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>
27 static struct kmem_cache *winode_slab;
29 static int gc_thread_func(void *data)
31 struct f2fs_sb_info *sbi = data;
32 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
35 wait_ms = GC_THREAD_MIN_SLEEP_TIME;
41 wait_event_interruptible_timeout(*wq,
42 kthread_should_stop(),
43 msecs_to_jiffies(wait_ms));
44 if (kthread_should_stop())
47 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
48 wait_ms = GC_THREAD_MAX_SLEEP_TIME;
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.
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.
65 if (!mutex_trylock(&sbi->gc_mutex))
69 wait_ms = increase_sleep_time(wait_ms);
70 mutex_unlock(&sbi->gc_mutex);
74 if (has_enough_invalid_blocks(sbi))
75 wait_ms = decrease_sleep_time(wait_ms);
77 wait_ms = increase_sleep_time(wait_ms);
81 if (f2fs_gc(sbi) == GC_NONE)
82 wait_ms = GC_THREAD_NOGC_SLEEP_TIME;
83 else if (wait_ms == GC_THREAD_NOGC_SLEEP_TIME)
84 wait_ms = GC_THREAD_MAX_SLEEP_TIME;
86 } while (!kthread_should_stop());
90 int start_gc_thread(struct f2fs_sb_info *sbi)
92 struct f2fs_gc_kthread *gc_th;
93 dev_t dev = sbi->sb->s_bdev->bd_dev;
95 if (!test_opt(sbi, BG_GC))
97 gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
101 sbi->gc_thread = gc_th;
102 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
103 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
104 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
105 if (IS_ERR(gc_th->f2fs_gc_task)) {
107 sbi->gc_thread = NULL;
113 void stop_gc_thread(struct f2fs_sb_info *sbi)
115 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
118 kthread_stop(gc_th->f2fs_gc_task);
120 sbi->gc_thread = NULL;
123 static int select_gc_type(int gc_type)
125 return (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
128 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
129 int type, struct victim_sel_policy *p)
131 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
134 p->gc_mode = GC_GREEDY;
135 p->dirty_segmap = dirty_i->dirty_segmap[type];
138 p->gc_mode = select_gc_type(gc_type);
139 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
140 p->ofs_unit = sbi->segs_per_sec;
142 p->offset = sbi->last_victim[p->gc_mode];
145 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
146 struct victim_sel_policy *p)
148 if (p->gc_mode == GC_GREEDY)
149 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
150 else if (p->gc_mode == GC_CB)
152 else /* No other gc_mode */
156 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
158 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
162 * If the gc_type is FG_GC, we can select victim segments
163 * selected by background GC before.
164 * Those segments guarantee they have small valid blocks.
166 segno = find_next_bit(dirty_i->victim_segmap[BG_GC],
168 if (segno < TOTAL_SEGS(sbi)) {
169 clear_bit(segno, dirty_i->victim_segmap[BG_GC]);
175 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
177 struct sit_info *sit_i = SIT_I(sbi);
178 unsigned int secno = GET_SECNO(sbi, segno);
179 unsigned int start = secno * sbi->segs_per_sec;
180 unsigned long long mtime = 0;
181 unsigned int vblocks;
182 unsigned char age = 0;
186 for (i = 0; i < sbi->segs_per_sec; i++)
187 mtime += get_seg_entry(sbi, start + i)->mtime;
188 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
190 mtime = div_u64(mtime, sbi->segs_per_sec);
191 vblocks = div_u64(vblocks, sbi->segs_per_sec);
193 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
195 /* Handle if the system time is changed by user */
196 if (mtime < sit_i->min_mtime)
197 sit_i->min_mtime = mtime;
198 if (mtime > sit_i->max_mtime)
199 sit_i->max_mtime = mtime;
200 if (sit_i->max_mtime != sit_i->min_mtime)
201 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
202 sit_i->max_mtime - sit_i->min_mtime);
204 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
207 static unsigned int get_gc_cost(struct f2fs_sb_info *sbi, unsigned int segno,
208 struct victim_sel_policy *p)
210 if (p->alloc_mode == SSR)
211 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
213 /* alloc_mode == LFS */
214 if (p->gc_mode == GC_GREEDY)
215 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
217 return get_cb_cost(sbi, segno);
221 * This function is called from two pathes.
