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 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);
192 * If the gc_type is FG_GC, we can select victim segments
193 * selected by background GC before.
194 * Those segments guarantee they have small valid blocks.
196 for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
197 if (sec_usage_check(sbi, secno))
199 clear_bit(secno, dirty_i->victim_secmap);
200 return secno * sbi->segs_per_sec;
205 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
207 struct sit_info *sit_i = SIT_I(sbi);
208 unsigned int secno = GET_SECNO(sbi, segno);
209 unsigned int start = secno * sbi->segs_per_sec;
210 unsigned long long mtime = 0;
211 unsigned int vblocks;
212 unsigned char age = 0;
216 for (i = 0; i < sbi->segs_per_sec; i++)
217 mtime += get_seg_entry(sbi, start + i)->mtime;
218 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
220 mtime = div_u64(mtime, sbi->segs_per_sec);
221 vblocks = div_u64(vblocks, sbi->segs_per_sec);
223 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
225 /* Handle if the system time has changed by the user */
226 if (mtime < sit_i->min_mtime)
227 sit_i->min_mtime = mtime;
228 if (mtime > sit_i->max_mtime)
229 sit_i->max_mtime = mtime;
230 if (sit_i->max_mtime != sit_i->min_mtime)
231 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
232 sit_i->max_mtime - sit_i->min_mtime);
234 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
237 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
238 unsigned int segno, struct victim_sel_policy *p)
240 if (p->alloc_mode == SSR)
241 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
243 /* alloc_mode == LFS */
244 if (p->gc_mode == GC_GREEDY)
245 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
247 return get_cb_cost(sbi, segno);
251 * This function is called from two paths.
252 * One is garbage collection and the other is SSR segment selection.
253 * When it is called during GC, it just gets a victim segment
254 * and it does not remove it from dirty seglist.
255 * When it is called from SSR segment selection, it finds a segment
256 * which has minimum valid blocks and removes it from dirty seglist.
258 static int get_victim_by_default(struct f2fs_sb_info *sbi,
259 unsigned int *result, int gc_type, int type, char alloc_mode)
261 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
262 struct victim_sel_policy p;
263 unsigned int secno, max_cost;
266 mutex_lock(&dirty_i->seglist_lock);
268 p.alloc_mode = alloc_mode;
269 select_policy(sbi, gc_type, type, &p);
271 p.min_segno = NULL_SEGNO;
272 p.min_cost = max_cost = get_max_cost(sbi, &p);
274 if (p.alloc_mode == LFS && gc_type == FG_GC) {
275 p.min_segno = check_bg_victims(sbi);
276 if (p.min_segno != NULL_SEGNO)
284 segno = find_next_bit(p.dirty_segmap, MAIN_SEGS(sbi), p.offset);
285 if (segno >= MAIN_SEGS(sbi)) {
286 if (sbi->last_victim[p.gc_mode]) {
287 sbi->last_victim[p.gc_mode] = 0;
294 p.offset = segno + p.ofs_unit;
296 p.offset -= segno % p.ofs_unit;
298 secno = GET_SECNO(sbi, segno);
300 if (sec_usage_check(sbi, secno))
302 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
305 cost = get_gc_cost(sbi, segno, &p);
307 if (p.min_cost > cost) {
310 } else if (unlikely(cost == max_cost)) {
314 if (nsearched++ >= p.max_search) {
315 sbi->last_victim[p.gc_mode] = segno;
319 if (p.min_segno != NULL_SEGNO) {
321 if (p.alloc_mode == LFS) {
322 secno = GET_SECNO(sbi, p.min_segno);
323 if (gc_type == FG_GC)
324 sbi->cur_victim_sec = secno;
326 set_bit(secno, dirty_i->victim_secmap);
328 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
330 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
332 prefree_segments(sbi), free_segments(sbi));
334 mutex_unlock(&dirty_i->seglist_lock);
336 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
339 static const struct victim_selection default_v_ops = {
340 .get_victim = get_victim_by_default,
343 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist)
345 struct inode_entry *ie;
347 list_for_each_entry(ie, ilist, list)
348 if (ie->inode->i_ino == ino)
353 static void add_gc_inode(struct inode *inode, struct list_head *ilist)
355 struct inode_entry *new_ie;
357 if (inode == find_gc_inode(inode->i_ino, ilist)) {
362 new_ie = f2fs_kmem_cache_alloc(winode_slab, GFP_NOFS);
363 new_ie->inode = inode;
364 list_add_tail(&new_ie->list, ilist);
367 static void put_gc_inode(struct list_head *ilist)
369 struct inode_entry *ie, *next_ie;
370 list_for_each_entry_safe(ie, next_ie, ilist, list) {
373 kmem_cache_free(winode_slab, ie);
377 static int check_valid_map(struct f2fs_sb_info *sbi,
378 unsigned int segno, int offset)
380 struct sit_info *sit_i = SIT_I(sbi);
381 struct seg_entry *sentry;
384 mutex_lock(&sit_i->sentry_lock);
385 sentry = get_seg_entry(sbi, segno);
386 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
387 mutex_unlock(&sit_i->sentry_lock);
392 * This function compares node address got in summary with that in NAT.
