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())
49 if (!test_opt(sbi, BG_GC))
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;
94 gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
98 sbi->gc_thread = gc_th;
99 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
100 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
102 if (IS_ERR(gc_th->f2fs_gc_task)) {
109 void stop_gc_thread(struct f2fs_sb_info *sbi)
111 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
114 kthread_stop(gc_th->f2fs_gc_task);
116 sbi->gc_thread = NULL;
119 static int select_gc_type(int gc_type)
121 return (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
124 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
125 int type, struct victim_sel_policy *p)
127 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
130 p->gc_mode = GC_GREEDY;
131 p->dirty_segmap = dirty_i->dirty_segmap[type];
134 p->gc_mode = select_gc_type(gc_type);
135 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
136 p->ofs_unit = sbi->segs_per_sec;
138 p->offset = sbi->last_victim[p->gc_mode];
141 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
142 struct victim_sel_policy *p)
144 if (p->gc_mode == GC_GREEDY)
145 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
146 else if (p->gc_mode == GC_CB)
148 else /* No other gc_mode */
152 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
154 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
158 * If the gc_type is FG_GC, we can select victim segments
159 * selected by background GC before.
160 * Those segments guarantee they have small valid blocks.
162 segno = find_next_bit(dirty_i->victim_segmap[BG_GC],
164 if (segno < TOTAL_SEGS(sbi)) {
165 clear_bit(segno, dirty_i->victim_segmap[BG_GC]);
171 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
173 struct sit_info *sit_i = SIT_I(sbi);
174 unsigned int secno = GET_SECNO(sbi, segno);
175 unsigned int start = secno * sbi->segs_per_sec;
176 unsigned long long mtime = 0;
177 unsigned int vblocks;
178 unsigned char age = 0;
182 for (i = 0; i < sbi->segs_per_sec; i++)
183 mtime += get_seg_entry(sbi, start + i)->mtime;
184 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
186 mtime = div_u64(mtime, sbi->segs_per_sec);
187 vblocks = div_u64(vblocks, sbi->segs_per_sec);
189 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
191 /* Handle if the system time is changed by user */
192 if (mtime < sit_i->min_mtime)
193 sit_i->min_mtime = mtime;
194 if (mtime > sit_i->max_mtime)
195 sit_i->max_mtime = mtime;
196 if (sit_i->max_mtime != sit_i->min_mtime)
197 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
198 sit_i->max_mtime - sit_i->min_mtime);
200 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
203 static unsigned int get_gc_cost(struct f2fs_sb_info *sbi, unsigned int segno,
204 struct victim_sel_policy *p)
206 if (p->alloc_mode == SSR)
207 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
209 /* alloc_mode == LFS */
210 if (p->gc_mode == GC_GREEDY)
211 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
213 return get_cb_cost(sbi, segno);
217 * This function is called from two pathes.
218 * One is garbage collection and the other is SSR segment selection.
219 * When it is called during GC, it just gets a victim segment
220 * and it does not remove it from dirty seglist.
221 * When it is called from SSR segment selection, it finds a segment
222 * which has minimum valid blocks and removes it from dirty seglist.
224 static int get_victim_by_default(struct f2fs_sb_info *sbi,
225 unsigned int *result, int gc_type, int type, char alloc_mode)
227 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
228 struct victim_sel_policy p;
232 p.alloc_mode = alloc_mode;
233 select_policy(sbi, gc_type, type, &p);
235 p.min_segno = NULL_SEGNO;
236 p.min_cost = get_max_cost(sbi, &p);
238 mutex_lock(&dirty_i->seglist_lock);
240 if (p.alloc_mode == LFS && gc_type == FG_GC) {
241 p.min_segno = check_bg_victims(sbi);
242 if (p.min_segno != NULL_SEGNO)
249 segno = find_next_bit(p.dirty_segmap,
250 TOTAL_SEGS(sbi), p.offset);
251 if (segno >= TOTAL_SEGS(sbi)) {
252 if (sbi->last_victim[p.gc_mode]) {
253 sbi->last_victim[p.gc_mode] = 0;
259 p.offset = ((segno / p.ofs_unit) * p.ofs_unit) + p.ofs_unit;
261 if (test_bit(segno, dirty_i->victim_segmap[FG_GC]))
263 if (gc_type == BG_GC &&
264 test_bit(segno, dirty_i->victim_segmap[BG_GC]))
266 if (IS_CURSEC(sbi, GET_SECNO(sbi, segno)))
269 cost = get_gc_cost(sbi, segno, &p);
271 if (p.min_cost > cost) {
276 if (cost == get_max_cost(sbi, &p))
279 if (nsearched++ >= MAX_VICTIM_SEARCH) {
280 sbi->last_victim[p.gc_mode] = segno;
285 if (p.min_segno != NULL_SEGNO) {
286 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
287 if (p.alloc_mode == LFS) {
289 for (i = 0; i < p.ofs_unit; i++)
