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>
24 #include <trace/events/f2fs.h>
26 static int gc_thread_func(void *data)
28 struct f2fs_sb_info *sbi = data;
29 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
30 wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
33 wait_ms = gc_th->min_sleep_time;
37 wait_event_interruptible_timeout(*wq,
38 kthread_should_stop() || freezing(current) ||
40 msecs_to_jiffies(wait_ms));
42 /* give it a try one time */
48 if (kthread_should_stop())
51 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
52 increase_sleep_time(gc_th, &wait_ms);
56 if (time_to_inject(sbi, FAULT_CHECKPOINT)) {
57 f2fs_show_injection_info(FAULT_CHECKPOINT);
58 f2fs_stop_checkpoint(sbi, false);
61 if (!sb_start_write_trylock(sbi->sb))
65 * [GC triggering condition]
66 * 0. GC is not conducted currently.
67 * 1. There are enough dirty segments.
68 * 2. IO subsystem is idle by checking the # of writeback pages.
69 * 3. IO subsystem is idle by checking the # of requests in
70 * bdev's request list.
72 * Note) We have to avoid triggering GCs frequently.
73 * Because it is possible that some segments can be
74 * invalidated soon after by user update or deletion.
75 * So, I'd like to wait some time to collect dirty segments.
77 if (sbi->gc_mode == GC_URGENT) {
78 wait_ms = gc_th->urgent_sleep_time;
79 mutex_lock(&sbi->gc_mutex);
83 if (!mutex_trylock(&sbi->gc_mutex))
87 increase_sleep_time(gc_th, &wait_ms);
88 mutex_unlock(&sbi->gc_mutex);
92 if (has_enough_invalid_blocks(sbi))
93 decrease_sleep_time(gc_th, &wait_ms);
95 increase_sleep_time(gc_th, &wait_ms);
97 stat_inc_bggc_count(sbi);
99 /* if return value is not zero, no victim was selected */
100 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true, NULL_SEGNO))
101 wait_ms = gc_th->no_gc_sleep_time;
103 trace_f2fs_background_gc(sbi->sb, wait_ms,
104 prefree_segments(sbi), free_segments(sbi));
106 /* balancing f2fs's metadata periodically */
107 f2fs_balance_fs_bg(sbi);
109 sb_end_write(sbi->sb);
111 } while (!kthread_should_stop());
115 int f2fs_start_gc_thread(struct f2fs_sb_info *sbi)
117 struct f2fs_gc_kthread *gc_th;
118 dev_t dev = sbi->sb->s_bdev->bd_dev;
121 gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
127 gc_th->urgent_sleep_time = DEF_GC_THREAD_URGENT_SLEEP_TIME;
128 gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
129 gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
130 gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
134 sbi->gc_thread = gc_th;
135 init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
136 sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
137 "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
138 if (IS_ERR(gc_th->f2fs_gc_task)) {
139 err = PTR_ERR(gc_th->f2fs_gc_task);
141 sbi->gc_thread = NULL;
147 void f2fs_stop_gc_thread(struct f2fs_sb_info *sbi)
149 struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
152 kthread_stop(gc_th->f2fs_gc_task);
154 sbi->gc_thread = NULL;
157 static int select_gc_type(struct f2fs_sb_info *sbi, int gc_type)
159 int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
161 switch (sbi->gc_mode) {
173 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
174 int type, struct victim_sel_policy *p)
176 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
178 if (p->alloc_mode == SSR) {
179 p->gc_mode = GC_GREEDY;
180 p->dirty_segmap = dirty_i->dirty_segmap[type];
181 p->max_search = dirty_i->nr_dirty[type];
184 p->gc_mode = select_gc_type(sbi, gc_type);
185 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
186 p->max_search = dirty_i->nr_dirty[DIRTY];
187 p->ofs_unit = sbi->segs_per_sec;
190 /* we need to check every dirty segments in the FG_GC case */
191 if (gc_type != FG_GC &&
192 (sbi->gc_mode != GC_URGENT) &&
193 p->max_search > sbi->max_victim_search)
194 p->max_search = sbi->max_victim_search;
196 /* let's select beginning hot/small space first in no_heap mode*/
197 if (test_opt(sbi, NOHEAP) &&
198 (type == CURSEG_HOT_DATA || IS_NODESEG(type)))
201 p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
204 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
205 struct victim_sel_policy *p)
207 /* SSR allocates in a segment unit */
208 if (p->alloc_mode == SSR)
209 return sbi->blocks_per_seg;
210 if (p->gc_mode == GC_GREEDY)
211 return 2 * sbi->blocks_per_seg * p->ofs_unit;
212 else if (p->gc_mode == GC_CB)
214 else /* No other gc_mode */
218 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
220 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
224 * If the gc_type is FG_GC, we can select victim segments
225 * selected by background GC before.
