Re: [PATCH] mm: fix unevictable page reclaim when calling madvise_pageout

[Michal Hocko <mhocko@kernel.org>]

On Wed 30-10-19 15:33:07, Johannes Weiner wrote:
> On Wed, Oct 30, 2019 at 06:45:33PM +0100, Michal Hocko wrote:
> > On Wed 30-10-19 09:52:39, Minchan Kim wrote:
[...]
> > > madvise_pageout could work with a shared page and one of the vmas among processes
> > > could do mlock so it could pass Unevictable LRU pages into shrink_page_list.
> > > It's pointless to try reclaim unevictable pages from the beginning so I want to fix
> > > madvise_pageout via introducing only_evictable flag into the API so that
> > > madvise_pageout uses it as "true".
> > > 
> > > If we want to remove the PageUnevictable VM_BUG_ON_PAGE in shrink_page_list,
> > > I want to see more strong reason why it happens and why caller couldn't
> > > filter them out from the beginning.
> > 
> > Why is this preferable over removing the VM_BUG_ON condition? In other
> > words why should we keep PageUnevictable check there?
> 
> The mlock LRU shuffling is a bit tricky and can race with page reclaim
> or others isolating the page from the LRU list. If another isolator
> wins, it has to move the page during putback on behalf of mlock.
> 
> See the implementation and comments in __pagevec_lru_add_fn().
> 
> That's why page reclaim can see !page_evictable(), but it must not see
> pages that have the PageUnevictable lru bit already set. Because that
> would mean the isolation/putback machinery messed up somewhere and the
> page LRU state is corrupt.
> 
> As that machinery is non-trivial, it's useful to have that sanity
> check in page reclaim.

Thanks for the clarification! This sounds reasonable (as much as the
mlock juggling does) to me. This is probably worth a comment right above
the bug_on.

I have to confess that I am still not clear on all the details here,
though. E.g. migrate_misplaced_transhuge_page sets the flag without
lru_lock and relies only on page lock IIUC and the bug on is done right
after the lock is released. Maybe I am just confused or maybe the race
window is too small to matter but isn't this race possible at least
theoretically?

-- 
Michal Hocko
SUSE Labs