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/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>
26 #include <trace/events/f2fs.h>
28 static struct kmem_cache *winode_slab;
30 static int gc_thread_func(void *data)
32 struct f2fs_sb_info *sbi = data;
33 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
36 wait_ms = GC_THREAD_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 = GC_THREAD_MAX_SLEEP_TIME;
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(wait_ms);
71 mutex_unlock(&sbi->gc_mutex);
75 if (has_enough_invalid_blocks(sbi))
76 wait_ms = decrease_sleep_time(wait_ms);
78 wait_ms = increase_sleep_time(wait_ms);
82 /* if return value is not zero, no victim was selected */
84 wait_ms = GC_THREAD_NOGC_SLEEP_TIME;
85 else if (wait_ms == GC_THREAD_NOGC_SLEEP_TIME)
86 wait_ms = GC_THREAD_MAX_SLEEP_TIME;
88 } while (!kthread_should_stop());
92 int start_gc_thread(struct f2fs_sb_info *sbi)
94 struct f2fs_gc_kthread *gc_th;
95 dev_t dev = sbi->sb->s_bdev->bd_dev;
97 if (!test_opt(sbi, BG_GC))
99 gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
103 sbi->gc_thread = gc_th;
104 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
105 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
106 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
107 if (IS_ERR(gc_th->f2fs_gc_task)) {
109 sbi->gc_thread = NULL;
115 void stop_gc_thread(struct f2fs_sb_info *sbi)
117 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
120 kthread_stop(gc_th->f2fs_gc_task);
122 sbi->gc_thread = NULL;
125 static int select_gc_type(int gc_type)
127 return (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
130 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
131 int type, struct victim_sel_policy *p)
133 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
135 if (p->alloc_mode == SSR) {
136 p->gc_mode = GC_GREEDY;
137 p->dirty_segmap = dirty_i->dirty_segmap[type];
140 p->gc_mode = select_gc_type(gc_type);
141 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
142 p->ofs_unit = sbi->segs_per_sec;
144 p->offset = sbi->last_victim[p->gc_mode];
147 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
148 struct victim_sel_policy *p)
150 /* SSR allocates in a segment unit */
151 if (p->alloc_mode == SSR)
152 return 1 << sbi->log_blocks_per_seg;
153 if (p->gc_mode == GC_GREEDY)
154 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
155 else if (p->gc_mode == GC_CB)
157 else /* No other gc_mode */
161 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
163 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
164 unsigned int hint = 0;
168 * If the gc_type is FG_GC, we can select victim segments
169 * selected by background GC before.
170 * Those segments guarantee they have small valid blocks.
173 secno = find_next_bit(dirty_i->victim_secmap, TOTAL_SECS(sbi), hint++);
174 if (secno < TOTAL_SECS(sbi)) {
175 if (sec_usage_check(sbi, secno))
177 clear_bit(secno, dirty_i->victim_secmap);
178 return secno * sbi->segs_per_sec;
183 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
185 struct sit_info *sit_i = SIT_I(sbi);
186 unsigned int secno = GET_SECNO(sbi, segno);
187 unsigned int start = secno * sbi->segs_per_sec;
188 unsigned long long mtime = 0;
189 unsigned int vblocks;
190 unsigned char age = 0;
194 for (i = 0; i < sbi->segs_per_sec; i++)
195 mtime += get_seg_entry(sbi, start + i)->mtime;
196 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
198 mtime = div_u64(mtime, sbi->segs_per_sec);
199 vblocks = div_u64(vblocks, sbi->segs_per_sec);
201 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
203 /* Handle if the system time is changed by user */
204 if (mtime < sit_i->min_mtime)
205 sit_i->min_mtime = mtime;
206 if (mtime > sit_i->max_mtime)
207 sit_i->max_mtime = mtime;
208 if (sit_i->max_mtime != sit_i->min_mtime)
209 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
210 sit_i->max_mtime - sit_i->min_mtime);
212 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
215 static unsigned int get_gc_cost(struct f2fs_sb_info *sbi, unsigned int segno,
216 struct victim_sel_policy *p)
218 if (p->alloc_mode == SSR)
219 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
221 /* alloc_mode == LFS */
222 if (p->gc_mode == GC_GREEDY)
223 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
225 return get_cb_cost(sbi, segno);
229 * This function is called from two paths.
230 * One is garbage collection and the other is SSR segment selection.
231 * When it is called during GC, it just gets a victim segment
232 * and it does not remove it from dirty seglist.
233 * When it is called from SSR segment selection, it finds a segment
234 * which has minimum valid blocks and removes it from dirty seglist.
