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 int gc_thread_func(void *data)
29 struct f2fs_sb_info *sbi = data;
30 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
31 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
34 wait_ms = gc_th->min_sleep_time;
40 wait_event_interruptible_timeout(*wq,
41 kthread_should_stop(),
42 msecs_to_jiffies(wait_ms));
43 if (kthread_should_stop())
46 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
47 increase_sleep_time(gc_th, &wait_ms);
52 * [GC triggering condition]
53 * 0. GC is not conducted currently.
54 * 1. There are enough dirty segments.
55 * 2. IO subsystem is idle by checking the # of writeback pages.
56 * 3. IO subsystem is idle by checking the # of requests in
57 * bdev's request list.
59 * Note) We have to avoid triggering GCs frequently.
60 * Because it is possible that some segments can be
61 * invalidated soon after by user update or deletion.
62 * So, I'd like to wait some time to collect dirty segments.
64 if (!mutex_trylock(&sbi->gc_mutex))
68 increase_sleep_time(gc_th, &wait_ms);
69 mutex_unlock(&sbi->gc_mutex);
73 if (has_enough_invalid_blocks(sbi))
74 decrease_sleep_time(gc_th, &wait_ms);
76 increase_sleep_time(gc_th, &wait_ms);
78 stat_inc_bggc_count(sbi);
80 /* if return value is not zero, no victim was selected */
82 wait_ms = gc_th->no_gc_sleep_time;
84 /* balancing f2fs's metadata periodically */
85 f2fs_balance_fs_bg(sbi);
87 } while (!kthread_should_stop());
91 int start_gc_thread(struct f2fs_sb_info *sbi)
93 struct f2fs_gc_kthread *gc_th;
94 dev_t dev = sbi->sb->s_bdev->bd_dev;
97 gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
103 gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
104 gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
105 gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
109 sbi->gc_thread = gc_th;
110 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
111 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
112 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
113 if (IS_ERR(gc_th->f2fs_gc_task)) {
114 err = PTR_ERR(gc_th->f2fs_gc_task);
116 sbi->gc_thread = NULL;
122 void stop_gc_thread(struct f2fs_sb_info *sbi)
124 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
127 kthread_stop(gc_th->f2fs_gc_task);
129 sbi->gc_thread = NULL;
132 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
134 int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
136 if (gc_th && gc_th->gc_idle) {
137 if (gc_th->gc_idle == 1)
139 else if (gc_th->gc_idle == 2)
145 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
146 int type, struct victim_sel_policy *p)
148 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
150 if (p->alloc_mode == SSR) {
151 p->gc_mode = GC_GREEDY;
152 p->dirty_segmap = dirty_i->dirty_segmap[type];
153 p->max_search = dirty_i->nr_dirty[type];
156 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
157 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
158 p->max_search = dirty_i->nr_dirty[DIRTY];
159 p->ofs_unit = sbi->segs_per_sec;
162 if (p->max_search > sbi->max_victim_search)
163 p->max_search = sbi->max_victim_search;
165 p->offset = sbi->last_victim[p->gc_mode];
168 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
169 struct victim_sel_policy *p)
171 /* SSR allocates in a segment unit */
172 if (p->alloc_mode == SSR)
173 return 1 << sbi->log_blocks_per_seg;
174 if (p->gc_mode == GC_GREEDY)
175 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
176 else if (p->gc_mode == GC_CB)
178 else /* No other gc_mode */
182 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
184 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
188 * If the gc_type is FG_GC, we can select victim segments
189 * selected by background GC before.
190 * Those segments guarantee they have small valid blocks.
192 for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
193 if (sec_usage_check(sbi, secno))
195 clear_bit(secno, dirty_i->victim_secmap);
196 return secno * sbi->segs_per_sec;
201 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
203 struct sit_info *sit_i = SIT_I(sbi);
204 unsigned int secno = GET_SECNO(sbi, segno);
205 unsigned int start = secno * sbi->segs_per_sec;
206 unsigned long long mtime = 0;
207 unsigned int vblocks;
208 unsigned char age = 0;
212 for (i = 0; i < sbi->segs_per_sec; i++)
213 mtime += get_seg_entry(sbi, start + i)->mtime;
214 vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
216 mtime = div_u64(mtime, sbi->segs_per_sec);
217 vblocks = div_u64(vblocks, sbi->segs_per_sec);
219 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
221 /* Handle if the system time has changed by the user */
222 if (mtime < sit_i->min_mtime)
223 sit_i->min_mtime = mtime;
224 if (mtime > sit_i->max_mtime)
225 sit_i->max_mtime = mtime;
226 if (sit_i->max_mtime != sit_i->min_mtime)
227 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
228 sit_i->max_mtime - sit_i->min_mtime);
230 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
233 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
234 unsigned int segno, struct victim_sel_policy *p)
236 if (p->alloc_mode == SSR)
237 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
239 /* alloc_mode == LFS */
240 if (p->gc_mode == GC_GREEDY)
241 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
243 return get_cb_cost(sbi, segno);
247 * This function is called from two paths.
