4 * Copyright (c) 2012 Samsung Electronics Co., Ltd.
5 * http://www.samsung.com/
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.
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 #include <linux/blkdev.h>
25 #include <trace/events/f2fs.h>
27 static struct kmem_cache *winode_slab;
29 static int gc_thread_func(void *data)
31 struct f2fs_sb_info *sbi = data;
32 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
33 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
36 wait_ms = gc_th->min_sleep_time;
42 wait_event_interruptible_timeout(*wq,
43 kthread_should_stop(),
44 msecs_to_jiffies(wait_ms));
45 if (kthread_should_stop())
48 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
49 wait_ms = increase_sleep_time(gc_th, wait_ms);
54 * [GC triggering condition]
55 * 0. GC is not conducted currently.
56 * 1. There are enough dirty segments.
57 * 2. IO subsystem is idle by checking the # of writeback pages.
58 * 3. IO subsystem is idle by checking the # of requests in
59 * bdev's request list.
61 * Note) We have to avoid triggering GCs too much frequently.
62 * Because it is possible that some segments can be
63 * invalidated soon after by user update or deletion.
64 * So, I'd like to wait some time to collect dirty segments.
66 if (!mutex_trylock(&sbi->gc_mutex))
70 wait_ms = increase_sleep_time(gc_th, wait_ms);
71 mutex_unlock(&sbi->gc_mutex);
75 if (has_enough_invalid_blocks(sbi))
76 wait_ms = decrease_sleep_time(gc_th, wait_ms);
78 wait_ms = increase_sleep_time(gc_th, wait_ms);
80 stat_inc_bggc_count(sbi);
82 /* if return value is not zero, no victim was selected */
84 wait_ms = gc_th->no_gc_sleep_time;
86 /* balancing f2fs's metadata periodically */
87 f2fs_balance_fs_bg(sbi);
89 } while (!kthread_should_stop());
93 int start_gc_thread(struct f2fs_sb_info *sbi)
95 struct f2fs_gc_kthread *gc_th;
96 dev_t dev = sbi->sb->s_bdev->bd_dev;
99 if (!test_opt(sbi, BG_GC))
101 gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
107 gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
108 gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
109 gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
113 sbi->gc_thread = gc_th;
114 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
115 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
116 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
117 if (IS_ERR(gc_th->f2fs_gc_task)) {
118 err = PTR_ERR(gc_th->f2fs_gc_task);
120 sbi->gc_thread = NULL;
126 void stop_gc_thread(struct f2fs_sb_info *sbi)
128 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
131 kthread_stop(gc_th->f2fs_gc_task);
133 sbi->gc_thread = NULL;
136 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
138 int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
140 if (gc_th && gc_th->gc_idle) {
141 if (gc_th->gc_idle == 1)
143 else if (gc_th->gc_idle == 2)
149 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
150 int type, struct victim_sel_policy *p)
152 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
154 if (p->alloc_mode == SSR) {
155 p->gc_mode = GC_GREEDY;
156 p->dirty_segmap = dirty_i->dirty_segmap[type];
157 p->max_search = dirty_i->nr_dirty[type];
160 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
161 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
162 p->max_search = dirty_i->nr_dirty[DIRTY];
163 p->ofs_unit = sbi->segs_per_sec;
166 if (p->max_search > sbi->max_victim_search)
167 p->max_search = sbi->max_victim_search;
169 p->offset = sbi->last_victim[p->gc_mode];
172 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
173 struct victim_sel_policy *p)
175 /* SSR allocates in a segment unit */
176 if (p->alloc_mode == SSR)
177 return 1 << sbi->log_blocks_per_seg;
178 if (p->gc_mode == GC_GREEDY)
179 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
180 else if (p->gc_mode == GC_CB)
182 else /* No other gc_mode */
186 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
188 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
189 unsigned int hint = 0;
193 * If the gc_type is FG_GC, we can select victim segments
194 * selected by background GC before.
195 * Those segments guarantee they have small valid blocks.