291 dirty_i->victim_segmap[gc_type]);
294 mutex_unlock(&dirty_i->seglist_lock);
296 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
299 static const struct victim_selection default_v_ops = {
300 .get_victim = get_victim_by_default,
303 static struct inode *find_gc_inode(nid_t ino, struct list_head *ilist)
305 struct list_head *this;
306 struct inode_entry *ie;
308 list_for_each(this, ilist) {
309 ie = list_entry(this, struct inode_entry, list);
310 if (ie->inode->i_ino == ino)
316 static void add_gc_inode(struct inode *inode, struct list_head *ilist)
318 struct list_head *this;
319 struct inode_entry *new_ie, *ie;
321 list_for_each(this, ilist) {
322 ie = list_entry(this, struct inode_entry, list);
323 if (ie->inode == inode) {
329 new_ie = kmem_cache_alloc(winode_slab, GFP_NOFS);
334 new_ie->inode = inode;
335 list_add_tail(&new_ie->list, ilist);
338 static void put_gc_inode(struct list_head *ilist)
340 struct inode_entry *ie, *next_ie;
341 list_for_each_entry_safe(ie, next_ie, ilist, list) {
344 kmem_cache_free(winode_slab, ie);
348 static int check_valid_map(struct f2fs_sb_info *sbi,
349 unsigned int segno, int offset)
351 struct sit_info *sit_i = SIT_I(sbi);
352 struct seg_entry *sentry;
355 mutex_lock(&sit_i->sentry_lock);
356 sentry = get_seg_entry(sbi, segno);
357 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
358 mutex_unlock(&sit_i->sentry_lock);
359 return ret ? GC_OK : GC_NEXT;
363 * This function compares node address got in summary with that in NAT.
364 * On validity, copy that node with cold status, otherwise (invalid node)
367 static int gc_node_segment(struct f2fs_sb_info *sbi,
368 struct f2fs_summary *sum, unsigned int segno, int gc_type)
371 struct f2fs_summary *entry;
376 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
377 nid_t nid = le32_to_cpu(entry->nid);
378 struct page *node_page;
382 * It makes sure that free segments are able to write
383 * all the dirty node pages before CP after this CP.
384 * So let's check the space of dirty node pages.
386 if (should_do_checkpoint(sbi)) {
387 mutex_lock(&sbi->cp_mutex);
388 block_operations(sbi);
392 err = check_valid_map(sbi, segno, off);
397 ra_node_page(sbi, nid);
400 node_page = get_node_page(sbi, nid);
401 if (IS_ERR(node_page))
404 /* set page dirty and write it */
405 if (!PageWriteback(node_page))
406 set_page_dirty(node_page);
407 f2fs_put_page(node_page, 1);
408 stat_inc_node_blk_count(sbi, 1);
415 if (gc_type == FG_GC) {
416 struct writeback_control wbc = {
417 .sync_mode = WB_SYNC_ALL,
418 .nr_to_write = LONG_MAX,
421 sync_node_pages(sbi, 0, &wbc);
427 * Calculate start block index indicating the given node offset.
428 * Be careful, caller should give this node offset only indicating direct node
429 * blocks. If any node offsets, which point the other types of node blocks such
430 * as indirect or double indirect node blocks, are given, it must be a caller's
433 block_t start_bidx_of_node(unsigned int node_ofs)
435 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
443 } else if (node_ofs <= indirect_blks) {
444 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
445 bidx = node_ofs - 2 - dec;
447 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
448 bidx = node_ofs - 5 - dec;
450 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE;
453 static int check_dnode(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
454 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
456 struct page *node_page;
458 unsigned int ofs_in_node;
459 block_t source_blkaddr;
461 nid = le32_to_cpu(sum->nid);
462 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
464 node_page = get_node_page(sbi, nid);
465 if (IS_ERR(node_page))
468 get_node_info(sbi, nid, dni);
470 if (sum->version != dni->version) {
471 f2fs_put_page(node_page, 1);
475 *nofs = ofs_of_node(node_page);
476 source_blkaddr = datablock_addr(node_page, ofs_in_node);
477 f2fs_put_page(node_page, 1);
479 if (source_blkaddr != blkaddr)
484 static void move_data_page(struct inode *inode, struct page *page, int gc_type)
486 if (page->mapping != inode->i_mapping)
489 if (inode != page->mapping->host)
492 if (PageWriteback(page))
495 if (gc_type == BG_GC) {
496 set_page_dirty(page);
499 struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb);
500 mutex_lock_op(sbi, DATA_WRITE);
501 if (clear_page_dirty_for_io(page) &&
502 S_ISDIR(inode->i_mode)) {
503 dec_page_count(sbi, F2FS_DIRTY_DENTS);
504 inode_dec_dirty_dents(inode);
507 do_write_data_page(page);
508 mutex_unlock_op(sbi, DATA_WRITE);
509 clear_cold_data(page);
512 f2fs_put_page(page, 1);
516 * This function tries to get parent node of victim data block, and identifies
517 * data block validity. If the block is valid, copy that with cold status and
518 * modify parent node.