248 * One is garbage collection and the other is SSR segment selection.
249 * When it is called during GC, it just gets a victim segment
250 * and it does not remove it from dirty seglist.
251 * When it is called from SSR segment selection, it finds a segment
252 * which has minimum valid blocks and removes it from dirty seglist.
254 static int get_victim_by_default(struct f2fs_sb_info *sbi,
255 unsigned int *result, int gc_type, int type, char alloc_mode)
257 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
258 struct victim_sel_policy p;
259 unsigned int secno, max_cost;
262 mutex_lock(&dirty_i->seglist_lock);
264 p.alloc_mode = alloc_mode;
265 select_policy(sbi, gc_type, type, &p);
267 p.min_segno = NULL_SEGNO;
268 p.min_cost = max_cost = get_max_cost(sbi, &p);
270 if (p.alloc_mode == LFS && gc_type == FG_GC) {
271 p.min_segno = check_bg_victims(sbi);
272 if (p.min_segno != NULL_SEGNO)
280 segno = find_next_bit(p.dirty_segmap, MAIN_SEGS(sbi), p.offset);
281 if (segno >= MAIN_SEGS(sbi)) {
282 if (sbi->last_victim[p.gc_mode]) {
283 sbi->last_victim[p.gc_mode] = 0;
290 p.offset = segno + p.ofs_unit;
292 p.offset -= segno % p.ofs_unit;
294 secno = GET_SECNO(sbi, segno);
296 if (sec_usage_check(sbi, secno))
298 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
301 cost = get_gc_cost(sbi, segno, &p);
303 if (p.min_cost > cost) {
306 } else if (unlikely(cost == max_cost)) {
310 if (nsearched++ >= p.max_search) {
311 sbi->last_victim[p.gc_mode] = segno;
315 if (p.min_segno != NULL_SEGNO) {
317 if (p.alloc_mode == LFS) {
318 secno = GET_SECNO(sbi, p.min_segno);
319 if (gc_type == FG_GC)
320 sbi->cur_victim_sec = secno;
322 set_bit(secno, dirty_i->victim_secmap);
324 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
326 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
328 prefree_segments(sbi), free_segments(sbi));
330 mutex_unlock(&dirty_i->seglist_lock);
332 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
335 static const struct victim_selection default_v_ops = {
336 .get_victim = get_victim_by_default,
339 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
341 struct inode_entry *ie;
343 ie = radix_tree_lookup(&gc_list->iroot, ino);
349 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
351 struct inode_entry *new_ie;
353 if (inode == find_gc_inode(gc_list, inode->i_ino)) {
357 new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
358 new_ie->inode = inode;
360 f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
361 list_add_tail(&new_ie->list, &gc_list->ilist);
364 static void put_gc_inode(struct gc_inode_list *gc_list)
366 struct inode_entry *ie, *next_ie;
367 list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
368 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
371 kmem_cache_free(inode_entry_slab, ie);
375 static int check_valid_map(struct f2fs_sb_info *sbi,
376 unsigned int segno, int offset)
378 struct sit_info *sit_i = SIT_I(sbi);
379 struct seg_entry *sentry;
382 mutex_lock(&sit_i->sentry_lock);
383 sentry = get_seg_entry(sbi, segno);
384 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
385 mutex_unlock(&sit_i->sentry_lock);
390 * This function compares node address got in summary with that in NAT.
391 * On validity, copy that node with cold status, otherwise (invalid node)
394 static int gc_node_segment(struct f2fs_sb_info *sbi,
395 struct f2fs_summary *sum, unsigned int segno, int gc_type)
398 struct f2fs_summary *entry;
402 start_addr = START_BLOCK(sbi, segno);
407 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
408 nid_t nid = le32_to_cpu(entry->nid);
409 struct page *node_page;
412 /* stop BG_GC if there is not enough free sections. */
413 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
416 if (check_valid_map(sbi, segno, off) == 0)
420 ra_node_page(sbi, nid);
423 node_page = get_node_page(sbi, nid);
424 if (IS_ERR(node_page))
427 /* block may become invalid during get_node_page */
428 if (check_valid_map(sbi, segno, off) == 0) {
429 f2fs_put_page(node_page, 1);
433 get_node_info(sbi, nid, &ni);
434 if (ni.blk_addr != start_addr + off) {
435 f2fs_put_page(node_page, 1);
439 /* set page dirty and write it */
440 if (gc_type == FG_GC) {
441 f2fs_wait_on_page_writeback(node_page, NODE);
442 set_page_dirty(node_page);
444 if (!PageWriteback(node_page))
445 set_page_dirty(node_page);
447 f2fs_put_page(node_page, 1);
448 stat_inc_node_blk_count(sbi, 1, gc_type);
456 if (gc_type == FG_GC) {
457 struct writeback_control wbc = {
458 .sync_mode = WB_SYNC_ALL,
459 .nr_to_write = LONG_MAX,
462 sync_node_pages(sbi, 0, &wbc);
464 /* return 1 only if FG_GC succefully reclaimed one */
465 if (get_valid_blocks(sbi, segno, 1) == 0)
472 * Calculate start block index indicating the given node offset.
