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;
127 void stop_gc_thread(struct f2fs_sb_info *sbi)
129 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
132 kthread_stop(gc_th->f2fs_gc_task);
134 sbi->gc_thread = NULL;
137 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
139 int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
141 if (gc_th && gc_th->gc_idle) {
142 if (gc_th->gc_idle == 1)
144 else if (gc_th->gc_idle == 2)
150 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
151 int type, struct victim_sel_policy *p)
153 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
155 if (p->alloc_mode == SSR) {
156 p->gc_mode = GC_GREEDY;
157 p->dirty_segmap = dirty_i->dirty_segmap[type];
158 p->max_search = dirty_i->nr_dirty[type];
161 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
162 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
163 p->max_search = dirty_i->nr_dirty[DIRTY];
164 p->ofs_unit = sbi->segs_per_sec;
167 if (p->max_search > MAX_VICTIM_SEARCH)
168 p->max_search = MAX_VICTIM_SEARCH;
170 p->offset = sbi->last_victim[p->gc_mode];
173 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
174 struct victim_sel_policy *p)
176 /* SSR allocates in a segment unit */
177 if (p->alloc_mode == SSR)
178 return 1 << sbi->log_blocks_per_seg;
179 if (p->gc_mode == GC_GREEDY)
180 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
181 else if (p->gc_mode == GC_CB)
183 else /* No other gc_mode */
187 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
189 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
190 unsigned int hint = 0;
194 * If the gc_type is FG_GC, we can select victim segments
195 * selected by background GC before.
196 * Those segments guarantee they have small valid blocks.
199 secno = find_next_bit(dirty_i->victim_secmap, TOTAL_SECS(sbi), hint++);
200 if (secno < TOTAL_SECS(sbi)) {
201 if (sec_usage_check(sbi, secno))
203 clear_bit(secno, dirty_i->victim_secmap);
204 return secno * sbi->segs_per_sec;
209 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
211 struct sit_info *sit_i = SIT_I(sbi);
212 unsigned int secno = GET_SECNO(sbi, segno);
213 unsigned int start = secno * sbi->segs_per_sec;
214 unsigned long long mtime = 0;
215 unsigned int vblocks;
216 unsigned char age = 0;
220 for (i = 0; i < sbi->segs_per_sec; i++)
221 mtime += get_seg_entry(sbi, start + i)->mtime;
222 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
224 mtime = div_u64(mtime, sbi->segs_per_sec);
225 vblocks = div_u64(vblocks, sbi->segs_per_sec);
227 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
229 /* Handle if the system time is changed by user */
230 if (mtime < sit_i->min_mtime)
231 sit_i->min_mtime = mtime;
232 if (mtime > sit_i->max_mtime)
233 sit_i->max_mtime = mtime;
234 if (sit_i->max_mtime != sit_i->min_mtime)
235 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
236 sit_i->max_mtime - sit_i->min_mtime);
238 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
241 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
242 unsigned int segno, struct victim_sel_policy *p)
244 if (p->alloc_mode == SSR)
245 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
247 /* alloc_mode == LFS */
248 if (p->gc_mode == GC_GREEDY)
249 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
251 return get_cb_cost(sbi, segno);
255 * This function is called from two paths.
256 * One is garbage collection and the other is SSR segment selection.
257 * When it is called during GC, it just gets a victim segment
258 * and it does not remove it from dirty seglist.
259 * When it is called from SSR segment selection, it finds a segment
260 * which has minimum valid blocks and removes it from dirty seglist.
