Re: [f2fs-dev] [PATCH] f2fs: avoid race condition for shinker count

[Chao Yu <yuchao0@huawei.com>]

On 2020/11/10 1:00, Jaegeuk Kim wrote:
> Light reported sometimes shinker gets nat_cnt < dirty_nat_cnt resulting in

I didn't get the problem clearly, did you mean __count_nat_entries() will
give the wrong shrink count due to race condition? should there be a lock
while reading these two variables?

> wrong do_shinker work. Basically the two counts should not happen like that.
> 
> So, I suspect this race condtion where:
> - f2fs_try_to_free_nats            __flush_nat_entry_set
>   nat_cnt=2, dirty_nat_cnt=2
>                                     __clear_nat_cache_dirty
>                                      spin_lock(nat_list_lock)
>                                      list_move()
>                                      spin_unlock(nat_list_lock)
>   spin_lock(nat_list_lock)
>   list_del()
>   spin_unlock(nat_list_lock)
>   nat_cnt=1, dirty_nat_cnt=2
>                                     nat_cnt=1, dirty_nat_cnt=1

nm_i->nat_cnt and nm_i->dirty_nat_cnt were protected by
nm_i->nat_tree_lock, I didn't see why expanding nat_list_lock range
will help... since there are still places nat_list_lock() didn't
cover these two reference counts.

Thanks,

> 
> Reported-by: Light Hsieh <Light.Hsieh@mediatek.com>
> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> ---
>   fs/f2fs/node.c | 3 +--
>   1 file changed, 1 insertion(+), 2 deletions(-)
> 
> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
> index 42394de6c7eb..e8ec65e40f06 100644
> --- a/fs/f2fs/node.c
> +++ b/fs/f2fs/node.c
> @@ -269,11 +269,10 @@ static void __clear_nat_cache_dirty(struct f2fs_nm_info *nm_i,
>   {
>   	spin_lock(&nm_i->nat_list_lock);
>   	list_move_tail(&ne->list, &nm_i->nat_entries);
> -	spin_unlock(&nm_i->nat_list_lock);
> -
>   	set_nat_flag(ne, IS_DIRTY, false);
>   	set->entry_cnt--;
>   	nm_i->dirty_nat_cnt--;
> +	spin_unlock(&nm_i->nat_list_lock);
>   }
>   
>   static unsigned int __gang_lookup_nat_set(struct f2fs_nm_info *nm_i,
>