226 * Those segments guarantee they have small valid blocks.
228 for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
229 if (sec_usage_check(sbi, secno))
231 clear_bit(secno, dirty_i->victim_secmap);
232 return GET_SEG_FROM_SEC(sbi, secno);
237 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
239 struct sit_info *sit_i = SIT_I(sbi);
240 unsigned int secno = GET_SEC_FROM_SEG(sbi, segno);
241 unsigned int start = GET_SEG_FROM_SEC(sbi, secno);
242 unsigned long long mtime = 0;
243 unsigned int vblocks;
244 unsigned char age = 0;
248 for (i = 0; i < sbi->segs_per_sec; i++)
249 mtime += get_seg_entry(sbi, start + i)->mtime;
250 vblocks = get_valid_blocks(sbi, segno, true);
252 mtime = div_u64(mtime, sbi->segs_per_sec);
253 vblocks = div_u64(vblocks, sbi->segs_per_sec);
255 u = (vblocks * 100) >> sbi->log_blocks_per_seg;
257 /* Handle if the system time has changed by the user */
258 if (mtime < sit_i->min_mtime)
259 sit_i->min_mtime = mtime;
260 if (mtime > sit_i->max_mtime)
261 sit_i->max_mtime = mtime;
262 if (sit_i->max_mtime != sit_i->min_mtime)
263 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
264 sit_i->max_mtime - sit_i->min_mtime);
266 return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
269 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
270 unsigned int segno, struct victim_sel_policy *p)
272 if (p->alloc_mode == SSR)
273 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
275 /* alloc_mode == LFS */
276 if (p->gc_mode == GC_GREEDY)
277 return get_valid_blocks(sbi, segno, true);
279 return get_cb_cost(sbi, segno);
282 static unsigned int count_bits(const unsigned long *addr,
283 unsigned int offset, unsigned int len)
285 unsigned int end = offset + len, sum = 0;
287 while (offset < end) {
288 if (test_bit(offset++, addr))
295 * This function is called from two paths.
296 * One is garbage collection and the other is SSR segment selection.
297 * When it is called during GC, it just gets a victim segment
298 * and it does not remove it from dirty seglist.
299 * When it is called from SSR segment selection, it finds a segment
300 * which has minimum valid blocks and removes it from dirty seglist.
302 static int get_victim_by_default(struct f2fs_sb_info *sbi,
303 unsigned int *result, int gc_type, int type, char alloc_mode)
305 struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
306 struct sit_info *sm = SIT_I(sbi);
307 struct victim_sel_policy p;
308 unsigned int secno, last_victim;
309 unsigned int last_segment = MAIN_SEGS(sbi);
310 unsigned int nsearched = 0;
312 mutex_lock(&dirty_i->seglist_lock);
314 p.alloc_mode = alloc_mode;
315 select_policy(sbi, gc_type, type, &p);
317 p.min_segno = NULL_SEGNO;
318 p.min_cost = get_max_cost(sbi, &p);
320 if (*result != NULL_SEGNO) {
321 if (get_valid_blocks(sbi, *result, false) &&
322 !sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result)))
323 p.min_segno = *result;
327 if (p.max_search == 0)
330 last_victim = sm->last_victim[p.gc_mode];
331 if (p.alloc_mode == LFS && gc_type == FG_GC) {
332 p.min_segno = check_bg_victims(sbi);
333 if (p.min_segno != NULL_SEGNO)
341 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
342 if (segno >= last_segment) {
343 if (sm->last_victim[p.gc_mode]) {
345 sm->last_victim[p.gc_mode];
346 sm->last_victim[p.gc_mode] = 0;
353 p.offset = segno + p.ofs_unit;
354 if (p.ofs_unit > 1) {
355 p.offset -= segno % p.ofs_unit;
356 nsearched += count_bits(p.dirty_segmap,
357 p.offset - p.ofs_unit,
363 secno = GET_SEC_FROM_SEG(sbi, segno);
