Skip to content

Commit

Permalink
f2fs: prevent the current section from being selected as a victim dur…
Browse files Browse the repository at this point in the history
…ing GC

When selecting a victim using next_victim_seg in a large section, the
selected section might already have been cleared and designated as the
new current section, making it actively in use.
This behavior causes inconsistency between the SIT and SSA.

F2FS-fs (dm-54): Inconsistent segment (70961) type [0, 1] in SSA and SIT
Call trace:
dump_backtrace+0xe8/0x10c
show_stack+0x18/0x28
dump_stack_lvl+0x50/0x6c
dump_stack+0x18/0x28
f2fs_stop_checkpoint+0x1c/0x3c
do_garbage_collect+0x41c/0x271c
f2fs_gc+0x27c/0x828
gc_thread_func+0x290/0x88c
kthread+0x11c/0x164
ret_from_fork+0x10/0x20

issue scenario
segs_per_sec=2
- seg#0 and seg#1 are all dirty
- all valid blocks are removed in seg#1
- gc select this sec and next_victim_seg=seg#0
- migrate seg#0, next_victim_seg=seg#1
- checkpoint -> sec(seg#0, seg#1)  becomes free
- allocator assigns sec(seg#0, seg#1) to curseg
- gc tries to migrate seg#1

Fixes: e3080b0 ("f2fs: support subsectional garbage collection")
Signed-off-by: yohan.joung <yohan.joung@sk.com>
Signed-off-by: Chao Yu <chao@kernel.org>
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
  • Loading branch information
yohan.joung authored and Jaegeuk Kim committed Apr 12, 2025
1 parent 05d3273 commit d26fecb
Showing 1 changed file with 5 additions and 0 deletions.
5 changes: 5 additions & 0 deletions fs/f2fs/segment.h
Original file line number Diff line number Diff line change
Expand Up @@ -486,6 +486,11 @@ static inline void __set_test_and_free(struct f2fs_sb_info *sbi,

free_i->free_sections++;

if (GET_SEC_FROM_SEG(sbi, sbi->next_victim_seg[BG_GC]) == secno)
sbi->next_victim_seg[BG_GC] = NULL_SEGNO;
if (GET_SEC_FROM_SEG(sbi, sbi->next_victim_seg[FG_GC]) == secno)
sbi->next_victim_seg[FG_GC] = NULL_SEGNO;

unlock_out:
spin_unlock(&free_i->segmap_lock);
}
Expand Down

0 comments on commit d26fecb

Please sign in to comment.