473 * Be careful, caller should give this node offset only indicating direct node
474 * blocks. If any node offsets, which point the other types of node blocks such
475 * as indirect or double indirect node blocks, are given, it must be a caller's
478 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
480 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
488 } else if (node_ofs <= indirect_blks) {
489 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
490 bidx = node_ofs - 2 - dec;
492 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
493 bidx = node_ofs - 5 - dec;
495 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
498 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
499 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
501 struct page *node_page;
503 unsigned int ofs_in_node;
504 block_t source_blkaddr;
506 nid = le32_to_cpu(sum->nid);
507 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
509 node_page = get_node_page(sbi, nid);
510 if (IS_ERR(node_page))
513 get_node_info(sbi, nid, dni);
515 if (sum->version != dni->version) {
516 f2fs_put_page(node_page, 1);
520 *nofs = ofs_of_node(node_page);
521 source_blkaddr = datablock_addr(node_page, ofs_in_node);
522 f2fs_put_page(node_page, 1);
524 if (source_blkaddr != blkaddr)
529 static void move_encrypted_block(struct inode *inode, block_t bidx)
531 struct f2fs_io_info fio = {
532 .sbi = F2FS_I_SB(inode),
535 .encrypted_page = NULL,
537 struct dnode_of_data dn;
538 struct f2fs_summary sum;
543 /* do not read out */
544 page = grab_cache_page(inode->i_mapping, bidx);
548 set_new_dnode(&dn, inode, NULL, NULL, 0);
549 err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
553 if (unlikely(dn.data_blkaddr == NULL_ADDR))
556 get_node_info(fio.sbi, dn.nid, &ni);
557 set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
561 fio.blk_addr = dn.data_blkaddr;
563 fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi),
567 if (!fio.encrypted_page)
570 err = f2fs_submit_page_bio(&fio);
575 lock_page(fio.encrypted_page);
577 if (unlikely(!PageUptodate(fio.encrypted_page)))
579 if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi)))
582 set_page_dirty(fio.encrypted_page);
583 f2fs_wait_on_page_writeback(fio.encrypted_page, META);
584 if (clear_page_dirty_for_io(fio.encrypted_page))
585 dec_page_count(fio.sbi, F2FS_DIRTY_META);
587 set_page_writeback(fio.encrypted_page);
589 /* allocate block address */
590 f2fs_wait_on_page_writeback(dn.node_page, NODE);
591 allocate_data_block(fio.sbi, NULL, fio.blk_addr,
592 &fio.blk_addr, &sum, CURSEG_COLD_DATA);
594 f2fs_submit_page_mbio(&fio);
596 dn.data_blkaddr = fio.blk_addr;
597 set_data_blkaddr(&dn);
598 f2fs_update_extent_cache(&dn);
599 set_inode_flag(F2FS_I(inode), FI_APPEND_WRITE);
600 if (page->index == 0)
601 set_inode_flag(F2FS_I(inode), FI_FIRST_BLOCK_WRITTEN);
603 f2fs_put_page(fio.encrypted_page, 1);
607 f2fs_put_page(page, 1);
610 static void move_data_page(struct inode *inode, block_t bidx, int gc_type)
614 page = get_lock_data_page(inode, bidx);
618 if (gc_type == BG_GC) {
619 if (PageWriteback(page))
621 set_page_dirty(page);
624 struct f2fs_io_info fio = {
625 .sbi = F2FS_I_SB(inode),
629 .encrypted_page = NULL,
631 set_page_dirty(page);
632 f2fs_wait_on_page_writeback(page, DATA);
633 if (clear_page_dirty_for_io(page))
634 inode_dec_dirty_pages(inode);
636 do_write_data_page(&fio);
637 clear_cold_data(page);
640 f2fs_put_page(page, 1);
644 * This function tries to get parent node of victim data block, and identifies
645 * data block validity. If the block is valid, copy that with cold status and
646 * modify parent node.