365 if (sec_usage_check(sbi, secno))
367 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
370 cost = get_gc_cost(sbi, segno, &p);
372 if (p.min_cost > cost) {
377 if (nsearched >= p.max_search) {
378 if (!sm->last_victim[p.gc_mode] && segno <= last_victim)
379 sm->last_victim[p.gc_mode] = last_victim + 1;
381 sm->last_victim[p.gc_mode] = segno + 1;
382 sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
386 if (p.min_segno != NULL_SEGNO) {
388 if (p.alloc_mode == LFS) {
389 secno = GET_SEC_FROM_SEG(sbi, p.min_segno);
390 if (gc_type == FG_GC)
391 sbi->cur_victim_sec = secno;
393 set_bit(secno, dirty_i->victim_secmap);
395 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
397 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
399 prefree_segments(sbi), free_segments(sbi));
402 mutex_unlock(&dirty_i->seglist_lock);
404 return (p.min_segno == NULL_SEGNO) ? 0 : 1;
407 static const struct victim_selection default_v_ops = {
408 .get_victim = get_victim_by_default,
411 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
413 struct inode_entry *ie;
415 ie = radix_tree_lookup(&gc_list->iroot, ino);
421 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
423 struct inode_entry *new_ie;
425 if (inode == find_gc_inode(gc_list, inode->i_ino)) {
429 new_ie = f2fs_kmem_cache_alloc(f2fs_inode_entry_slab, GFP_NOFS);
430 new_ie->inode = inode;
432 f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
433 list_add_tail(&new_ie->list, &gc_list->ilist);
436 static void put_gc_inode(struct gc_inode_list *gc_list)
438 struct inode_entry *ie, *next_ie;
439 list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
440 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
443 kmem_cache_free(f2fs_inode_entry_slab, ie);
447 static int check_valid_map(struct f2fs_sb_info *sbi,
448 unsigned int segno, int offset)
450 struct sit_info *sit_i = SIT_I(sbi);
451 struct seg_entry *sentry;
454 down_read(&sit_i->sentry_lock);
455 sentry = get_seg_entry(sbi, segno);
456 ret = f2fs_test_bit(offset, sentry->cur_valid_map);
457 up_read(&sit_i->sentry_lock);
462 * This function compares node address got in summary with that in NAT.
463 * On validity, copy that node with cold status, otherwise (invalid node)
466 static void gc_node_segment(struct f2fs_sb_info *sbi,
467 struct f2fs_summary *sum, unsigned int segno, int gc_type)
469 struct f2fs_summary *entry;
473 bool fggc = (gc_type == FG_GC);
475 start_addr = START_BLOCK(sbi, segno);
480 if (fggc && phase == 2)
481 atomic_inc(&sbi->wb_sync_req[NODE]);
483 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
484 nid_t nid = le32_to_cpu(entry->nid);
485 struct page *node_page;
488 /* stop BG_GC if there is not enough free sections. */
489 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
492 if (check_valid_map(sbi, segno, off) == 0)
496 f2fs_ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
502 f2fs_ra_node_page(sbi, nid);
507 node_page = f2fs_get_node_page(sbi, nid);
508 if (IS_ERR(node_page))
511 /* block may become invalid during f2fs_get_node_page */
512 if (check_valid_map(sbi, segno, off) == 0) {
513 f2fs_put_page(node_page, 1);
517 if (f2fs_get_node_info(sbi, nid, &ni)) {
518 f2fs_put_page(node_page, 1);
522 if (ni.blk_addr != start_addr + off) {
523 f2fs_put_page(node_page, 1);
527 f2fs_move_node_page(node_page, gc_type);
528 stat_inc_node_blk_count(sbi, 1, gc_type);
535 atomic_dec(&sbi->wb_sync_req[NODE]);
539 * Calculate start block index indicating the given node offset.
