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/proc_fs.h> 15 #include <linux/init.h> 16 #include <linux/f2fs_fs.h> 17 #include <linux/kthread.h> 18 #include <linux/delay.h> 19 #include <linux/freezer.h> 20 #include <linux/blkdev.h> 21 22 #include "f2fs.h" 23 #include "node.h" 24 #include "segment.h" 25 #include "gc.h" 26 27 static struct kmem_cache *winode_slab; 28 29 static int gc_thread_func(void *data) 30 { 31 struct f2fs_sb_info *sbi = data; 32 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head; 33 long wait_ms; 34 35 wait_ms = GC_THREAD_MIN_SLEEP_TIME; 36 37 do { 38 if (try_to_freeze()) 39 continue; 40 else 41 wait_event_interruptible_timeout(*wq, 42 kthread_should_stop(), 43 msecs_to_jiffies(wait_ms)); 44 if (kthread_should_stop()) 45 break; 46 47 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) { 48 wait_ms = GC_THREAD_MAX_SLEEP_TIME; 49 continue; 50 } 51 52 /* 53 * [GC triggering condition] 54 * 0. GC is not conducted currently. 55 * 1. There are enough dirty segments. 56 * 2. IO subsystem is idle by checking the # of writeback pages. 57 * 3. IO subsystem is idle by checking the # of requests in 58 * bdev's request list. 59 * 60 * Note) We have to avoid triggering GCs too much frequently. 61 * Because it is possible that some segments can be 62 * invalidated soon after by user update or deletion. 63 * So, I'd like to wait some time to collect dirty segments. 64 */ 65 if (!mutex_trylock(&sbi->gc_mutex)) 66 continue; 67 68 if (!is_idle(sbi)) { 69 wait_ms = increase_sleep_time(wait_ms); 70 mutex_unlock(&sbi->gc_mutex); 71 continue; 72 } 73 74 if (has_enough_invalid_blocks(sbi)) 75 wait_ms = decrease_sleep_time(wait_ms); 76 else 77 wait_ms = increase_sleep_time(wait_ms); 78 79 sbi->bg_gc++; 80 81 /* if return value is not zero, no victim was selected */ 82 if (f2fs_gc(sbi)) 83 wait_ms = GC_THREAD_NOGC_SLEEP_TIME; 84 else if (wait_ms == GC_THREAD_NOGC_SLEEP_TIME) 85 wait_ms = GC_THREAD_MAX_SLEEP_TIME; 86 87 } while (!kthread_should_stop()); 88 return 0; 89 } 90 91 int start_gc_thread(struct f2fs_sb_info *sbi) 92 { 93 struct f2fs_gc_kthread *gc_th; 94 dev_t dev = sbi->sb->s_bdev->bd_dev; 95 96 if (!test_opt(sbi, BG_GC)) 97 return 0; 98 gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL); 99 if (!gc_th) 100 return -ENOMEM; 101 102 sbi->gc_thread = gc_th; 103 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head); 104 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi, 105 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev)); 106 if (IS_ERR(gc_th->f2fs_gc_task)) { 107 kfree(gc_th); 108 sbi->gc_thread = NULL; 109 return -ENOMEM; 110 } 111 return 0; 112 } 113 114 void stop_gc_thread(struct f2fs_sb_info *sbi) 115 { 116 struct f2fs_gc_kthread *gc_th = sbi->gc_thread; 117 if (!gc_th) 118 return; 119 kthread_stop(gc_th->f2fs_gc_task); 120 kfree(gc_th); 121 sbi->gc_thread = NULL; 122 } 123 124 static int select_gc_type(int gc_type) 125 { 126 return (gc_type == BG_GC) ? GC_CB : GC_GREEDY; 127 } 128 129 static void select_policy(struct f2fs_sb_info *sbi, int gc_type, 130 int type, struct victim_sel_policy *p) 131 { 132 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 133 134 if (p->alloc_mode) { 135 p->gc_mode = GC_GREEDY; 136 p->dirty_segmap = dirty_i->dirty_segmap[type]; 137 p->ofs_unit = 1; 138 } else { 139 p->gc_mode = select_gc_type(gc_type); 140 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY]; 141 p->ofs_unit = sbi->segs_per_sec; 142 } 143 p->offset = sbi->last_victim[p->gc_mode]; 144 } 145 146 static unsigned int get_max_cost(struct f2fs_sb_info *sbi, 147 struct victim_sel_policy *p) 148 { 149 /* SSR allocates in a segment unit */ 150 if (p->alloc_mode == SSR) 151 return 1 << sbi->log_blocks_per_seg; 152 if (p->gc_mode == GC_GREEDY) 153 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit; 154 else if (p->gc_mode == GC_CB) 155 return UINT_MAX; 156 else /* No other gc_mode */ 157 return 0; 158 } 159 160 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi) 161 { 162 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 163 unsigned int segno; 164 165 /* 166 * If the gc_type is FG_GC, we can select victim segments 167 * selected by background GC before. 