198 secno = find_next_bit(dirty_i->victim_secmap, TOTAL_SECS(sbi), hint++);
199 if (secno < TOTAL_SECS(sbi)) {
200 if (sec_usage_check(sbi, secno))
202 clear_bit(secno, dirty_i->victim_secmap);
203 return secno * sbi->segs_per_sec;
208 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
210 struct sit_info *sit_i = SIT_I(sbi);
211 unsigned int secno = GET_SECNO(sbi, segno);
212 unsigned int start = secno * sbi->segs_per_sec;
213 unsigned long long mtime = 0;
214 unsigned int vblocks;
215 unsigned char age = 0;
219 for (i = 0; i < sbi->segs_per_sec; i++)
220 mtime += get_seg_entry(sbi, start + i)->mtime;
221 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
223 mtime = div_u64(mtime, sbi->segs_per_sec);
224 vblocks = div_u64(vblocks, sbi->segs_per_sec);
226 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
228 /* Handle if the system time is changed by user */
229 if (mtime < sit_i->min_mtime)
230 sit_i->min_mtime = mtime;
231 if (mtime > sit_i->max_mtime)
232 sit_i->max_mtime = mtime;
233 if (sit_i->max_mtime != sit_i->min_mtime)
234 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
235 sit_i->max_mtime - sit_i->min_mtime);
237 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
240 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
241 unsigned int segno, struct victim_sel_policy *p)
243 if (p->alloc_mode == SSR)
244 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
246 /* alloc_mode == LFS */
247 if (p->gc_mode == GC_GREEDY)
248 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
250 return get_cb_cost(sbi, segno);
254 * This function is called from two paths.
255 * One is garbage collection and the other is SSR segment selection.
256 * When it is called during GC, it just gets a victim segment
257 * and it does not remove it from dirty seglist.
258 * When it is called from SSR segment selection, it finds a segment
259 * which has minimum valid blocks and removes it from dirty seglist.
261 static int get_victim_by_default(struct f2fs_sb_info *sbi,
262 unsigned int *result, int gc_type, int type, char alloc_mode)
264 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
265 struct victim_sel_policy p;
266 unsigned int secno, max_cost;
269 p.alloc_mode = alloc_mode;
270 select_policy(sbi, gc_type, type, &p);
272 p.min_segno = NULL_SEGNO;
273 p.min_cost = max_cost = get_max_cost(sbi, &p);
275 mutex_lock(&dirty_i->seglist_lock);
277 if (p.alloc_mode == LFS && gc_type == FG_GC) {
278 p.min_segno = check_bg_victims(sbi);
279 if (p.min_segno != NULL_SEGNO)
287 segno = find_next_bit(p.dirty_segmap,
288 TOTAL_SEGS(sbi), p.offset);
289 if (segno >= TOTAL_SEGS(sbi)) {
290 if (sbi->last_victim[p.gc_mode]) {
291 sbi->last_victim[p.gc_mode] = 0;
298 p.offset = segno + p.ofs_unit;
300 p.offset -= segno % p.ofs_unit;
302 secno = GET_SECNO(sbi, segno);
304 if (sec_usage_check(sbi, secno))
306 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
309 cost = get_gc_cost(sbi, segno, &p);
311 if (p.min_cost > cost) {
314 } else if (unlikely(cost == max_cost)) {
318 if (nsearched++ >= p.max_search) {
319 sbi->last_victim[p.gc_mode] = segno;
323 if (p.min_segno != NULL_SEGNO) {
325 if (p.alloc_mode == LFS) {
326 secno = GET_SECNO(sbi, p.min_segno);
327 if (gc_type == FG_GC)
328 sbi->cur_victim_sec = secno;
330 set_bit(secno, dirty_i->victim_secmap);
332 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
334 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
336 prefree_segments(sbi), free_segments(sbi));
338 mutex_unlock(&dirty_i->seglist_lock);
340 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
343 static const struct victim_selection default_v_ops = {
344 .get_victim = get_victim_by_default,
347 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist)
349 struct inode_entry *ie;
351 list_for_each_entry(ie, ilist, list)
352 if (ie->inode->i_ino == ino)
357 static void add_gc_inode(struct inode *inode, struct list_head *ilist)
359 struct inode_entry *new_ie;
361 if (inode == find_gc_inode(inode->i_ino, ilist)) {
366 new_ie = f2fs_kmem_cache_alloc(winode_slab, GFP_NOFS);
367 new_ie->inode = inode;
368 list_add_tail(&new_ie->list, ilist);
371 static void put_gc_inode(struct list_head *ilist)
373 struct inode_entry *ie, *next_ie;
374 list_for_each_entry_safe(ie, next_ie, ilist, list) {
377 kmem_cache_free(winode_slab, ie);
381 static int check_valid_map(struct f2fs_sb_info *sbi,
382 unsigned int segno, int offset)
384 struct sit_info *sit_i = SIT_I(sbi);
385 struct seg_entry *sentry;
388 mutex_lock(&sit_i->sentry_lock);
389 sentry = get_seg_entry(sbi, segno);
390 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
391 mutex_unlock(&sit_i->sentry_lock);
396 * This function compares node address got in summary with that in NAT.