222 * One is garbage collection and the other is SSR segment selection.
223 * When it is called during GC, it just gets a victim segment
224 * and it does not remove it from dirty seglist.
225 * When it is called from SSR segment selection, it finds a segment
226 * which has minimum valid blocks and removes it from dirty seglist.
228 static int get_victim_by_default(struct f2fs_sb_info *sbi,
229 unsigned int *result, int gc_type, int type, char alloc_mode)
231 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
232 struct victim_sel_policy p;
236 p.alloc_mode = alloc_mode;
237 select_policy(sbi, gc_type, type, &p);
239 p.min_segno = NULL_SEGNO;
240 p.min_cost = get_max_cost(sbi, &p);
242 mutex_lock(&dirty_i->seglist_lock);
244 if (p.alloc_mode == LFS && gc_type == FG_GC) {
245 p.min_segno = check_bg_victims(sbi);
246 if (p.min_segno != NULL_SEGNO)
253 segno = find_next_bit(p.dirty_segmap,
254 TOTAL_SEGS(sbi), p.offset);
255 if (segno >= TOTAL_SEGS(sbi)) {
256 if (sbi->last_victim[p.gc_mode]) {
257 sbi->last_victim[p.gc_mode] = 0;
263 p.offset = ((segno / p.ofs_unit) * p.ofs_unit) + p.ofs_unit;
265 if (test_bit(segno, dirty_i->victim_segmap[FG_GC]))
267 if (gc_type == BG_GC &&
268 test_bit(segno, dirty_i->victim_segmap[BG_GC]))
270 if (IS_CURSEC(sbi, GET_SECNO(sbi, segno)))
273 cost = get_gc_cost(sbi, segno, &p);
275 if (p.min_cost > cost) {
280 if (cost == get_max_cost(sbi, &p))
283 if (nsearched++ >= MAX_VICTIM_SEARCH) {
284 sbi->last_victim[p.gc_mode] = segno;
289 if (p.min_segno != NULL_SEGNO) {
290 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
291 if (p.alloc_mode == LFS) {
293 for (i = 0; i < p.ofs_unit; i++)
295 dirty_i->victim_segmap[gc_type]);
298 mutex_unlock(&dirty_i->seglist_lock);
300 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
303 static const struct victim_selection default_v_ops = {
304 .get_victim = get_victim_by_default,
307 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist)
309 struct list_head *this;
310 struct inode_entry *ie;
312 list_for_each(this, ilist) {
313 ie = list_entry(this, struct inode_entry, list);
314 if (ie->inode->i_ino == ino)
320 static void add_gc_inode(struct inode *inode, struct list_head *ilist)
322 struct list_head *this;
323 struct inode_entry *new_ie, *ie;
325 list_for_each(this, ilist) {
326 ie = list_entry(this, struct inode_entry, list);
327 if (ie->inode == inode) {
333 new_ie = kmem_cache_alloc(winode_slab, GFP_NOFS);
338 new_ie->inode = inode;
339 list_add_tail(&new_ie->list, ilist);
342 static void put_gc_inode(struct list_head *ilist)
344 struct inode_entry *ie, *next_ie;
345 list_for_each_entry_safe(ie, next_ie, ilist, list) {
348 kmem_cache_free(winode_slab, ie);
352 static int check_valid_map(struct f2fs_sb_info *sbi,
353 unsigned int segno, int offset)
355 struct sit_info *sit_i = SIT_I(sbi);
356 struct seg_entry *sentry;
359 mutex_lock(&sit_i->sentry_lock);
360 sentry = get_seg_entry(sbi, segno);
361 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
362 mutex_unlock(&sit_i->sentry_lock);
363 return ret ? GC_OK : GC_NEXT;
367 * This function compares node address got in summary with that in NAT.
