inc_node_blk_count(sbi, 1, gc_type); 502 } 503 504 if (++phase < 3) 505 goto next_step; 506 } 507 508 /* 509 * Calculate start block index indicating the given node offset. 510 * Be careful, caller should give this node offset only indicating direct node 511 * blocks. If any node offsets, which point the other types of node blocks such 512 * as indirect or double indirect node blocks, are given, it must be a caller's 513 * bug. 514 */ 515 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode) 516 { 517 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4; 518 unsigned int bidx; 519 520 if (node_ofs == 0) 521 return 0; 522 523 if (node_ofs <= 2) { 524 bidx = node_ofs - 1; 525 } else if (node_ofs <= indirect_blks) { 526 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1); 527 bidx = node_ofs - 2 - dec; 528 } else { 529 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1); 530 bidx = node_ofs - 5 - dec; 531 } 532 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode); 533 } 534 535 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 536 struct node_info *dni, block_t blkaddr, unsigned int *nofs) 537 { 538 struct page *node_page; 539 nid_t nid; 540 unsigned int ofs_in_node; 541 block_t source_blkaddr; 542 543 nid = le32_to_cpu(sum->nid); 544 ofs_in_node = le16_to_cpu(sum->ofs_in_node); 545 546 node_page = get_node_page(sbi, nid); 547 if (IS_ERR(node_page)) 548 return false; 549 550 get_node_info(sbi, nid, dni); 551 552 if (sum->version != dni->version) { 553 f2fs_put_page(node_page, 1); 554 return false; 555 } 556 557 *nofs = ofs_of_node(node_page); 558 source_blkaddr = datablock_addr(node_page, ofs_in_node); 559 f2fs_put_page(node_page, 1); 560 561 if (source_blkaddr != blkaddr) 562 return false; 563 return true; 564 } 565 566 static void move_encrypted_block(struct inode *inode, block_t bidx, 567 unsigned int segno, int off) 568 { 569 struct f2fs_io_info fio = { 570 .sbi = F2FS_I_SB(inode), 571 .type = DATA, 572 .op = REQ_OP_READ, 573 .op_flags = 0, 574 .encrypted_page = NULL, 575 }; 576 struct dnode_of_data dn; 577 struct f2fs_summary sum; 578 struct node_info ni; 579 struct page *page; 580 block_t newaddr; 581 int err; 582 583 /* do not read out */ 584 page = f2fs_grab_cache_page(inode->i_mapping, bidx, false); 585 if (!page) 586 return; 587 588 if (!check_valid_map(F2FS_I_SB(inode), segno, off)) 589 goto out; 590 591 if (f2fs_is_atomic_file(inode)) 592 goto out; 593 594 set_new_dnode(&dn, inode, NULL, NULL, 0); 595 err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE); 596 if (err) 597 goto out; 598 599 if (unlikely(dn.data_blkaddr == NULL_ADDR)) { 600 ClearPageUptodate(page); 601 goto put_out; 602 } 603 604 /* 605 * don't cache encrypted data into meta inode until previous dirty 606 * data were writebacked to avoid racing between GC and flush. 607 */ 608 f2fs_wait_on_page_writeback(page, DATA, true); 609 610 get_node_info(fio.sbi, dn.nid, &ni); 611 set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version); 612 613 /* read page */ 614 fio.page = page; 615 fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr; 616 617 allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr, 618 &sum, CURSEG_COLD_DATA); 619 620 fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi), newaddr, 621 FGP_LOCK | FGP_CREAT, GFP_NOFS); 622 if (!fio.encrypted_page) { 623 err = -ENOMEM; 624 goto recover_block; 625 } 626 627 err = f2fs_submit_page_bio(&fio); 628 if (err) 629 goto put_page_out; 630 631 /* write page */ 632 lock_page(fio.encrypted_page); 633 634 if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) { 635 err = -EIO; 636 goto put_page_out; 637 } 638 if (unlikely(!PageUptodate(fio.encrypted_page))) { 639 err = -EIO; 640 goto put_page_out; 641 } 642 643 set_page_dirty(fio.encrypted_page); 644 f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true); 645 if (clear_page_dirty_for_io(fio.encrypted_page)) 646 dec_page_count(fio.sbi, F2FS_DIRTY_META); 647 648 set_page_writeback(fio.encrypted_page); 649 650 /* allocate block address */ 651 f2fs_wait_on_page_writeback(dn.node_page, NODE, true); 652 653 fio.op = REQ_OP_WRITE; 654 fio.op_flags = REQ_SYNC; 655 fio.new_blkaddr = newaddr; 656 f2fs_submit_page_mbio(&fio); 657 658 f2fs_update_data_blkaddr(&dn, newaddr); 659 set_inode_flag(inode, FI_APPEND_WRITE); 660 if (page->index == 0) 661 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN); 662 put_page_out: 663 f2fs_put_page(fio.encrypted_page, 1); 664 recover_block: 665 if (err) 666 __f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr, 667 true, true); 668 put_out: 669 f2fs_put_dnode(&dn); 670 out: 671 f2fs_put_page(page, 1); 672 } 673 674 static void move_data_page(struct inode *inode, block_t bidx, int gc_type, 675 unsigned int segno, int off) 676 { 677 struct page *page; 678 679 page = get_lock_data_page(inode, bidx, true); 680 if (IS_ERR(page)) 681 return; 682 683 if (!check_valid_map(F2FS_I_SB(inode), segno, off)) 684 goto out; 685 686 if (f2fs_is_atomic_file(inode)) 687 goto out; 688 689 if (gc_type == BG_GC) { 690 if (PageWriteback(page)) 691 goto out; 692 set_page_dirty(page); 693 set_cold_data(page); 694 } else { 695 struct f2fs_io_info fio = { 696 .sbi = F2FS_I_SB(inode), 697 .type = DATA, 698 .op = REQ_OP_WRITE, 699 .op_flags = REQ_SYNC, 700 .page = page, 701 .encrypted_page = NULL, 702 }; 703 bool is_dirty = PageDirty(page); 704 int err; 705 706 retry: 707 set_page_dirty(page); 708 f2fs_wait_on_page_writeback(page, DATA, true); 709 if (clear_page_dirty_for_io(page)) { 710 inode_dec_dirty_pages(inode); 711 remove_dirty_inode(inode); 712 } 713 714 set_cold_data(page); 715 716 err = do_write_data_page(&fio); 717 if (err == -ENOMEM && is_dirty) { 718 congestion_wait(BLK_RW_ASYNC, HZ/50); 719 goto retry; 720 } 721 } 722 out: 723 f2fs_put_page(page, 1); 724 } 725 726 /* 727 * This function tries to get parent node of victim data block, and identifies 728 * data block validity. If the block is valid, copy that with cold status and 729 * modify parent node. 730 * If the parent node is not valid or the data block address is different, 731 * the victim data block is ignored. 732 */ 733 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 734 struct gc_inode_list *gc_list, unsigned int segno, int gc_type) 735 { 736 struct super_block *sb = sbi->sb; 737 struct f2fs_summary *entry; 738 block_t start_addr; 739 int off; 740 int phase = 0; 741 742 start_addr = START_BLOCK(sbi, segno); 743 744 next_step: 745 entry = sum; 746 747 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) { 748 struct page *data_page; 749 struct inode *inode; 750 struct node_info dni; /* dnode info for the data */ 751 unsigned int ofs_in_node, nofs; 752 block_t start_bidx; 753 nid_t nid = le32_to_cpu(entry->nid); 754 755 /* stop BG_GC if there is not enough free sections. */ 756 