236 static int get_victim_by_default(struct f2fs_sb_info *sbi,
237 unsigned int *result, int gc_type, int type, char alloc_mode)
239 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
240 struct victim_sel_policy p;
244 p.alloc_mode = alloc_mode;
245 select_policy(sbi, gc_type, type, &p);
247 p.min_segno = NULL_SEGNO;
248 p.min_cost = get_max_cost(sbi, &p);
250 mutex_lock(&dirty_i->seglist_lock);
252 if (p.alloc_mode == LFS && gc_type == FG_GC) {
253 p.min_segno = check_bg_victims(sbi);
254 if (p.min_segno != NULL_SEGNO)
262 segno = find_next_bit(p.dirty_segmap,
263 TOTAL_SEGS(sbi), p.offset);
264 if (segno >= TOTAL_SEGS(sbi)) {
265 if (sbi->last_victim[p.gc_mode]) {
266 sbi->last_victim[p.gc_mode] = 0;
272 p.offset = ((segno / p.ofs_unit) * p.ofs_unit) + p.ofs_unit;
273 secno = GET_SECNO(sbi, segno);
275 if (sec_usage_check(sbi, secno))
277 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
280 cost = get_gc_cost(sbi, segno, &p);
282 if (p.min_cost > cost) {
287 if (cost == get_max_cost(sbi, &p))
290 if (nsearched++ >= MAX_VICTIM_SEARCH) {
291 sbi->last_victim[p.gc_mode] = segno;
296 if (p.min_segno != NULL_SEGNO) {
297 if (p.alloc_mode == LFS) {
298 secno = GET_SECNO(sbi, p.min_segno);
299 if (gc_type == FG_GC)
300 sbi->cur_victim_sec = secno;
302 set_bit(secno, dirty_i->victim_secmap);
304 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
306 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
308 prefree_segments(sbi), free_segments(sbi));
310 mutex_unlock(&dirty_i->seglist_lock);
312 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
315 static const struct victim_selection default_v_ops = {
316 .get_victim = get_victim_by_default,
319 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist)
321 struct list_head *this;
322 struct inode_entry *ie;
324 list_for_each(this, ilist) {
325 ie = list_entry(this, struct inode_entry, list);
326 if (ie->inode->i_ino == ino)
332 static void add_gc_inode(struct inode *inode, struct list_head *ilist)
334 struct list_head *this;
335 struct inode_entry *new_ie, *ie;
337 list_for_each(this, ilist) {
338 ie = list_entry(this, struct inode_entry, list);
339 if (ie->inode == inode) {
345 new_ie = kmem_cache_alloc(winode_slab, GFP_NOFS);
350 new_ie->inode = inode;
351 list_add_tail(&new_ie->list, ilist);
354 static void put_gc_inode(struct list_head *ilist)
356 struct inode_entry *ie, *next_ie;
357 list_for_each_entry_safe(ie, next_ie, ilist, list) {
360 kmem_cache_free(winode_slab, ie);
364 static int check_valid_map(struct f2fs_sb_info *sbi,
365 unsigned int segno, int offset)
367 struct sit_info *sit_i = SIT_I(sbi);
368 struct seg_entry *sentry;
371 mutex_lock(&sit_i->sentry_lock);
372 sentry = get_seg_entry(sbi, segno);
373 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
374 mutex_unlock(&sit_i->sentry_lock);
379 * This function compares node address got in summary with that in NAT.
380 * On validity, copy that node with cold status, otherwise (invalid node)
383 static void gc_node_segment(struct f2fs_sb_info *sbi,
384 struct f2fs_summary *sum, unsigned int segno, int gc_type)
387 struct f2fs_summary *entry;
392 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
393 nid_t nid = le32_to_cpu(entry->nid);
394 struct page *node_page;
396 /* stop BG_GC if there is not enough free sections. */
397 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
400 if (check_valid_map(sbi, segno, off) == 0)
404 ra_node_page(sbi, nid);
407 node_page = get_node_page(sbi, nid);
408 if (IS_ERR(node_page))
411 /* set page dirty and write it */
412 if (gc_type == FG_GC) {
413 f2fs_submit_bio(sbi, NODE, true);
414 wait_on_page_writeback(node_page);
415 set_page_dirty(node_page);
417 if (!PageWriteback(node_page))
418 set_page_dirty(node_page);
420 f2fs_put_page(node_page, 1);
421 stat_inc_node_blk_count(sbi, 1);
428 if (gc_type == FG_GC) {
429 struct writeback_control wbc = {
430 .sync_mode = WB_SYNC_ALL,
431 .nr_to_write = LONG_MAX,
434 sync_node_pages(sbi, 0, &wbc);
437 * In the case of FG_GC, it'd be better to reclaim this victim
440 if (get_valid_blocks(sbi, segno, 1) != 0)
446 * Calculate start block index indicating the given node offset.