368 * On validity, copy that node with cold status, otherwise (invalid node)
371 static int gc_node_segment(struct f2fs_sb_info *sbi,
372 struct f2fs_summary *sum, unsigned int segno, int gc_type)
375 struct f2fs_summary *entry;
380 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
381 nid_t nid = le32_to_cpu(entry->nid);
382 struct page *node_page;
386 * It makes sure that free segments are able to write
387 * all the dirty node pages before CP after this CP.
388 * So let's check the space of dirty node pages.
390 if (should_do_checkpoint(sbi)) {
391 mutex_lock(&sbi->cp_mutex);
392 block_operations(sbi);
396 err = check_valid_map(sbi, segno, off);
401 ra_node_page(sbi, nid);
404 node_page = get_node_page(sbi, nid);
405 if (IS_ERR(node_page))
408 /* set page dirty and write it */
409 if (!PageWriteback(node_page))
410 set_page_dirty(node_page);
411 f2fs_put_page(node_page, 1);
412 stat_inc_node_blk_count(sbi, 1);
419 if (gc_type == FG_GC) {
420 struct writeback_control wbc = {
421 .sync_mode = WB_SYNC_ALL,
422 .nr_to_write = LONG_MAX,
425 sync_node_pages(sbi, 0, &wbc);
431 * Calculate start block index indicating the given node offset.
432 * Be careful, caller should give this node offset only indicating direct node
433 * blocks. If any node offsets, which point the other types of node blocks such
434 * as indirect or double indirect node blocks, are given, it must be a caller's
437 block_t start_bidx_of_node(unsigned int node_ofs)
439 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
447 } else if (node_ofs <= indirect_blks) {
448 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
449 bidx = node_ofs - 2 - dec;
451 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
452 bidx = node_ofs - 5 - dec;
454 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE;
457 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
458 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
460 struct page *node_page;
462 unsigned int ofs_in_node;
463 block_t source_blkaddr;
465 nid = le32_to_cpu(sum->nid);
466 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
468 node_page = get_node_page(sbi, nid);
469 if (IS_ERR(node_page))
472 get_node_info(sbi, nid, dni);
474 if (sum->version != dni->version) {
475 f2fs_put_page(node_page, 1);
479 *nofs = ofs_of_node(node_page);
480 source_blkaddr = datablock_addr(node_page, ofs_in_node);
481 f2fs_put_page(node_page, 1);
483 if (source_blkaddr != blkaddr)
488 static void move_data_page(struct inode *inode, struct page *page, int gc_type)
490 if (page->mapping != inode->i_mapping)
493 if (inode != page->mapping->host)
496 if (PageWriteback(page))
499 if (gc_type == BG_GC) {
500 set_page_dirty(page);
503 struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb);
504 mutex_lock_op(sbi, DATA_WRITE);
505 if (clear_page_dirty_for_io(page) &&
506 S_ISDIR(inode->i_mode)) {
507 dec_page_count(sbi, F2FS_DIRTY_DENTS);
508 inode_dec_dirty_dents(inode);
511 do_write_data_page(page);
512 mutex_unlock_op(sbi, DATA_WRITE);
513 clear_cold_data(page);
516 f2fs_put_page(page, 1);
520 * This function tries to get parent node of victim data block, and identifies
521 * data block validity. If the block is valid, copy that with cold status and
522 * modify parent node.
523 * If the parent node is not valid or the data block address is different,
524 * the victim data block is ignored.
526 static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
527 struct list_head *ilist, unsigned int segno, int gc_type)
529 struct super_block *sb = sbi->sb;
530 struct f2fs_summary *entry;
535 start_addr = START_BLOCK(sbi, segno);
539 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
540 struct page *data_page;
542 struct node_info dni; /* dnode info for the data */
543 unsigned int ofs_in_node, nofs;
547 * It makes sure that free segments are able to write
548 * all the dirty node pages before CP after this CP.