647 * If the parent node is not valid or the data block address is different,
648 * the victim data block is ignored.
650 static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
651 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
653 struct super_block *sb = sbi->sb;
654 struct f2fs_summary *entry;
659 start_addr = START_BLOCK(sbi, segno);
664 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
665 struct page *data_page;
667 struct node_info dni; /* dnode info for the data */
668 unsigned int ofs_in_node, nofs;
671 /* stop BG_GC if there is not enough free sections. */
672 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
675 if (check_valid_map(sbi, segno, off) == 0)
679 ra_node_page(sbi, le32_to_cpu(entry->nid));
683 /* Get an inode by ino with checking validity */
684 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
688 ra_node_page(sbi, dni.ino);
692 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
695 inode = f2fs_iget(sb, dni.ino);
696 if (IS_ERR(inode) || is_bad_inode(inode))
699 /* if encrypted inode, let's go phase 3 */
700 if (f2fs_encrypted_inode(inode) &&
701 S_ISREG(inode->i_mode)) {
702 add_gc_inode(gc_list, inode);
706 start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
707 data_page = get_read_data_page(inode,
708 start_bidx + ofs_in_node, READA);
709 if (IS_ERR(data_page)) {
714 f2fs_put_page(data_page, 0);
715 add_gc_inode(gc_list, inode);
720 inode = find_gc_inode(gc_list, dni.ino);
722 start_bidx = start_bidx_of_node(nofs, F2FS_I(inode))
724 if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
725 move_encrypted_block(inode, start_bidx);
727 move_data_page(inode, start_bidx, gc_type);
728 stat_inc_data_blk_count(sbi, 1, gc_type);
735 if (gc_type == FG_GC) {
736 f2fs_submit_merged_bio(sbi, DATA, WRITE);
738 /* return 1 only if FG_GC succefully reclaimed one */
739 if (get_valid_blocks(sbi, segno, 1) == 0)
745 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
748 struct sit_info *sit_i = SIT_I(sbi);
751 mutex_lock(&sit_i->sentry_lock);
752 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
754 mutex_unlock(&sit_i->sentry_lock);
758 static int do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
759 struct gc_inode_list *gc_list, int gc_type)
761 struct page *sum_page;
762 struct f2fs_summary_block *sum;
763 struct blk_plug plug;
766 /* read segment summary of victim */
767 sum_page = get_sum_page(sbi, segno);
769 blk_start_plug(&plug);
771 sum = page_address(sum_page);
774 * this is to avoid deadlock:
775 * - lock_page(sum_page) - f2fs_replace_block
776 * - check_valid_map() - mutex_lock(sentry_lock)
777 * - mutex_lock(sentry_lock) - change_curseg()
778 * - lock_page(sum_page)
780 unlock_page(sum_page);
782 switch (GET_SUM_TYPE((&sum->footer))) {
784 nfree = gc_node_segment(sbi, sum->entries, segno, gc_type);
787 nfree = gc_data_segment(sbi, sum->entries, gc_list,
791 blk_finish_plug(&plug);
793 stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)), gc_type);
794 stat_inc_call_count(sbi->stat_info);
796 f2fs_put_page(sum_page, 0);
800 int f2fs_gc(struct f2fs_sb_info *sbi)
802 unsigned int segno = NULL_SEGNO;
807 struct cp_control cpc;
808 struct gc_inode_list gc_list = {
809 .ilist = LIST_HEAD_INIT(gc_list.ilist),
810 .iroot = RADIX_TREE_INIT(GFP_NOFS),
813 cpc.reason = __get_cp_reason(sbi);
815 if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
817 if (unlikely(f2fs_cp_error(sbi)))
820 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
822 if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
823 write_checkpoint(sbi, &cpc);
826 if (segno == NULL_SEGNO && !__get_victim(sbi, &segno, gc_type))
830 /* readahead multi ssa blocks those have contiguous address */
831 if (sbi->segs_per_sec > 1)
832 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), sbi->segs_per_sec,
835 for (i = 0; i < sbi->segs_per_sec; i++)
836 nfree += do_garbage_collect(sbi, segno + i, &gc_list, gc_type);
838 if (gc_type == FG_GC)
839 sbi->cur_victim_sec = NULL_SEGNO;
841 if (has_not_enough_free_secs(sbi, nfree))
844 if (gc_type == FG_GC)
845 write_checkpoint(sbi, &cpc);
847 mutex_unlock(&sbi->gc_mutex);
849 put_gc_inode(&gc_list);
853 void build_gc_manager(struct f2fs_sb_info *sbi)
855 DIRTY_I(sbi)->v_ops = &default_v_ops;