540 * Be careful, caller should give this node offset only indicating direct node
541 * blocks. If any node offsets, which point the other types of node blocks such
542 * as indirect or double indirect node blocks, are given, it must be a caller's
545 block_t f2fs_start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
547 unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
555 } else if (node_ofs <= indirect_blks) {
556 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
557 bidx = node_ofs - 2 - dec;
559 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
560 bidx = node_ofs - 5 - dec;
562 return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
565 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
566 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
568 struct page *node_page;
570 unsigned int ofs_in_node;
571 block_t source_blkaddr;
573 nid = le32_to_cpu(sum->nid);
574 ofs_in_node = le16_to_cpu(sum->ofs_in_node);
576 node_page = f2fs_get_node_page(sbi, nid);
577 if (IS_ERR(node_page))
580 if (f2fs_get_node_info(sbi, nid, dni)) {
581 f2fs_put_page(node_page, 1);
585 if (sum->version != dni->version) {
586 f2fs_msg(sbi->sb, KERN_WARNING,
587 "%s: valid data with mismatched node version.",
589 set_sbi_flag(sbi, SBI_NEED_FSCK);
592 *nofs = ofs_of_node(node_page);
593 source_blkaddr = datablock_addr(NULL, node_page, ofs_in_node);
594 f2fs_put_page(node_page, 1);
596 if (source_blkaddr != blkaddr)
601 static int ra_data_block(struct inode *inode, pgoff_t index)
603 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
604 struct address_space *mapping = inode->i_mapping;
605 struct dnode_of_data dn;
607 struct extent_info ei = {0, 0, 0};
608 struct f2fs_io_info fio = {
615 .encrypted_page = NULL,
621 page = f2fs_grab_cache_page(mapping, index, true);
625 if (f2fs_lookup_extent_cache(inode, index, &ei)) {
626 dn.data_blkaddr = ei.blk + index - ei.fofs;
630 set_new_dnode(&dn, inode, NULL, NULL, 0);
631 err = f2fs_get_dnode_of_data(&dn, index, LOOKUP_NODE);
636 if (unlikely(!f2fs_is_valid_blkaddr(sbi, dn.data_blkaddr,
644 fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
646 fio.encrypted_page = f2fs_pagecache_get_page(META_MAPPING(sbi),
648 FGP_LOCK | FGP_CREAT, GFP_NOFS);
649 if (!fio.encrypted_page) {
654 err = f2fs_submit_page_bio(&fio);
656 goto put_encrypted_page;
657 f2fs_put_page(fio.encrypted_page, 0);
658 f2fs_put_page(page, 1);
661 f2fs_put_page(fio.encrypted_page, 1);
663 f2fs_put_page(page, 1);
668 * Move data block via META_MAPPING while keeping locked data page.
669 * This can be used to move blocks, aka LBAs, directly on disk.
671 static void move_data_block(struct inode *inode, block_t bidx,
672 int gc_type, unsigned int segno, int off)
674 struct f2fs_io_info fio = {
675 .sbi = F2FS_I_SB(inode),
681 .encrypted_page = NULL,
685 struct dnode_of_data dn;
686 struct f2fs_summary sum;
688 struct page *page, *mpage;
691 bool lfs_mode = test_opt(fio.sbi, LFS);
693 /* do not read out */
694 page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
698 if (!check_valid_map(F2FS_I_SB(inode), segno, off))
701 if (f2fs_is_atomic_file(inode)) {
702 F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
703 F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
707 if (f2fs_is_pinned_file(inode)) {
708 f2fs_pin_file_control(inode, true);
712 set_new_dnode(&dn, inode, NULL, NULL, 0);
713 err = f2fs_get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
717 if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
718 ClearPageUptodate(page);
723 * don't cache encrypted data into meta inode until previous dirty
724 * data were writebacked to avoid racing between GC and flush.