549 * So let's check the space of dirty node pages.
551 if (should_do_checkpoint(sbi)) {
552 mutex_lock(&sbi->cp_mutex);
553 block_operations(sbi);
558 err = check_valid_map(sbi, segno, off);
563 ra_node_page(sbi, le32_to_cpu(entry->nid));
567 /* Get an inode by ino with checking validity */
568 err = check_dnode(sbi, entry, &dni, start_addr + off, &nofs);
573 ra_node_page(sbi, dni.ino);
577 start_bidx = start_bidx_of_node(nofs);
578 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
581 inode = f2fs_iget(sb, dni.ino);
585 data_page = find_data_page(inode,
586 start_bidx + ofs_in_node);
587 if (IS_ERR(data_page))
590 f2fs_put_page(data_page, 0);
591 add_gc_inode(inode, ilist);
593 inode = find_gc_inode(dni.ino, ilist);
595 data_page = get_lock_data_page(inode,
596 start_bidx + ofs_in_node);
597 if (IS_ERR(data_page))
599 move_data_page(inode, data_page, gc_type);
600 stat_inc_data_blk_count(sbi, 1);
611 if (gc_type == FG_GC)
612 f2fs_submit_bio(sbi, DATA, true);
616 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
617 int gc_type, int type)
619 struct sit_info *sit_i = SIT_I(sbi);
621 mutex_lock(&sit_i->sentry_lock);
622 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS);
623 mutex_unlock(&sit_i->sentry_lock);
627 static int do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
628 struct list_head *ilist, int gc_type)
630 struct page *sum_page;
631 struct f2fs_summary_block *sum;
634 /* read segment summary of victim */
635 sum_page = get_sum_page(sbi, segno);
636 if (IS_ERR(sum_page))
640 * CP needs to lock sum_page. In this time, we don't need
641 * to lock this page, because this summary page is not gone anywhere.
642 * Also, this page is not gonna be updated before GC is done.
644 unlock_page(sum_page);
645 sum = page_address(sum_page);
647 switch (GET_SUM_TYPE((&sum->footer))) {
649 ret = gc_node_segment(sbi, sum->entries, segno, gc_type);
652 ret = gc_data_segment(sbi, sum->entries, ilist, segno, gc_type);
655 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)));
656 stat_inc_call_count(sbi->stat_info);
658 f2fs_put_page(sum_page, 0);
662 int f2fs_gc(struct f2fs_sb_info *sbi)
664 struct list_head ilist;
665 unsigned int segno, i;
667 int gc_status = GC_NONE;
669 INIT_LIST_HEAD(&ilist);
671 if (!(sbi->sb->s_flags & MS_ACTIVE))
674 if (gc_type == BG_GC && has_not_enough_free_secs(sbi))
677 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE))
680 for (i = 0; i < sbi->segs_per_sec; i++) {
682 * do_garbage_collect will give us three gc_status:
683 * GC_ERROR, GC_DONE, and GC_BLOCKED.
684 * If GC is finished uncleanly, we have to return
685 * the victim to dirty segment list.
687 gc_status = do_garbage_collect(sbi, segno + i, &ilist, gc_type);
688 if (gc_status != GC_DONE)
691 if (has_not_enough_free_secs(sbi)) {
692 write_checkpoint(sbi, (gc_status == GC_BLOCKED), false);
693 if (has_not_enough_free_secs(sbi))
697 mutex_unlock(&sbi->gc_mutex);
699 put_gc_inode(&ilist);
703 void build_gc_manager(struct f2fs_sb_info *sbi)
705 DIRTY_I(sbi)->v_ops = &default_v_ops;
708 int __init create_gc_caches(void)
710 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes",
711 sizeof(struct inode_entry), NULL);
717 void destroy_gc_caches(void)
719 kmem_cache_destroy(winode_slab);