168 * Those segments guarantee they have small valid blocks. 169 */ 170 segno = find_next_bit(dirty_i->victim_segmap[BG_GC], 171 TOTAL_SEGS(sbi), 0); 172 if (segno < TOTAL_SEGS(sbi)) { 173 clear_bit(segno, dirty_i->victim_segmap[BG_GC]); 174 return segno; 175 } 176 return NULL_SEGNO; 177 } 178 179 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno) 180 { 181 struct sit_info *sit_i = SIT_I(sbi); 182 unsigned int secno = GET_SECNO(sbi, segno); 183 unsigned int start = secno * sbi->segs_per_sec; 184 unsigned long long mtime = 0; 185 unsigned int vblocks; 186 unsigned char age = 0; 187 unsigned char u; 188 unsigned int i; 189 190 for (i = 0; i < sbi->segs_per_sec; i++) 191 mtime += get_seg_entry(sbi, start + i)->mtime; 192 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec); 193 194 mtime = div_u64(mtime, sbi->segs_per_sec); 195 vblocks = div_u64(vblocks, sbi->segs_per_sec); 196 197 u = (vblocks * 100) >> sbi->log_blocks_per_seg; 198 199 /* Handle if the system time is changed by user */ 200 if (mtime < sit_i->min_mtime) 201 sit_i->min_mtime = mtime; 202 if (mtime > sit_i->max_mtime) 203 sit_i->max_mtime = mtime; 204 if (sit_i->max_mtime != sit_i->min_mtime) 205 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime), 206 sit_i->max_mtime - sit_i->min_mtime); 207 208 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u)); 209 } 210 211 static unsigned int get_gc_cost(struct f2fs_sb_info *sbi, unsigned int segno, 212 struct victim_sel_policy *p) 213 { 214 if (p->alloc_mode == SSR) 215 return get_seg_entry(sbi, segno)->ckpt_valid_blocks; 216 217 /* alloc_mode == LFS */ 218 if (p->gc_mode == GC_GREEDY) 219 return get_valid_blocks(sbi, segno, sbi->segs_per_sec); 220 else 221 return get_cb_cost(sbi, segno); 222 } 223 224 /* 225 * This function is called from two pathes. 226 * One is garbage collection and the other is SSR segment selection. 227 * When it is called during GC, it just gets a victim segment 228 * and it does not remove it from dirty seglist. 229 * When it is called from SSR segment selection, it finds a segment 230 * which has minimum valid blocks and removes it from dirty seglist. 231 */ 232 static int get_victim_by_default(struct f2fs_sb_info *sbi, 233 unsigned int *result, int gc_type, int type, char alloc_mode) 234 { 235 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi); 236 struct victim_sel_policy p; 237 unsigned int segno; 238 int nsearched = 0; 239 240 p.alloc_mode = alloc_mode; 241 select_policy(sbi, gc_type, type, &p); 242 243 p.min_segno = NULL_SEGNO; 244 p.min_cost = get_max_cost(sbi, &p); 245 246 mutex_lock(&dirty_i->seglist_lock); 247 248 if (p.alloc_mode == LFS && gc_type == FG_GC) { 249 p.min_segno = check_bg_victims(sbi); 250 if (p.min_segno != NULL_SEGNO) 251 goto got_it; 252 } 253 254 while (1) { 255 unsigned long cost; 256 257 segno = find_next_bit(p.dirty_segmap, 258 TOTAL_SEGS(sbi), p.offset); 259 if (segno >= TOTAL_SEGS(sbi)) { 260 if (sbi->last_victim[p.gc_mode]) { 261 sbi->last_victim[p.gc_mode] = 0; 262 p.offset = 0; 263 continue; 264 } 265 break; 266 } 267 p.offset = ((segno / p.ofs_unit) * p.ofs_unit) + p.ofs_unit; 268 269 if (test_bit(segno, dirty_i->victim_segmap[FG_GC])) 