726 f2fs_wait_on_page_writeback(page, DATA, true);
728 err = f2fs_get_node_info(fio.sbi, dn.nid, &ni);
732 set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
736 fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
739 down_write(&fio.sbi->io_order_lock);
741 f2fs_allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
742 &sum, CURSEG_COLD_DATA, NULL, false);
744 fio.encrypted_page = f2fs_pagecache_get_page(META_MAPPING(fio.sbi),
745 newaddr, FGP_LOCK | FGP_CREAT, GFP_NOFS);
746 if (!fio.encrypted_page) {
751 mpage = f2fs_pagecache_get_page(META_MAPPING(fio.sbi),
752 fio.old_blkaddr, FGP_LOCK, GFP_NOFS);
754 bool updated = false;
756 if (PageUptodate(mpage)) {
757 memcpy(page_address(fio.encrypted_page),
758 page_address(mpage), PAGE_SIZE);
761 f2fs_put_page(mpage, 1);
762 invalidate_mapping_pages(META_MAPPING(fio.sbi),
763 fio.old_blkaddr, fio.old_blkaddr);
768 err = f2fs_submit_page_bio(&fio);
773 lock_page(fio.encrypted_page);
775 if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
779 if (unlikely(!PageUptodate(fio.encrypted_page))) {
785 set_page_dirty(fio.encrypted_page);
786 f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
787 if (clear_page_dirty_for_io(fio.encrypted_page))
788 dec_page_count(fio.sbi, F2FS_DIRTY_META);
790 set_page_writeback(fio.encrypted_page);
791 ClearPageError(page);
793 /* allocate block address */
794 f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
796 fio.op = REQ_OP_WRITE;
797 fio.op_flags = REQ_SYNC;
798 fio.new_blkaddr = newaddr;
799 f2fs_submit_page_write(&fio);
801 if (PageWriteback(fio.encrypted_page))
802 end_page_writeback(fio.encrypted_page);
806 f2fs_update_iostat(fio.sbi, FS_GC_DATA_IO, F2FS_BLKSIZE);
808 f2fs_update_data_blkaddr(&dn, newaddr);
809 set_inode_flag(inode, FI_APPEND_WRITE);
810 if (page->index == 0)
811 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
813 f2fs_put_page(fio.encrypted_page, 1);
816 up_write(&fio.sbi->io_order_lock);
818 f2fs_do_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
823 f2fs_put_page(page, 1);
826 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
827 unsigned int segno, int off)
831 page = f2fs_get_lock_data_page(inode, bidx, true);
835 if (!check_valid_map(F2FS_I_SB(inode), segno, off))
838 if (f2fs_is_atomic_file(inode)) {
839 F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
840 F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
843 if (f2fs_is_pinned_file(inode)) {
844 if (gc_type == FG_GC)
845 f2fs_pin_file_control(inode, true);
849 if (gc_type == BG_GC) {
850 if (PageWriteback(page))
852 set_page_dirty(page);
855 struct f2fs_io_info fio = {
856 .sbi = F2FS_I_SB(inode),
861 .op_flags = REQ_SYNC,
862 .old_blkaddr = NULL_ADDR,
864 .encrypted_page = NULL,
865 .need_lock = LOCK_REQ,
866 .io_type = FS_GC_DATA_IO,
868 bool is_dirty = PageDirty(page);
872 set_page_dirty(page);
873 f2fs_wait_on_page_writeback(page, DATA, true);
874 if (clear_page_dirty_for_io(page)) {
875 inode_dec_dirty_pages(inode);
876 f2fs_remove_dirty_inode(inode);
881 err = f2fs_do_write_data_page(&fio);
883 clear_cold_data(page);
884 if (err == -ENOMEM) {
885 congestion_wait(BLK_RW_ASYNC, HZ/50);
889 set_page_dirty(page);
893 f2fs_put_page(page, 1);
897 * This function tries to get parent node of victim data block, and identifies
898 * data block validity. If the block is valid, copy that with cold status and
899 * modify parent node.
900 * If the parent node is not valid or the data block address is different,
901 * the victim data block is ignored.
