Re: [PATCH] thp: use is_zero_pfn after pte_present check

[Minchan Kim]

Hello,

On Mon, Oct 12, 2015 at 01:13:20PM +0300, Kirill A. Shutemov wrote:
> On Mon, Oct 12, 2015 at 10:54:16AM +0900, Minchan Kim wrote:
> > Use is_zero_pfn on pteval only after pte_present check on pteval
> > (It might be better idea to introduce is_zero_pte where checks
> > pte_present first). Otherwise, it could work with swap or
> > migration entry and if pte_pfn's result is equal to zero_pfn
> > by chance, we lose user's data in __collapse_huge_page_copy.
> > So if you're luck, the application is segfaulted and finally you
> > could see below message when the application is exit.
> > 
> > BUG: Bad rss-counter state mm:ffff88007f099300 idx:2 val:3
> 
> Did you acctually steped on the bug?
> If yes it's subject for stable@, I think.

Yes, I did with my testing program which made heavy swap-in/out/
swapoff with MADV_DONTNEED in a memcg.
Actually, I marked this patch as -stable but removed it right before
sending because my test program is artificial and didn't see any
report about rss bad counting with MM_SWAPENTS in linux-mm(Of course,
I might miss it).
In addition, sometime I saw someone insists on "It's not a stable
material if it's not a bug with real workload". I don't want to
involve such non-technical stuff so waited someone nudges me to
mark it as -stable and finally, you did. ;-)
If other reviewers are not against, I will Cc -stable in next spin.

> 
> > Signed-off-by: Minchan Kim <minchan@xxxxxxxxxx>
> > ---
> > 
> > I found this bug with MADV_FREE hard test. Sometime, I saw
> > "Bad rss-counter" message with MM_SWAPENTS but it's really
> > rare, once a day if I was luck or once in five days if I was
> > unlucky so I am doing test still and just pass a few days but
> > I hope it will fix the issue.
> > 
> >  mm/huge_memory.c | 12 +++++++++++-
> >  1 file changed, 11 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 4b06b8db9df2..349590aa4533 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -2665,15 +2665,25 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> >  	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
> >  	     _pte++, _address += PAGE_SIZE) {
> >  		pte_t pteval = *_pte;
> > -		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
> > +		if (pte_none(pteval)) {
> 
> In -mm tree we have is_swap_pte() check before this point in
> khugepaged_scan_pmd()

Actually, I tested this patch with v4.2 kernel so it doesn't have
the check.
Now, I look through optimistic check for swapin readahead patch
in current mmotm.
It seems the check couldn't prevent this problem because it releases
pte lock and anon_vma lock before being isolated the page in
__collapse_huge_page_isolate so the page could be swapped out again.

> 
> Also, what about similar pattern in __collapse_huge_page_isolate() and
> __collapse_huge_page_copy()? Shouldn't they be fixed as well?

I see what's wrong here.
/me slaps self.
The line I was about to change was in __collapse_huge_page_isolate
but I changed khugepaged_scan_pmd by mistake at last modification
since that part is almost same. :(
Fortunately my testing kernel is doing right version.
Here it goes.