Re: [PATCH v3] mm: fix trying to reclaim unevictable lru page when calling madvise_pageout

[Johannes Weiner <hannes@cmpxchg.org>]

On Tue, Nov 05, 2019 at 10:10:37PM +0800, zhong jiang wrote:
> On 2019/11/5 20:45, Michal Hocko wrote:
> > On Tue 05-11-19 20:28:58, zhong jiang wrote:
> >> On 2019/11/5 14:33, Michal Hocko wrote:
> >>> On Sat 02-11-19 15:36:55, zhong jiang wrote:
> >>>> On 2019/11/2 2:32, Michal Hocko wrote:
> >>> [...]
> >>>>> But I would really appreciate to add a comment for the BUG_ON and
> >>>>> explain why do we care about PageUnevictable so much when there is an
> >>>>> explicit page_evictable check in the reclaim path. In other words a
> >>>>> short summary of what Johannes explained in
> >>>>> http://lkml.kernel.org/r/20191030193307.GA48128@cmpxchg.org. Maybe in a
> >>>>> separate patch. Care to send one or should I send it?
> >>>> Hi,  Michal
> >>>>
> >>>> Actually,  I am not very clear about the words Johannes had said.   How the race to
> >>>> tirgger, it will result in an PgeMlocked page can be visible in shrink_page_list.
> >>>>
> >>>> Could you elaborate the race in detail further ?
> >>> I would go with the following comment
> >>>
> >>> 	/*
> >>> 	 * Page reclaim can see !page_evictable(), but it must not see pages that
> >>> 	 * have the PageUnevictable lru bit already set. See  __pagevec_lru_add_fn()
> >>> 	 * for more details.
> >>> 	 */
> >> But  the detail still confuses me in __pagevec_lru_add_fn()  to  see PageMlocked  in vmscan :-\ .
> > Which part does confuse you exactly?
> page reclaim can see !page_evictable()  means some race still exist in the kernel. Is there any race window .

Yes. mlock does this:

lock_page()
SetPageMlocked()
if (isolate_lru_page())
  putback_lru_page() // move to unevictable list
unlock_page()

and vmscan does this:

isolate_lru_pages()
for_each_page()
  if (!try_lock_page())
    continue
  if (!page_evictable())
    continue
putback_lru_pages()

It's possible that mlock locks the page and sets PG_mlocked, but
vmscan has the page already isolated and mlock cannot move it to the
unevictable list itself. In that case, vmscan will either fail to lock
the page or see !page_unevictable() and move the page on the
unevictable list on behalf of mlock.