393 * On validity, copy that node with cold status, otherwise (invalid node)
396 static void gc_node_segment(struct f2fs_sb_info *sbi,
397 struct f2fs_summary *sum, unsigned int segno, int gc_type)
400 struct f2fs_summary *entry;
406 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
407 nid_t nid = le32_to_cpu(entry->nid);
408 struct page *node_page;
410 /* stop BG_GC if there is not enough free sections. */
411 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
414 if (check_valid_map(sbi, segno, off) == 0)
418 ra_node_page(sbi, nid);
421 node_page = get_node_page(sbi, nid);
422 if (IS_ERR(node_page))
425 /* block may become invalid during get_node_page */
426 if (check_valid_map(sbi, segno, off) == 0) {
427 f2fs_put_page(node_page, 1);
431 /* set page dirty and write it */
432 if (gc_type == FG_GC) {
433 f2fs_wait_on_page_writeback(node_page, NODE);
434 set_page_dirty(node_page);
436 if (!PageWriteback(node_page))
437 set_page_dirty(node_page);
439 f2fs_put_page(node_page, 1);
440 stat_inc_node_blk_count(sbi, 1);
448 if (gc_type == FG_GC) {
449 struct writeback_control wbc = {
450 .sync_mode = WB_SYNC_ALL,
451 .nr_to_write = LONG_MAX,
454 sync_node_pages(sbi, 0, &wbc);
457 * In the case of FG_GC, it'd be better to reclaim this victim
460 if (get_valid_blocks(sbi, segno, 1) != 0)
466 * Calculate start block index indicating the given node offset.
467 * Be careful, caller should give this node offset only indicating direct node
468 * blocks. If any node offsets, which point the other types of node blocks such
469 * as indirect or double indirect node blocks, are given, it must be a caller's
472 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
474 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
482 } else if (node_ofs <= indirect_blks) {
483 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
484 bidx = node_ofs - 2 - dec;
486 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
487 bidx = node_ofs - 5 - dec;
489 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
492 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
493 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
495 struct page *node_page;
497 unsigned int ofs_in_node;
498 block_t source_blkaddr;
500 nid = le32_to_cpu(sum->nid);
501 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
503 node_page = get_node_page(sbi, nid);
504 if (IS_ERR(node_page))
507 get_node_info(sbi, nid, dni);
509 if (sum->version != dni->version) {
510 f2fs_put_page(node_page, 1);
514 *nofs = ofs_of_node(node_page);
515 source_blkaddr = datablock_addr(node_page, ofs_in_node);
516 f2fs_put_page(node_page, 1);
518 if (source_blkaddr != blkaddr)
523 static void move_data_page(struct inode *inode, struct page *page, int gc_type)
525 struct f2fs_io_info fio = {
530 if (gc_type == BG_GC) {
531 if (PageWriteback(page))
533 set_page_dirty(page);
536 f2fs_wait_on_page_writeback(page, DATA);
538 if (clear_page_dirty_for_io(page))
539 inode_dec_dirty_pages(inode);
541 do_write_data_page(page, &fio);
542 clear_cold_data(page);
545 f2fs_put_page(page, 1);
549 * This function tries to get parent node of victim data block, and identifies
550 * data block validity. If the block is valid, copy that with cold status and
551 * modify parent node.