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) 757 return; 758 759 if (check_valid_map(sbi, segno, off) == 0) 760 continue; 761 762 if (phase == 0) { 763 ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1, 764 META_NAT, true); 765 continue; 766 } 767 768 if (phase == 1) { 769 ra_node_page(sbi, nid); 770 continue; 771 } 772 773 /* Get an inode by ino with checking validity */ 774 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs)) 775 continue; 776 777 if (phase == 2) { 778 ra_node_page(sbi, dni.ino); 779 continue; 780 } 781 782 ofs_in_node = le16_to_cpu(entry->ofs_in_node); 783 784 if (phase == 3) { 785 inode = f2fs_iget(sb, dni.ino); 786 if (IS_ERR(inode) || is_bad_inode(inode)) 787 continue; 788 789 /* if encrypted inode, let's go phase 3 */ 790 if (f2fs_encrypted_inode(inode) && 791 S_ISREG(inode->i_mode)) { 792 add_gc_inode(gc_list, inode); 793 continue; 794 } 795 796 start_bidx = start_bidx_of_node(nofs, inode); 797 data_page = get_read_data_page(inode, 798 start_bidx + ofs_in_node, REQ_RAHEAD, 799 true); 800 if (IS_ERR(data_page)) { 801 iput(inode); 802 continue; 803 } 804 805 f2fs_put_page(data_page, 0); 806 add_gc_inode(gc_list, inode); 807 continue; 808 } 809 810 /* phase 4 */ 811 inode = find_gc_inode(gc_list, dni.ino); 812 if (inode) { 813 struct f2fs_inode_info *fi = F2FS_I(inode); 814 bool locked = false; 815 816 if (S_ISREG(inode->i_mode)) { 817 if (!down_write_trylock(&fi->dio_rwsem[READ])) 818 continue; 819 if (!down_write_trylock( 820 &fi->dio_rwsem[WRITE])) { 821 up_write(&fi->dio_rwsem[READ]); 822 continue; 823 } 824 locked = true; 825 } 826 827 start_bidx = start_bidx_of_node(nofs, inode) 828 + ofs_in_node; 829 if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode)) 830 move_encrypted_block(inode, start_bidx, segno, off); 831 else 832 move_data_page(inode, start_bidx, gc_type, segno, off); 833 834 if (locked) { 835 up_write(&fi->dio_rwsem[WRITE]); 836 up_write(&fi->dio_rwsem[READ]); 837 } 838 839 stat_inc_data_blk_count(sbi, 1, gc_type); 840 } 841 } 842 843 if (++phase < 5) 844 goto next_step; 845 } 846 847 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim, 848 int gc_type) 849 { 850 struct sit_info *sit_i = SIT_I(sbi); 851 int ret; 852 853 mutex_lock(&sit_i->sentry_lock); 854 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, 855 NO_CHECK_TYPE, LFS); 856 mutex_unlock(&sit_i->sentry_lock); 857 return ret; 858 } 859 860 static int do_garbage_collect(struct f2fs_sb_info *sbi, 861 unsigned int start_segno, 862 struct gc_inode_list *gc_list, int gc_type) 863 { 864 struct page *sum_page; 865 struct f2fs_summary_block *sum; 866 struct blk_plug plug; 867 unsigned int segno = start_segno; 868 unsigned int end_segno = start_segno + sbi->segs_per_sec; 869 int sec_freed = 0; 870 unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ? 871 SUM_TYPE_DATA : SUM_TYPE_NODE; 872 873 /* readahead multi ssa blocks those have contiguous address */ 874 if (sbi->segs_per_sec > 1) 875 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), 876 sbi->segs_per_sec, META_SSA, true); 877 878 /* reference all summary page */ 879 while (segno < end_segno) { 880 sum_page = get_sum_page(sbi, segno++); 881 unlock_page(sum_page); 882 } 883 884 blk_start_plug(&plug); 885 886 for (segno = start_segno; segno < end_segno; segno++) { 887 888 /* find segment summary of victim */ 889 sum_page = find_get_page(META_MAPPING(sbi), 890 