397 * On validity, copy that node with cold status, otherwise (invalid node)
400 static void gc_node_segment(struct f2fs_sb_info *sbi,
401 struct f2fs_summary *sum, unsigned int segno, int gc_type)
404 struct f2fs_summary *entry;
410 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
411 nid_t nid = le32_to_cpu(entry->nid);
412 struct page *node_page;
414 /* stop BG_GC if there is not enough free sections. */
415 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
418 if (check_valid_map(sbi, segno, off) == 0)
422 ra_node_page(sbi, nid);
425 node_page = get_node_page(sbi, nid);
426 if (IS_ERR(node_page))
429 /* set page dirty and write it */
430 if (gc_type == FG_GC) {
431 f2fs_wait_on_page_writeback(node_page, NODE);
432 set_page_dirty(node_page);
434 if (!PageWriteback(node_page))
435 set_page_dirty(node_page);
437 f2fs_put_page(node_page, 1);
438 stat_inc_node_blk_count(sbi, 1);
446 if (gc_type == FG_GC) {
447 struct writeback_control wbc = {
448 .sync_mode = WB_SYNC_ALL,
449 .nr_to_write = LONG_MAX,
452 sync_node_pages(sbi, 0, &wbc);
455 * In the case of FG_GC, it'd be better to reclaim this victim
458 if (get_valid_blocks(sbi, segno, 1) != 0)
464 * Calculate start block index indicating the given node offset.
465 * Be careful, caller should give this node offset only indicating direct node
466 * blocks. If any node offsets, which point the other types of node blocks such
467 * as indirect or double indirect node blocks, are given, it must be a caller's
470 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
472 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
480 } else if (node_ofs <= indirect_blks) {
481 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
482 bidx = node_ofs - 2 - dec;
484 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
485 bidx = node_ofs - 5 - dec;
487 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
490 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
491 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
493 struct page *node_page;
495 unsigned int ofs_in_node;
496 block_t source_blkaddr;
498 nid = le32_to_cpu(sum->nid);
499 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
501 node_page = get_node_page(sbi, nid);
502 if (IS_ERR(node_page))
505 get_node_info(sbi, nid, dni);
507 if (sum->version != dni->version) {
508 f2fs_put_page(node_page, 1);
512 *nofs = ofs_of_node(node_page);
513 source_blkaddr = datablock_addr(node_page, ofs_in_node);
514 f2fs_put_page(node_page, 1);
516 if (source_blkaddr != blkaddr)
521 static void move_data_page(struct inode *inode, struct page *page, int gc_type)
523 struct f2fs_io_info fio = {
528 if (gc_type == BG_GC) {
529 if (PageWriteback(page))
531 set_page_dirty(page);
534 struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb);
536 f2fs_wait_on_page_writeback(page, DATA);
538 if (clear_page_dirty_for_io(page) &&
539 S_ISDIR(inode->i_mode)) {
540 dec_page_count(sbi, F2FS_DIRTY_DENTS);
541 inode_dec_dirty_dents(inode);
544 do_write_data_page(page, &fio);
545 clear_cold_data(page);
548 f2fs_put_page(page, 1);
552 * This function tries to get parent node of victim data block, and identifies
553 * data block validity. If the block is valid, copy that with cold status and
554 * modify parent node.