552 * If the parent node is not valid or the data block address is different,
553 * the victim data block is ignored.
555 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
556 struct list_head *ilist, unsigned int segno, int gc_type)
558 struct super_block *sb = sbi->sb;
559 struct f2fs_summary *entry;
564 start_addr = START_BLOCK(sbi, segno);
569 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
570 struct page *data_page;
572 struct node_info dni; /* dnode info for the data */
573 unsigned int ofs_in_node, nofs;
576 /* stop BG_GC if there is not enough free sections. */
577 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
580 if (check_valid_map(sbi, segno, off) == 0)
584 ra_node_page(sbi, le32_to_cpu(entry->nid));
588 /* Get an inode by ino with checking validity */
589 if (check_dnode(sbi, entry, &dni, start_addr + off, &nofs) == 0)
593 ra_node_page(sbi, dni.ino);
597 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
600 inode = f2fs_iget(sb, dni.ino);
601 if (IS_ERR(inode) || is_bad_inode(inode))
604 start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
606 data_page = find_data_page(inode,
607 start_bidx + ofs_in_node, false);
608 if (IS_ERR(data_page))
611 f2fs_put_page(data_page, 0);
612 add_gc_inode(inode, ilist);
614 inode = find_gc_inode(dni.ino, ilist);
616 start_bidx = start_bidx_of_node(nofs,
618 data_page = get_lock_data_page(inode,
619 start_bidx + ofs_in_node);
620 if (IS_ERR(data_page))
622 move_data_page(inode, data_page, gc_type);
623 stat_inc_data_blk_count(sbi, 1);
634 if (gc_type == FG_GC) {
635 f2fs_submit_merged_bio(sbi, DATA, WRITE);
638 * In the case of FG_GC, it'd be better to reclaim this victim
641 if (get_valid_blocks(sbi, segno, 1) != 0) {
648 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
649 int gc_type, int type)
651 struct sit_info *sit_i = SIT_I(sbi);
653 mutex_lock(&sit_i->sentry_lock);
654 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS);
655 mutex_unlock(&sit_i->sentry_lock);
659 static void do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
660 struct list_head *ilist, int gc_type)
662 struct page *sum_page;
663 struct f2fs_summary_block *sum;
664 struct blk_plug plug;
666 /* read segment summary of victim */
667 sum_page = get_sum_page(sbi, segno);
669 blk_start_plug(&plug);
671 sum = page_address(sum_page);
673 switch (GET_SUM_TYPE((&sum->footer))) {
675 gc_node_segment(sbi, sum->entries, segno, gc_type);
678 gc_data_segment(sbi, sum->entries, ilist, segno, gc_type);
681 blk_finish_plug(&plug);
683 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)));
684 stat_inc_call_count(sbi->stat_info);
686 f2fs_put_page(sum_page, 1);
689 int f2fs_gc(struct f2fs_sb_info *sbi)
691 struct list_head ilist;
692 unsigned int segno, i;
696 struct cp_control cpc = {
700 INIT_LIST_HEAD(&ilist);
702 if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
704 if (unlikely(f2fs_cp_error(sbi)))
707 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
709 write_checkpoint(sbi, &cpc);
712 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE))
716 /* readahead multi ssa blocks those have contiguous address */
717 if (sbi->segs_per_sec > 1)
718 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), sbi->segs_per_sec,
721 for (i = 0; i < sbi->segs_per_sec; i++)
722 do_garbage_collect(sbi, segno + i, &ilist, gc_type);
724 if (gc_type == FG_GC) {
725 sbi->cur_victim_sec = NULL_SEGNO;
727 WARN_ON(get_valid_blocks(sbi, segno, sbi->segs_per_sec));
730 if (has_not_enough_free_secs(sbi, nfree))
733 if (gc_type == FG_GC)
734 write_checkpoint(sbi, &cpc);
736 mutex_unlock(&sbi->gc_mutex);
738 put_gc_inode(&ilist);
742 void build_gc_manager(struct f2fs_sb_info *sbi)
744 DIRTY_I(sbi)->v_ops = &default_v_ops;
747 int __init create_gc_caches(void)
749 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes",
750 sizeof(struct inode_entry));
756 void destroy_gc_caches(void)
758 kmem_cache_destroy(winode_slab);