447 * Be careful, caller should give this node offset only indicating direct node
448 * blocks. If any node offsets, which point the other types of node blocks such
449 * as indirect or double indirect node blocks, are given, it must be a caller's
452 block_t start_bidx_of_node(unsigned int node_ofs)
454 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
462 } else if (node_ofs <= indirect_blks) {
463 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
464 bidx = node_ofs - 2 - dec;
466 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
467 bidx = node_ofs - 5 - dec;
469 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE;
472 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
473 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
475 struct page *node_page;
477 unsigned int ofs_in_node;
478 block_t source_blkaddr;
480 nid = le32_to_cpu(sum->nid);
481 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
483 node_page = get_node_page(sbi, nid);
484 if (IS_ERR(node_page))
487 get_node_info(sbi, nid, dni);
489 if (sum->version != dni->version) {
490 f2fs_put_page(node_page, 1);
494 *nofs = ofs_of_node(node_page);
495 source_blkaddr = datablock_addr(node_page, ofs_in_node);
496 f2fs_put_page(node_page, 1);
498 if (source_blkaddr != blkaddr)
503 static void move_data_page(struct inode *inode, struct page *page, int gc_type)
505 if (gc_type == BG_GC) {
506 if (PageWriteback(page))
508 set_page_dirty(page);
511 struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb);
513 if (PageWriteback(page)) {
514 f2fs_submit_bio(sbi, DATA, true);
515 wait_on_page_writeback(page);
518 if (clear_page_dirty_for_io(page) &&
519 S_ISDIR(inode->i_mode)) {
520 dec_page_count(sbi, F2FS_DIRTY_DENTS);
521 inode_dec_dirty_dents(inode);
524 do_write_data_page(page);
525 clear_cold_data(page);
528 f2fs_put_page(page, 1);
532 * This function tries to get parent node of victim data block, and identifies
533 * data block validity. If the block is valid, copy that with cold status and
534 * modify parent node.
535 * If the parent node is not valid or the data block address is different,
536 * the victim data block is ignored.
538 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
539 struct list_head *ilist, unsigned int segno, int gc_type)
541 struct super_block *sb = sbi->sb;
542 struct f2fs_summary *entry;
547 start_addr = START_BLOCK(sbi, segno);
551 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
552 struct page *data_page;
554 struct node_info dni; /* dnode info for the data */
555 unsigned int ofs_in_node, nofs;
558 /* stop BG_GC if there is not enough free sections. */
559 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
562 if (check_valid_map(sbi, segno, off) == 0)
566 ra_node_page(sbi, le32_to_cpu(entry->nid));
570 /* Get an inode by ino with checking validity */
571 if (check_dnode(sbi, entry, &dni, start_addr + off, &nofs) == 0)
575 ra_node_page(sbi, dni.ino);
579 start_bidx = start_bidx_of_node(nofs);
580 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
583 inode = f2fs_iget(sb, dni.ino);
587 data_page = find_data_page(inode,
588 start_bidx + ofs_in_node);
589 if (IS_ERR(data_page))
592 f2fs_put_page(data_page, 0);
593 add_gc_inode(inode, ilist);
595 inode = find_gc_inode(dni.ino, ilist);
597 data_page = get_lock_data_page(inode,
598 start_bidx + ofs_in_node);
599 if (IS_ERR(data_page))
601 move_data_page(inode, data_page, gc_type);
602 stat_inc_data_blk_count(sbi, 1);
612 if (gc_type == FG_GC) {
613 f2fs_submit_bio(sbi, DATA, true);
616 * In the case of FG_GC, it'd be better to reclaim this victim
619 if (get_valid_blocks(sbi, segno, 1) != 0) {
626 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
627 int gc_type, int type)
629 struct sit_info *sit_i = SIT_I(sbi);
631 mutex_lock(&sit_i->sentry_lock);
632 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS);
633 mutex_unlock(&sit_i->sentry_lock);
637 static void do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
638 struct list_head *ilist, int gc_type)
640 struct page *sum_page;
641 struct f2fs_summary_block *sum;
643 /* read segment summary of victim */
644 sum_page = get_sum_page(sbi, segno);
645 if (IS_ERR(sum_page))
648 sum = page_address(sum_page);
650 switch (GET_SUM_TYPE((&sum->footer))) {
652 gc_node_segment(sbi, sum->entries, segno, gc_type);
655 gc_data_segment(sbi, sum->entries, ilist, segno, gc_type);
658 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)));
659 stat_inc_call_count(sbi->stat_info);
661 f2fs_put_page(sum_page, 1);
664 int f2fs_gc(struct f2fs_sb_info *sbi)
666 struct list_head ilist;
667 unsigned int segno, i;
672 INIT_LIST_HEAD(&ilist);
674 if (!(sbi->sb->s_flags & MS_ACTIVE))
677 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
679 write_checkpoint(sbi, false);
682 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE))
686 for (i = 0; i < sbi->segs_per_sec; i++)
687 do_garbage_collect(sbi, segno + i, &ilist, gc_type);
689 if (gc_type == FG_GC) {
690 sbi->cur_victim_sec = NULL_SEGNO;
692 WARN_ON(get_valid_blocks(sbi, segno, sbi->segs_per_sec));
695 if (has_not_enough_free_secs(sbi, nfree))
698 if (gc_type == FG_GC)
699 write_checkpoint(sbi, false);
701 mutex_unlock(&sbi->gc_mutex);
703 put_gc_inode(&ilist);
707 void build_gc_manager(struct f2fs_sb_info *sbi)
709 DIRTY_I(sbi)->v_ops = &default_v_ops;
712 int __init create_gc_caches(void)
714 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes",
715 sizeof(struct inode_entry), NULL);
721 void destroy_gc_caches(void)
723 kmem_cache_destroy(winode_slab);