GET_SUM_BLOCK(sbi, segno)); 891 f2fs_put_page(sum_page, 0); 892 893 if (get_valid_blocks(sbi, segno, 1) == 0 || 894 !PageUptodate(sum_page) || 895 unlikely(f2fs_cp_error(sbi))) 896 goto next; 897 898 sum = page_address(sum_page); 899 f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer))); 900 901 /* 902 * this is to avoid deadlock: 903 * - lock_page(sum_page) - f2fs_replace_block 904 * - check_valid_map() - mutex_lock(sentry_lock) 905 * - mutex_lock(sentry_lock) - change_curseg() 906 * - lock_page(sum_page) 907 */ 908 909 if (type == SUM_TYPE_NODE) 910 gc_node_segment(sbi, sum->entries, segno, gc_type); 911 else 912 gc_data_segment(sbi, sum->entries, gc_list, segno, 913 gc_type); 914 915 stat_inc_seg_count(sbi, type, gc_type); 916 next: 917 f2fs_put_page(sum_page, 0); 918 } 919 920 if (gc_type == FG_GC) 921 f2fs_submit_merged_bio(sbi, 922 (type == SUM_TYPE_NODE) ? NODE : DATA, WRITE); 923 924 blk_finish_plug(&plug); 925 926 if (gc_type == FG_GC && 927 get_valid_blocks(sbi, start_segno, sbi->segs_per_sec) == 0) 928 sec_freed = 1; 929 930 stat_inc_call_count(sbi->stat_info); 931 932 return sec_freed; 933 } 934 935 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background) 936 { 937 unsigned int segno; 938 int gc_type = sync ? FG_GC : BG_GC; 939 int sec_freed = 0; 940 int ret = -EINVAL; 941 struct cp_control cpc; 942 struct gc_inode_list gc_list = { 943 .ilist = LIST_HEAD_INIT(gc_list.ilist), 944 .iroot = RADIX_TREE_INIT(GFP_NOFS), 945 }; 946 947 cpc.reason = __get_cp_reason(sbi); 948 gc_more: 949 if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE))) 950 goto stop; 951 if (unlikely(f2fs_cp_error(sbi))) { 952 ret = -EIO; 953 goto stop; 954 } 955 956 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) { 957 /* 958 * For example, if there are many prefree_segments below given 959 * threshold, we can make them free by checkpoint. Then, we 960 * secure free segments which doesn't need fggc any more. 961 */ 962 ret = write_checkpoint(sbi, &cpc); 963 if (ret) 964 goto stop; 965 if (has_not_enough_free_secs(sbi, 0, 0)) 966 gc_type = FG_GC; 967 } 968 969 /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */ 970 if (gc_type == BG_GC && !background) 971 goto stop; 972 if (!__get_victim(sbi, &segno, gc_type)) 973 goto stop; 974 ret = 0; 975 976 if (do_garbage_collect(sbi, segno, &gc_list, gc_type) && 977 gc_type == FG_GC) 978 sec_freed++; 979 980 if (gc_type == FG_GC) 981 sbi->cur_victim_sec = NULL_SEGNO; 982 983 if (!sync) { 984 if (has_not_enough_free_secs(sbi, sec_freed, 0)) 985 goto gc_more; 986 987 if (gc_type == FG_GC) 988 ret = write_checkpoint(sbi, &cpc); 989 } 990 stop: 991 mutex_unlock(&sbi->gc_mutex); 992 993 put_gc_inode(&gc_list); 994 995 if (sync) 996 ret = sec_freed ? 0 : -EAGAIN; 997 return ret; 998 } 999 1000 void build_gc_manager(struct f2fs_sb_info *sbi) 1001 { 1002 u64 main_count, resv_count, ovp_count, blocks_per_sec; 1003 1004 DIRTY_I(sbi)->v_ops = &default_v_ops; 1005 1006 /* threshold of # of valid blocks in a section for victims of FG_GC */ 1007 main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg; 1008 resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg; 1009 ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg; 1010 blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec; 1011 1012 sbi->fggc_threshold = div64_u64((main_count - ovp_count) * blocks_per_sec, 1013 (main_count - resv_count)); 1014 } 1015