270 continue; 271 if (gc_type == BG_GC && 272 test_bit(segno, dirty_i->victim_segmap[BG_GC])) 273 continue; 274 if (IS_CURSEC(sbi, GET_SECNO(sbi, segno))) 275 continue; 276 277 cost = get_gc_cost(sbi, segno, &p); 278 279 if (p.min_cost > cost) { 280 p.min_segno = segno; 281 p.min_cost = cost; 282 } 283 284 if (cost == get_max_cost(sbi, &p)) 285 continue; 286 287 if (nsearched++ >= MAX_VICTIM_SEARCH) { 288 sbi->last_victim[p.gc_mode] = segno; 289 break; 290 } 291 } 292 got_it: 293 if (p.min_segno != NULL_SEGNO) { 294 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit; 295 if (p.alloc_mode == LFS) { 296 int i; 297 for (i = 0; i < p.ofs_unit; i++) 298 set_bit(*result + i, 299 dirty_i->victim_segmap[gc_type]); 300 } 301 } 302 mutex_unlock(&dirty_i->seglist_lock); 303 304 return (p.min_segno == NULL_SEGNO) ? 0 : 1; 305 } 306 307 static const struct victim_selection default_v_ops = { 308 .get_victim = get_victim_by_default, 309 }; 310 311 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist) 312 { 313 struct list_head *this; 314 struct inode_entry *ie; 315 316 list_for_each(this, ilist) { 317 ie = list_entry(this, struct inode_entry, list); 318 if (ie->inode->i_ino == ino) 319 return ie->inode; 320 } 321 return NULL; 322 } 323 324 static void add_gc_inode(struct inode *inode, struct list_head *ilist) 325 { 326 struct list_head *this; 327 struct inode_entry *new_ie, *ie; 328 329 list_for_each(this, ilist) { 330 ie = list_entry(this, struct inode_entry, list); 331 if (ie->inode == inode) { 332 iput(inode); 333 return; 334 } 335 } 336 repeat: 337 new_ie = kmem_cache_alloc(winode_slab, GFP_NOFS); 338 if (!new_ie) { 339 cond_resched(); 340 goto repeat; 341 } 342 new_ie->inode = inode; 343 list_add_tail(&new_ie->list, ilist); 344 } 345 346 static void put_gc_inode(struct list_head *ilist) 347 { 348 struct inode_entry *ie, *next_ie; 349 list_for_each_entry_safe(ie, next_ie, ilist, list) { 350 iput(ie->inode); 351 list_del(&ie->list); 352 kmem_cache_free(winode_slab, ie); 353 } 354 } 355 356 static int check_valid_map(struct f2fs_sb_info *sbi, 357 unsigned int segno, int offset) 358 { 359 struct sit_info *sit_i = SIT_I(sbi); 360 struct seg_entry *sentry; 361 int ret; 362 363 mutex_lock(&sit_i->sentry_lock); 364 sentry = get_seg_entry(sbi, segno); 365 ret = f2fs_test_bit(offset, sentry->cur_valid_map); 366 mutex_unlock(&sit_i->sentry_lock); 367 return ret; 368 } 369 370 /* 371 * This function compares node address got in summary with that in NAT. 372 * On validity, copy that node with cold status, otherwise (invalid node) 373 * ignore that. 374 */ 375 static void gc_node_segment(struct f2fs_sb_info *sbi, 376 struct f2fs_summary *sum, unsigned int segno, int gc_type) 377 { 378 bool initial = true; 379 struct f2fs_summary *entry; 380 int off; 381 382 next_step: 383 entry = sum; 384 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) { 385 nid_t nid = le32_to_cpu(entry->nid); 386 struct page *node_page; 387 388 /* stop BG_GC if there is not enough free sections. */ 389 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0)) 390 return; 391 392 if (check_valid_map(sbi, segno, off) == 0) 393 continue; 394 395 if (initial) { 396 ra_node_page(sbi, nid); 397 continue; 398 } 399 node_page = get_node_page(sbi, nid); 400 if (IS_ERR(node_page)) 401 continue; 402 403 /* set page dirty and write it */ 404 if (!PageWriteback(node_page)) 405 set_page_dirty(node_page); 406 f2fs_put_page(node_page, 1); 407 stat_inc_node_blk_count(sbi, 1); 408 } 409 if (initial) { 410 initial = false; 411 goto next_step; 412 } 413 414 if (gc_type == FG_GC) { 415 struct writeback_control wbc = { 416 .sync_mode = WB_SYNC_ALL, 417 .nr_to_write = LONG_MAX, 418 .for_reclaim = 0, 419 }; 420 sync_node_pages(sbi, 0, &wbc); 421 } 422 } 423 424 /* 425 * Calculate start block index indicating the given node offset. 