903 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
904 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
906 struct super_block *sb = sbi->sb;
907 struct f2fs_summary *entry;
912 start_addr = START_BLOCK(sbi, segno);
917 for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
918 struct page *data_page;
920 struct node_info dni; /* dnode info for the data */
921 unsigned int ofs_in_node, nofs;
923 nid_t nid = le32_to_cpu(entry->nid);
925 /* stop BG_GC if there is not enough free sections. */
926 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
929 if (check_valid_map(sbi, segno, off) == 0)
933 f2fs_ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
939 f2fs_ra_node_page(sbi, nid);
943 /* Get an inode by ino with checking validity */
944 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
948 f2fs_ra_node_page(sbi, dni.ino);
952 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
955 inode = f2fs_iget(sb, dni.ino);
956 if (IS_ERR(inode) || is_bad_inode(inode))
959 if (!down_write_trylock(
960 &F2FS_I(inode)->i_gc_rwsem[WRITE])) {
962 sbi->skipped_gc_rwsem++;
966 start_bidx = f2fs_start_bidx_of_node(nofs, inode) +
969 if (f2fs_post_read_required(inode)) {
970 int err = ra_data_block(inode, start_bidx);
972 up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
977 add_gc_inode(gc_list, inode);
981 data_page = f2fs_get_read_data_page(inode,
982 start_bidx, REQ_RAHEAD, true);
983 up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
984 if (IS_ERR(data_page)) {
989 f2fs_put_page(data_page, 0);
990 add_gc_inode(gc_list, inode);
995 inode = find_gc_inode(gc_list, dni.ino);
997 struct f2fs_inode_info *fi = F2FS_I(inode);
1000 if (S_ISREG(inode->i_mode)) {
1001 if (!down_write_trylock(&fi->i_gc_rwsem[READ])) {
1002 sbi->skipped_gc_rwsem++;
1005 if (!down_write_trylock(
1006 &fi->i_gc_rwsem[WRITE])) {
1007 sbi->skipped_gc_rwsem++;
1008 up_write(&fi->i_gc_rwsem[READ]);
1013 /* wait for all inflight aio data */
1014 inode_dio_wait(inode);
1017 start_bidx = f2fs_start_bidx_of_node(nofs, inode)
1019 if (f2fs_post_read_required(inode))
1020 move_data_block(inode, start_bidx, gc_type,
1023 move_data_page(inode, start_bidx, gc_type,
1027 up_write(&fi->i_gc_rwsem[WRITE]);
1028 up_write(&fi->i_gc_rwsem[READ]);
1031 stat_inc_data_blk_count(sbi, 1, gc_type);
1039 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
1042 struct sit_info *sit_i = SIT_I(sbi);
1045 down_write(&sit_i->sentry_lock);
1046 ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
1047 NO_CHECK_TYPE, LFS);
1048 up_write(&sit_i->sentry_lock);
1052 static int do_garbage_collect(struct f2fs_sb_info *sbi,
1053 unsigned int start_segno,
1054 struct gc_inode_list *gc_list, int gc_type)
1056 struct page *sum_page;
1057 struct f2fs_summary_block *sum;
1058 struct blk_plug plug;
1059 unsigned int segno = start_segno;
1060 unsigned int end_segno = start_segno + sbi->segs_per_sec;
1062 unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
1063 SUM_TYPE_DATA : SUM_TYPE_NODE;
1065 /* readahead multi ssa blocks those have contiguous address */
1066 if (sbi->segs_per_sec > 1)
1067 f2fs_ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
1068 sbi->segs_per_sec, META_SSA, true);
1070 /* reference all summary page */
1071 while (segno < end_segno) {
1072 sum_page = f2fs_get_sum_page(sbi, segno++);
1073 unlock_page(sum_page);
1076 blk_start_plug(&plug);
1078 for (segno = start_segno; segno < end_segno; segno++) {
1080 /* find segment summary of victim */
1081 sum_page = find_get_page(META_MAPPING(sbi),
1082 GET_SUM_BLOCK(sbi, segno));
1083 f2fs_put_page(sum_page, 0);
1085 if (get_valid_blocks(sbi, segno, false) == 0)
1087 if (!PageUptodate(sum_page) || unlikely(f2fs_cp_error(sbi)))
1090 sum = page_address(sum_page);
1091 if (type != GET_SUM_TYPE((&sum->footer))) {
1092 f2fs_msg(sbi->sb, KERN_ERR, "Inconsistent segment (%u) "
1093 "type [%d, %d] in SSA and SIT",
