Re: [PATCH] mm: swap: prevent possible data-race in __try_to_reclaim_swap

[Jeongjun Park]

Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Fri, Oct 04, 2024 at 11:25:04PM +0900, Jeongjun Park wrote:
> > A report [1] was uploaded from syzbot.
> >
> > In the previous commit 862590ac3708 ("mm: swap: allow cache reclaim to skip
> > slot cache"), the __try_to_reclaim_swap() function reads offset and nr_pages
> > from folio without folio_lock protection.
>
> Umm.  You don't need folio_lock to read nr_pages.  Holding a refcount
> is sufficient to stabilise nr_pages.  I cannot speak to folio->swap
> though (and the KCSAN report does appear to be pointing to folio->swap).
>

That's right. It looks like KCSAN log occurs when reading folio->swap.
In fact, since most of the code reads folio->swap under the protection
of folio_lock, it is possible to modify only the part that reads folio->swap
and the code that reads offset to operate under the protection of
folio_lock.

However, even if reading nr_pages does not require folio_lock, I don't
think it is very desirable to modify only this code to not be protected
by folio_lock.

Regards,
Jeongjun Park