426 * Be careful, caller should give this node offset only indicating direct node 427 * blocks. If any node offsets, which point the other types of node blocks such 428 * as indirect or double indirect node blocks, are given, it must be a caller's 429 * bug. 430 */ 431 block_t start_bidx_of_node(unsigned int node_ofs) 432 { 433 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4; 434 unsigned int bidx; 435 436 if (node_ofs == 0) 437 return 0; 438 439 if (node_ofs <= 2) { 440 bidx = node_ofs - 1; 441 } else if (node_ofs <= indirect_blks) { 442 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1); 443 bidx = node_ofs - 2 - dec; 444 } else { 445 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1); 446 bidx = node_ofs - 5 - dec; 447 } 448 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE; 449 } 450 451 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 452 struct node_info *dni, block_t blkaddr, unsigned int *nofs) 453 { 454 struct page *node_page; 455 nid_t nid; 456 unsigned int ofs_in_node; 457 block_t source_blkaddr; 458 459 nid = le32_to_cpu(sum->nid); 460 ofs_in_node = le16_to_cpu(sum->ofs_in_node); 461 462 node_page = get_node_page(sbi, nid); 463 if (IS_ERR(node_page)) 464 return 0; 465 466 get_node_info(sbi, nid, dni); 467 468 if (sum->version != dni->version) { 469 f2fs_put_page(node_page, 1); 470 return 0; 471 } 472 473 *nofs = ofs_of_node(node_page); 474 source_blkaddr = datablock_addr(node_page, ofs_in_node); 475 f2fs_put_page(node_page, 1); 476 477 if (source_blkaddr != blkaddr) 478 return 0; 479 return 1; 480 } 481 482 static void move_data_page(struct inode *inode, struct page *page, int gc_type) 483 { 484 if (page->mapping != inode->i_mapping) 485 goto out; 486 487 if (inode != page->mapping->host) 488 goto out; 489 490 if (PageWriteback(page)) 491 goto out; 492 493 if (gc_type == BG_GC) { 494 set_page_dirty(page); 495 set_cold_data(page); 496 } else { 497 struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb); 498 mutex_lock_op(sbi, DATA_WRITE); 499 if (clear_page_dirty_for_io(page) && 500 S_ISDIR(inode->i_mode)) { 501 dec_page_count(sbi, F2FS_DIRTY_DENTS); 502 inode_dec_dirty_dents(inode); 503 } 504 set_cold_data(page); 505 do_write_data_page(page); 506 mutex_unlock_op(sbi, DATA_WRITE); 507 clear_cold_data(page); 508 } 509 out: 510 f2fs_put_page(page, 1); 511 } 512 513 /* 514 * This function tries to get parent node of victim data block, and identifies 515 * data block validity. If the block is valid, copy that with cold status and 516 * modify parent node. 517 * If the parent node is not valid or the data block address is different, 518 * the victim data block is ignored. 519 */ 520 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum, 521 struct list_head *ilist, unsigned int segno, int gc_type) 522 { 523 struct super_block *sb = sbi->sb; 524 struct f2fs_summary *entry; 525 block_t start_addr; 526 int off; 527 int phase = 0; 528 529 start_addr = START_BLOCK(sbi, segno); 530 531 next_step: 532 entry = sum; 533 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) { 534 struct page *data_page; 535 struct inode *inode; 536 struct node_info dni; /* dnode info for the data */ 537 unsigned int ofs_in_node, nofs; 538 block_t start_bidx; 539 540 /* stop BG_GC if there is not enough free sections. */ 541 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0)) 542 return; 543 544 if (check_valid_map(sbi, segno, off) == 0) 545 continue; 546 547 if (phase == 0) { 548 ra_node_page(sbi, le32_to_cpu(entry->nid)); 