262 static int get_victim_by_default(struct f2fs_sb_info *sbi,
263 unsigned int *result, int gc_type, int type, char alloc_mode)
265 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
266 struct victim_sel_policy p;
267 unsigned int secno, max_cost;
270 p.alloc_mode = alloc_mode;
271 select_policy(sbi, gc_type, type, &p);
273 p.min_segno = NULL_SEGNO;
274 p.min_cost = max_cost = get_max_cost(sbi, &p);
276 mutex_lock(&dirty_i->seglist_lock);
278 if (p.alloc_mode == LFS && gc_type == FG_GC) {
279 p.min_segno = check_bg_victims(sbi);
280 if (p.min_segno != NULL_SEGNO)
288 segno = find_next_bit(p.dirty_segmap,
289 TOTAL_SEGS(sbi), p.offset);
290 if (segno >= TOTAL_SEGS(sbi)) {
291 if (sbi->last_victim[p.gc_mode]) {
292 sbi->last_victim[p.gc_mode] = 0;
299 p.offset = segno + p.ofs_unit;
301 p.offset -= segno % p.ofs_unit;
303 secno = GET_SECNO(sbi, segno);
305 if (sec_usage_check(sbi, secno))
307 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
310 cost = get_gc_cost(sbi, segno, &p);
312 if (p.min_cost > cost) {
315 } else if (unlikely(cost == max_cost)) {
319 if (nsearched++ >= p.max_search) {
320 sbi->last_victim[p.gc_mode] = segno;
324 if (p.min_segno != NULL_SEGNO) {
326 if (p.alloc_mode == LFS) {
327 secno = GET_SECNO(sbi, p.min_segno);
328 if (gc_type == FG_GC)
329 sbi->cur_victim_sec = secno;
331 set_bit(secno, dirty_i->victim_secmap);
333 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
335 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
337 prefree_segments(sbi), free_segments(sbi));
339 mutex_unlock(&dirty_i->seglist_lock);
341 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
344 static const struct victim_selection default_v_ops = {
345 .get_victim = get_victim_by_default,
348 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist)
350 struct inode_entry *ie;
352 list_for_each_entry(ie, ilist, list)
353 if (ie->inode->i_ino == ino)
358 static void add_gc_inode(struct inode *inode, struct list_head *ilist)
360 struct inode_entry *new_ie;
362 if (inode == find_gc_inode(inode->i_ino, ilist)) {
367 new_ie = f2fs_kmem_cache_alloc(winode_slab, GFP_NOFS);
368 new_ie->inode = inode;
369 list_add_tail(&new_ie->list, ilist);
372 static void put_gc_inode(struct list_head *ilist)
374 struct inode_entry *ie, *next_ie;
375 list_for_each_entry_safe(ie, next_ie, ilist, list) {
378 kmem_cache_free(winode_slab, ie);
382 static int check_valid_map(struct f2fs_sb_info *sbi,
383 unsigned int segno, int offset)
385 struct sit_info *sit_i = SIT_I(sbi);
386 struct seg_entry *sentry;
389 mutex_lock(&sit_i->sentry_lock);
390 sentry = get_seg_entry(sbi, segno);
391 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
392 mutex_unlock(&sit_i->sentry_lock);
397 * This function compares node address got in summary with that in NAT.
398 * On validity, copy that node with cold status, otherwise (invalid node)
401 static void gc_node_segment(struct f2fs_sb_info *sbi,
402 struct f2fs_summary *sum, unsigned int segno, int gc_type)
405 struct f2fs_summary *entry;
411 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
412 nid_t nid = le32_to_cpu(entry->nid);
413 struct page *node_page;
415 /* stop BG_GC if there is not enough free sections. */
416 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
419 if (check_valid_map(sbi, segno, off) == 0)
423 ra_node_page(sbi, nid);
426 node_page = get_node_page(sbi, nid);
427 if (IS_ERR(node_page))
430 /* set page dirty and write it */
431 if (gc_type == FG_GC) {
432 f2fs_wait_on_page_writeback(node_page, NODE, true);
433 set_page_dirty(node_page);
435 if (!PageWriteback(node_page))
436 set_page_dirty(node_page);
438 f2fs_put_page(node_page, 1);
439 stat_inc_node_blk_count(sbi, 1);
447 if (gc_type == FG_GC) {
448 struct writeback_control wbc = {
449 .sync_mode = WB_SYNC_ALL,
450 .nr_to_write = LONG_MAX,
453 sync_node_pages(sbi, 0, &wbc);
456 * In the case of FG_GC, it'd be better to reclaim this victim
459 if (get_valid_blocks(sbi, segno, 1) != 0)
465 * Calculate start block index indicating the given node offset.
466 * Be careful, caller should give this node offset only indicating direct node
467 * blocks. If any node offsets, which point the other types of node blocks such
468 * as indirect or double indirect node blocks, are given, it must be a caller's
471 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
473 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
481 } else if (node_ofs <= indirect_blks) {
482 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
483 bidx = node_ofs - 2 - dec;
485 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
486 bidx = node_ofs - 5 - dec;
488 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
491 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
492 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
494 struct page *node_page;
496 unsigned int ofs_in_node;
497 block_t source_blkaddr;
499 nid = le32_to_cpu(sum->nid);
500 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
502 node_page = get_node_page(sbi, nid);
503 if (IS_ERR(node_page))
506 get_node_info(sbi, nid, dni);
508 if (sum->version != dni->version) {
509 f2fs_put_page(node_page, 1);
513 *nofs = ofs_of_node(node_page);
514 source_blkaddr = datablock_addr(node_page, ofs_in_node);
515 f2fs_put_page(node_page, 1);
517 if (source_blkaddr != blkaddr)
522 static void move_data_page(struct inode *inode, struct page *page, int gc_type)
524 if (gc_type == BG_GC) {
525 if (PageWriteback(page))
527 set_page_dirty(page);
530 struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb);
532 f2fs_wait_on_page_writeback(page, DATA, true);
534 if (clear_page_dirty_for_io(page) &&
535 S_ISDIR(inode->i_mode)) {
536 dec_page_count(sbi, F2FS_DIRTY_DENTS);
537 inode_dec_dirty_dents(inode);
540 do_write_data_page(page);
541 clear_cold_data(page);
544 f2fs_put_page(page, 1);
548 * This function tries to get parent node of victim data block, and identifies
549 * data block validity. If the block is valid, copy that with cold status and
550 * modify parent node.
