Re: [f2fs-dev] [PATCH v3] f2fs: avoid race condition for shinker count

[Jaegeuk Kim]

On 12/03, Chao Yu wrote:
> On 2020/12/3 14:07, Jaegeuk Kim wrote:
> > On 11/11, Jaegeuk Kim wrote:
> > > Light reported sometimes shinker gets nat_cnt < dirty_nat_cnt resulting in
> > > wrong do_shinker work. Let's avoid to get stale data by using nat_tree_lock.
> > > 
> > > Reported-by: Light Hsieh <Light.Hsieh@xxxxxxxxxxxx>
> > > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > ---
> > > v3:
> > >   - fix to use NM_I(sbi)
> > > 
> > >   fs/f2fs/shrinker.c | 6 +++++-
> > >   1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/f2fs/shrinker.c b/fs/f2fs/shrinker.c
> > > index d66de5999a26..555712ba06d8 100644
> > > --- a/fs/f2fs/shrinker.c
> > > +++ b/fs/f2fs/shrinker.c
> > > @@ -18,7 +18,11 @@ static unsigned int shrinker_run_no;
> > >   static unsigned long __count_nat_entries(struct f2fs_sb_info *sbi)
> > >   {
> > > -	long count = NM_I(sbi)->nat_cnt - NM_I(sbi)->dirty_nat_cnt;
> > > +	long count;
> > > +
> > > +	down_read(&NM_I(sbi)->nat_tree_lock);
> > > +	count = NM_I(sbi)->nat_cnt - NM_I(sbi)->dirty_nat_cnt;
> > > +	up_read(&NM_I(sbi)->nat_tree_lock);
> > 
> > I just fosund this can give kernel hang due to the following backtrace.
> > f2fs_shrink_count
> > shrink_slab
> > shrink_node
> > do_try_to_free_pages
> > try_to_free_pages
> > __alloc_pages_nodemask
> > alloc_pages_current
> > allocate_slab
> > __slab_alloc
> > __slab_alloc
> > kmem_cache_alloc
> > add_free_nid
> > f2fs_flush_nat_entries
> > f2fs_write_checkpoint
> 
> Oh, I missed that case.
> 
> > 
> > Let me just check like this.
> > 
> > > From 971163330224449d90aac90957ea38f77d494f0f Mon Sep 17 00:00:00 2001
> > From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > Date: Fri, 6 Nov 2020 13:22:05 -0800
> > Subject: [PATCH] f2fs: avoid race condition for shrinker count
> > 
> > Light reported sometimes shinker gets nat_cnt < dirty_nat_cnt resulting in
> > wrong do_shinker work. Let's avoid to return insane overflowed value.
> > 
> > Reported-by: Light Hsieh <Light.Hsieh@xxxxxxxxxxxx>
> > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > ---
> >   fs/f2fs/shrinker.c | 6 +++---
> >   1 file changed, 3 insertions(+), 3 deletions(-)
> > 
> > diff --git a/fs/f2fs/shrinker.c b/fs/f2fs/shrinker.c
> > index d66de5999a26..75b5b4aaed99 100644
> > --- a/fs/f2fs/shrinker.c
> > +++ b/fs/f2fs/shrinker.c
> > @@ -18,9 +18,9 @@ static unsigned int shrinker_run_no;
> >   static unsigned long __count_nat_entries(struct f2fs_sb_info *sbi)
> >   {
> > -	long count = NM_I(sbi)->nat_cnt - NM_I(sbi)->dirty_nat_cnt;
> > -
> > -	return count > 0 ? count : 0;
> > +	if (NM_I(sbi)->nat_cnt > NM_I(sbi)->dirty_nat_cnt)
> > +		return NM_I(sbi)->nat_cnt - NM_I(sbi)->dirty_nat_cnt;
> 
> Hmm.. in the case that we are not in checkpoint progress, nat_cnt and dirty_nat_cnt
> is mutable, how can we make sure the calculation is non-negative after the check
> condition? :(

How about adding another variable to monitor it?