549 continue; 550 } 551 552 /* Get an inode by ino with checking validity */ 553 if (check_dnode(sbi, entry, &dni, start_addr + off, &nofs) == 0) 554 continue; 555 556 if (phase == 1) { 557 ra_node_page(sbi, dni.ino); 558 continue; 559 } 560 561 start_bidx = start_bidx_of_node(nofs); 562 ofs_in_node = le16_to_cpu(entry->ofs_in_node); 563 564 if (phase == 2) { 565 inode = f2fs_iget(sb, dni.ino); 566 if (IS_ERR(inode)) 567 continue; 568 569 data_page = find_data_page(inode, 570 start_bidx + ofs_in_node); 571 if (IS_ERR(data_page)) 572 goto next_iput; 573 574 f2fs_put_page(data_page, 0); 575 add_gc_inode(inode, ilist); 576 } else { 577 inode = find_gc_inode(dni.ino, ilist); 578 if (inode) { 579 data_page = get_lock_data_page(inode, 580 start_bidx + ofs_in_node); 581 if (IS_ERR(data_page)) 582 continue; 583 move_data_page(inode, data_page, gc_type); 584 stat_inc_data_blk_count(sbi, 1); 585 } 586 } 587 continue; 588 next_iput: 589 iput(inode); 590 } 591 if (++phase < 4) 592 goto next_step; 593 594 if (gc_type == FG_GC) 595 f2fs_submit_bio(sbi, DATA, true); 596 } 597 598 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim, 599 int gc_type, int type) 600 { 601 struct sit_info *sit_i = SIT_I(sbi); 602 int ret; 603 mutex_lock(&sit_i->sentry_lock); 604 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS); 605 mutex_unlock(&sit_i->sentry_lock); 606 return ret; 607 } 608 609 static void do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno, 610 struct list_head *ilist, int gc_type) 611 { 612 struct page *sum_page; 613 struct f2fs_summary_block *sum; 614 615 /* read segment summary of victim */ 616 sum_page = get_sum_page(sbi, segno); 617 if (IS_ERR(sum_page)) 618 return; 619 620 /* 621 * CP needs to lock sum_page. In this time, we don't need 622 * to lock this page, because this summary page is not gone anywhere. 623 * Also, this page is not gonna be updated before GC is done. 624 */ 625 unlock_page(sum_page); 626 sum = page_address(sum_page); 627 628 switch (GET_SUM_TYPE((&sum->footer))) { 629 case SUM_TYPE_NODE: 630 gc_node_segment(sbi, sum->entries, segno, gc_type); 631 break; 632 case SUM_TYPE_DATA: 633 gc_data_segment(sbi, sum->entries, ilist, segno, gc_type); 634 break; 635 } 636 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer))); 637 stat_inc_call_count(sbi->stat_info); 638 639 f2fs_put_page(sum_page, 0); 640 } 641 642 int f2fs_gc(struct f2fs_sb_info *sbi) 643 { 644 struct list_head ilist; 645 unsigned int segno, i; 646 int gc_type = BG_GC; 647 int nfree = 0; 648 int ret = -1; 649 650 INIT_LIST_HEAD(&ilist); 651 gc_more: 652 if (!(sbi->sb->s_flags & MS_ACTIVE)) 653 goto stop; 654 655 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) 656 gc_type = FG_GC; 657 658 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE)) 659 goto stop; 660 ret = 0; 661 662 for (i = 0; i < sbi->segs_per_sec; i++) 663 do_garbage_collect(sbi, segno + i, &ilist, gc_type); 664 665 if (gc_type == FG_GC && 666 get_valid_blocks(sbi, segno, sbi->segs_per_sec) == 0) 667 nfree++; 668 669 if (has_not_enough_free_secs(sbi, nfree)) 670 goto gc_more; 671 672 if (gc_type == FG_GC) 673 write_checkpoint(sbi, false); 674 stop: 675 mutex_unlock(&sbi->gc_mutex); 676 677 put_gc_inode(&ilist); 678 return ret; 679 } 680 681 void build_gc_manager(struct f2fs_sb_info *sbi) 682 { 683 DIRTY_I(sbi)->v_ops = &default_v_ops; 684 } 685 686 int __init create_gc_caches(void) 687 { 688 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes", 689 sizeof(struct inode_entry), NULL); 690 if (!winode_slab) 691 return -ENOMEM; 692 return 0; 693 } 694 695 void destroy_gc_caches(void) 696 { 697 kmem_cache_destroy(winode_slab); 698 } 699