1094 segno, type, GET_SUM_TYPE((&sum->footer)));
1095 set_sbi_flag(sbi, SBI_NEED_FSCK);
1100 * this is to avoid deadlock:
1101 * - lock_page(sum_page) - f2fs_replace_block
1102 * - check_valid_map() - down_write(sentry_lock)
1103 * - down_read(sentry_lock) - change_curseg()
1104 * - lock_page(sum_page)
1106 if (type == SUM_TYPE_NODE)
1107 gc_node_segment(sbi, sum->entries, segno, gc_type);
1109 gc_data_segment(sbi, sum->entries, gc_list, segno,
1112 stat_inc_seg_count(sbi, type, gc_type);
1115 if (gc_type == FG_GC &&
1116 get_valid_blocks(sbi, segno, false) == 0)
1119 f2fs_put_page(sum_page, 0);
1122 if (gc_type == FG_GC)
1123 f2fs_submit_merged_write(sbi,
1124 (type == SUM_TYPE_NODE) ? NODE : DATA);
1126 blk_finish_plug(&plug);
1128 stat_inc_call_count(sbi->stat_info);
1133 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
1134 bool background, unsigned int segno)
1136 int gc_type = sync ? FG_GC : BG_GC;
1137 int sec_freed = 0, seg_freed = 0, total_freed = 0;
1139 struct cp_control cpc;
1140 unsigned int init_segno = segno;
1141 struct gc_inode_list gc_list = {
1142 .ilist = LIST_HEAD_INIT(gc_list.ilist),
1143 .iroot = RADIX_TREE_INIT(gc_list.iroot, GFP_NOFS),
1145 unsigned long long last_skipped = sbi->skipped_atomic_files[FG_GC];
1146 unsigned long long first_skipped;
1147 unsigned int skipped_round = 0, round = 0;
1149 trace_f2fs_gc_begin(sbi->sb, sync, background,
1150 get_pages(sbi, F2FS_DIRTY_NODES),
1151 get_pages(sbi, F2FS_DIRTY_DENTS),
1152 get_pages(sbi, F2FS_DIRTY_IMETA),
1155 reserved_segments(sbi),
1156 prefree_segments(sbi));
1158 cpc.reason = __get_cp_reason(sbi);
1159 sbi->skipped_gc_rwsem = 0;
1160 first_skipped = last_skipped;
1162 if (unlikely(!(sbi->sb->s_flags & SB_ACTIVE))) {
1166 if (unlikely(f2fs_cp_error(sbi))) {
1171 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
1173 * For example, if there are many prefree_segments below given
1174 * threshold, we can make them free by checkpoint. Then, we
1175 * secure free segments which doesn't need fggc any more.
1177 if (prefree_segments(sbi)) {
1178 ret = f2fs_write_checkpoint(sbi, &cpc);
1182 if (has_not_enough_free_secs(sbi, 0, 0))
1186 /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
1187 if (gc_type == BG_GC && !background) {
1191 if (!__get_victim(sbi, &segno, gc_type)) {
1196 seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type);
1197 if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec)
1199 total_freed += seg_freed;
1201 if (gc_type == FG_GC) {
1202 if (sbi->skipped_atomic_files[FG_GC] > last_skipped ||
1203 sbi->skipped_gc_rwsem)
1205 last_skipped = sbi->skipped_atomic_files[FG_GC];
1209 if (gc_type == FG_GC)
1210 sbi->cur_victim_sec = NULL_SEGNO;
1215 if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
1216 if (skipped_round <= MAX_SKIP_GC_COUNT ||
1217 skipped_round * 2 < round) {
1222 if (first_skipped < last_skipped &&
1223 (last_skipped - first_skipped) >
1224 sbi->skipped_gc_rwsem) {
1225 f2fs_drop_inmem_pages_all(sbi, true);
1229 if (gc_type == FG_GC)
1230 ret = f2fs_write_checkpoint(sbi, &cpc);
1233 SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
1234 SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno;
1236 trace_f2fs_gc_end(sbi->sb, ret, total_freed, sec_freed,
1237 get_pages(sbi, F2FS_DIRTY_NODES),
1238 get_pages(sbi, F2FS_DIRTY_DENTS),
1239 get_pages(sbi, F2FS_DIRTY_IMETA),
1242 reserved_segments(sbi),
1243 prefree_segments(sbi));
1245 mutex_unlock(&sbi->gc_mutex);
1247 put_gc_inode(&gc_list);
1250 ret = sec_freed ? 0 : -EAGAIN;
1254 void f2fs_build_gc_manager(struct f2fs_sb_info *sbi)
1256 DIRTY_I(sbi)->v_ops = &default_v_ops;
1258 sbi->gc_pin_file_threshold = DEF_GC_FAILED_PINNED_FILES;
1260 /* give warm/cold data area from slower device */
1261 if (f2fs_is_multi_device(sbi) && sbi->segs_per_sec == 1)
1262 SIT_I(sbi)->last_victim[ALLOC_NEXT] =
1263 GET_SEGNO(sbi, FDEV(0).end_blk) + 1;