Re: [PATCH] mm: fix race between MADV_FREE reclaim and blkdev direct IO read

[Mauricio Faria de Oliveira <mfo@canonical.com>]

On Thu, Dec 16, 2021 at 11:10 PM Huang, Ying <ying.huang@intel.com> wrote:
>
> Minchan Kim <minchan@kernel.org> writes:
...
> >
> > A madv_free marked page could be mapped at several processes so
> > it wouldn't be refcount two all the time, I think.
> > Shouldn't we check it with page_mapcount with page_refcount?
> >
> >     page_ref_count(page) - 1  > page_mapcount(page)
> >
>
> And should we consider page_count() too in madvise_free_pte_range()?
> That is, if the page has been used by GUP, we needn't to make its PTE
> clean?

Hey, thanks for reviewing!

That might not be sufficient time-wise, I guess, because the page can
be used by GUP after the madvise() call (e.g., this case), thus
checking for it during the call wouldn't catch it -- this may apply to
other cases too, where there's no guarantee/ordering between both
operations.

cheers,

-- 
Mauricio Faria de Oliveira