555 * If the parent node is not valid or the data block address is different,
556 * the victim data block is ignored.
558 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
559 struct list_head *ilist, unsigned int segno, int gc_type)
561 struct super_block *sb = sbi->sb;
562 struct f2fs_summary *entry;
567 start_addr = START_BLOCK(sbi, segno);
572 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
573 struct page *data_page;
575 struct node_info dni; /* dnode info for the data */
576 unsigned int ofs_in_node, nofs;
579 /* stop BG_GC if there is not enough free sections. */
580 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
583 if (check_valid_map(sbi, segno, off) == 0)
587 ra_node_page(sbi, le32_to_cpu(entry->nid));
591 /* Get an inode by ino with checking validity */
592 if (check_dnode(sbi, entry, &dni, start_addr + off, &nofs) == 0)
596 ra_node_page(sbi, dni.ino);
600 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
603 inode = f2fs_iget(sb, dni.ino);
607 start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
609 data_page = find_data_page(inode,
610 start_bidx + ofs_in_node, false);
611 if (IS_ERR(data_page))
614 f2fs_put_page(data_page, 0);
615 add_gc_inode(inode, ilist);
617 inode = find_gc_inode(dni.ino, ilist);
619 start_bidx = start_bidx_of_node(nofs,
621 data_page = get_lock_data_page(inode,
622 start_bidx + ofs_in_node);
623 if (IS_ERR(data_page))
625 move_data_page(inode, data_page, gc_type);
626 stat_inc_data_blk_count(sbi, 1);
637 if (gc_type == FG_GC) {
638 f2fs_submit_merged_bio(sbi, DATA, WRITE);
641 * In the case of FG_GC, it'd be better to reclaim this victim
644 if (get_valid_blocks(sbi, segno, 1) != 0) {
651 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
652 int gc_type, int type)
654 struct sit_info *sit_i = SIT_I(sbi);
656 mutex_lock(&sit_i->sentry_lock);
657 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS);
658 mutex_unlock(&sit_i->sentry_lock);
662 static void do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
663 struct list_head *ilist, int gc_type)
665 struct page *sum_page;
666 struct f2fs_summary_block *sum;
667 struct blk_plug plug;
669 /* read segment summary of victim */
670 sum_page = get_sum_page(sbi, segno);
672 blk_start_plug(&plug);
674 sum = page_address(sum_page);
676 switch (GET_SUM_TYPE((&sum->footer))) {
678 gc_node_segment(sbi, sum->entries, segno, gc_type);
681 gc_data_segment(sbi, sum->entries, ilist, segno, gc_type);
684 blk_finish_plug(&plug);
686 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)));
687 stat_inc_call_count(sbi->stat_info);
689 f2fs_put_page(sum_page, 1);
692 int f2fs_gc(struct f2fs_sb_info *sbi)
694 struct list_head ilist;
695 unsigned int segno, i;
700 INIT_LIST_HEAD(&ilist);
702 if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
705 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
707 write_checkpoint(sbi, false);
710 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE))
714 for (i = 0; i < sbi->segs_per_sec; i++)
715 do_garbage_collect(sbi, segno + i, &ilist, gc_type);
717 if (gc_type == FG_GC) {
718 sbi->cur_victim_sec = NULL_SEGNO;
720 WARN_ON(get_valid_blocks(sbi, segno, sbi->segs_per_sec));
723 if (has_not_enough_free_secs(sbi, nfree))
726 if (gc_type == FG_GC)
727 write_checkpoint(sbi, false);
729 mutex_unlock(&sbi->gc_mutex);
731 put_gc_inode(&ilist);
735 void build_gc_manager(struct f2fs_sb_info *sbi)
737 DIRTY_I(sbi)->v_ops = &default_v_ops;
740 int __init create_gc_caches(void)
742 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes",
743 sizeof(struct inode_entry), NULL);
749 void destroy_gc_caches(void)
751 kmem_cache_destroy(winode_slab);