551 * If the parent node is not valid or the data block address is different,
552 * the victim data block is ignored.
554 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
555 struct list_head *ilist, unsigned int segno, int gc_type)
557 struct super_block *sb = sbi->sb;
558 struct f2fs_summary *entry;
563 start_addr = START_BLOCK(sbi, segno);
568 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
569 struct page *data_page;
571 struct node_info dni; /* dnode info for the data */
572 unsigned int ofs_in_node, nofs;
575 /* stop BG_GC if there is not enough free sections. */
576 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
579 if (check_valid_map(sbi, segno, off) == 0)
583 ra_node_page(sbi, le32_to_cpu(entry->nid));
587 /* Get an inode by ino with checking validity */
588 if (check_dnode(sbi, entry, &dni, start_addr + off, &nofs) == 0)
592 ra_node_page(sbi, dni.ino);
596 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
599 inode = f2fs_iget(sb, dni.ino);
603 start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
605 data_page = find_data_page(inode,
606 start_bidx + ofs_in_node, false);
607 if (IS_ERR(data_page))
610 f2fs_put_page(data_page, 0);
611 add_gc_inode(inode, ilist);
613 inode = find_gc_inode(dni.ino, ilist);
615 start_bidx = start_bidx_of_node(nofs,
617 data_page = get_lock_data_page(inode,
618 start_bidx + ofs_in_node);
619 if (IS_ERR(data_page))
621 move_data_page(inode, data_page, gc_type);
622 stat_inc_data_blk_count(sbi, 1);
633 if (gc_type == FG_GC) {
634 f2fs_submit_bio(sbi, DATA, true);
637 * In the case of FG_GC, it'd be better to reclaim this victim
640 if (get_valid_blocks(sbi, segno, 1) != 0) {
647 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
648 int gc_type, int type)
650 struct sit_info *sit_i = SIT_I(sbi);
652 mutex_lock(&sit_i->sentry_lock);
653 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS);
654 mutex_unlock(&sit_i->sentry_lock);
658 static void do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
659 struct list_head *ilist, int gc_type)
661 struct page *sum_page;
662 struct f2fs_summary_block *sum;
663 struct blk_plug plug;
665 /* read segment summary of victim */
666 sum_page = get_sum_page(sbi, segno);
667 if (IS_ERR(sum_page))
670 blk_start_plug(&plug);
672 sum = page_address(sum_page);
674 switch (GET_SUM_TYPE((&sum->footer))) {
676 gc_node_segment(sbi, sum->entries, segno, gc_type);
679 gc_data_segment(sbi, sum->entries, ilist, segno, gc_type);
682 blk_finish_plug(&plug);
684 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)));
685 stat_inc_call_count(sbi->stat_info);
687 f2fs_put_page(sum_page, 1);
690 int f2fs_gc(struct f2fs_sb_info *sbi)
692 struct list_head ilist;
693 unsigned int segno, i;
698 INIT_LIST_HEAD(&ilist);
700 if (!(sbi->sb->s_flags & MS_ACTIVE))
703 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
705 write_checkpoint(sbi, false);
708 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE))
712 for (i = 0; i < sbi->segs_per_sec; i++)
713 do_garbage_collect(sbi, segno + i, &ilist, gc_type);
715 if (gc_type == FG_GC) {
716 sbi->cur_victim_sec = NULL_SEGNO;
718 WARN_ON(get_valid_blocks(sbi, segno, sbi->segs_per_sec));
721 if (has_not_enough_free_secs(sbi, nfree))
724 if (gc_type == FG_GC)
725 write_checkpoint(sbi, false);
727 mutex_unlock(&sbi->gc_mutex);
729 put_gc_inode(&ilist);
733 void build_gc_manager(struct f2fs_sb_info *sbi)
735 DIRTY_I(sbi)->v_ops = &default_v_ops;
738 int __init create_gc_caches(void)
740 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes",
741 sizeof(struct inode_entry), NULL);
747 void destroy_gc_caches(void)
749 kmem_cache_destroy(winode_slab);