519 * If the parent node is not valid or the data block address is different,
520 * the victim data block is ignored.
522 static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
523 struct list_head *ilist, unsigned int segno, int gc_type)
525 struct super_block *sb = sbi->sb;
526 struct f2fs_summary *entry;
531 start_addr = START_BLOCK(sbi, segno);
535 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
536 struct page *data_page;
538 struct node_info dni; /* dnode info for the data */
539 unsigned int ofs_in_node, nofs;
543 * It makes sure that free segments are able to write
544 * all the dirty node pages before CP after this CP.
545 * So let's check the space of dirty node pages.
547 if (should_do_checkpoint(sbi)) {
548 mutex_lock(&sbi->cp_mutex);
549 block_operations(sbi);
554 err = check_valid_map(sbi, segno, off);
559 ra_node_page(sbi, le32_to_cpu(entry->nid));
563 /* Get an inode by ino with checking validity */
564 err = check_dnode(sbi, entry, &dni, start_addr + off, &nofs);
569 ra_node_page(sbi, dni.ino);
573 start_bidx = start_bidx_of_node(nofs);
574 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
577 inode = f2fs_iget_nowait(sb, dni.ino);
581 data_page = find_data_page(inode,
582 start_bidx + ofs_in_node);
583 if (IS_ERR(data_page))
586 f2fs_put_page(data_page, 0);
587 add_gc_inode(inode, ilist);
589 inode = find_gc_inode(dni.ino, ilist);
591 data_page = get_lock_data_page(inode,
592 start_bidx + ofs_in_node);
593 if (IS_ERR(data_page))
595 move_data_page(inode, data_page, gc_type);
596 stat_inc_data_blk_count(sbi, 1);
607 if (gc_type == FG_GC)
608 f2fs_submit_bio(sbi, DATA, true);
612 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
613 int gc_type, int type)
615 struct sit_info *sit_i = SIT_I(sbi);
617 mutex_lock(&sit_i->sentry_lock);
618 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type, type, LFS);
619 mutex_unlock(&sit_i->sentry_lock);
623 static int do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
624 struct list_head *ilist, int gc_type)
626 struct page *sum_page;
627 struct f2fs_summary_block *sum;
630 /* read segment summary of victim */
631 sum_page = get_sum_page(sbi, segno);
632 if (IS_ERR(sum_page))
636 * CP needs to lock sum_page. In this time, we don't need
637 * to lock this page, because this summary page is not gone anywhere.
638 * Also, this page is not gonna be updated before GC is done.
640 unlock_page(sum_page);
641 sum = page_address(sum_page);
643 switch (GET_SUM_TYPE((&sum->footer))) {
645 ret = gc_node_segment(sbi, sum->entries, segno, gc_type);
648 ret = gc_data_segment(sbi, sum->entries, ilist, segno, gc_type);
651 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)));
652 stat_inc_call_count(sbi->stat_info);
654 f2fs_put_page(sum_page, 0);
658 int f2fs_gc(struct f2fs_sb_info *sbi)
660 struct list_head ilist;
661 unsigned int segno, i;
663 int gc_status = GC_NONE;
665 INIT_LIST_HEAD(&ilist);
667 if (!(sbi->sb->s_flags & MS_ACTIVE))
670 if (has_not_enough_free_secs(sbi))
673 if (!__get_victim(sbi, &segno, gc_type, NO_CHECK_TYPE))
676 for (i = 0; i < sbi->segs_per_sec; i++) {
678 * do_garbage_collect will give us three gc_status:
679 * GC_ERROR, GC_DONE, and GC_BLOCKED.
680 * If GC is finished uncleanly, we have to return
681 * the victim to dirty segment list.
683 gc_status = do_garbage_collect(sbi, segno + i, &ilist, gc_type);
684 if (gc_status != GC_DONE)
687 if (has_not_enough_free_secs(sbi)) {
688 write_checkpoint(sbi, (gc_status == GC_BLOCKED), false);
689 if (has_not_enough_free_secs(sbi))
693 mutex_unlock(&sbi->gc_mutex);
695 put_gc_inode(&ilist);
699 void build_gc_manager(struct f2fs_sb_info *sbi)
701 DIRTY_I(sbi)->v_ops = &default_v_ops;
704 int __init create_gc_caches(void)
706 winode_slab = f2fs_kmem_cache_create("f2fs_gc_inodes",
707 sizeof(struct inode_entry), NULL);
713 void destroy_gc_caches(void)
715 kmem_cache_destroy(winode_slab);