Re: [PATCH 3/4] mm: unreserve highatomic free pages fully before OOM

[Minchan Kim <minchan@kernel.org>]

Hi Michal,

On Mon, Oct 10, 2016 at 09:41:40AM +0200, Michal Hocko wrote:
> On Fri 07-10-16 23:43:45, Minchan Kim wrote:
> > On Fri, Oct 07, 2016 at 11:09:17AM +0200, Michal Hocko wrote:
> > > On Fri 07-10-16 14:45:35, Minchan Kim wrote:
> > > > After fixing the race of highatomic page count, I still encounter
> > > > OOM with many free memory reserved as highatomic.
> > > > 
> > > > One of reason in my testing was we unreserve free pages only if
> > > > reclaim has progress. Otherwise, we cannot have chance to unreseve.
> > > > 
> > > > Other problem after fixing it was it doesn't guarantee every pages
> > > > unreserving of highatomic pageblock because it just release *a*
> > > > pageblock which could have few free pages so other context could
> > > > steal it easily so that the process stucked with direct reclaim
> > > > finally can encounter OOM although there are free pages which can
> > > > be unreserved.
> > > > 
> > > > This patch changes the logic so that it unreserves pageblocks with
> > > > no_progress_loop proportionally. IOW, in first retrial of reclaim,
> > > > it will try to unreserve a pageblock. In second retrial of reclaim,
> > > > it will try to unreserve 1/MAX_RECLAIM_RETRIES * reserved_pageblock
> > > > and finally all reserved pageblock before the OOM.
> > > > 
> > > > Signed-off-by: Minchan Kim <minchan@kernel.org>
> > > > ---
> > > >  mm/page_alloc.c | 57 ++++++++++++++++++++++++++++++++++++++++++++-------------
> > > >  1 file changed, 44 insertions(+), 13 deletions(-)
> > > 
> > > This sounds much more complex then it needs to be IMHO. Why something as
> > > simple as thhe following wouldn't work? Please note that I even didn't
> > > try to compile this. It is just give you an idea.
> > > ---
> > >  mm/page_alloc.c | 26 ++++++++++++++++++++------
> > >  1 file changed, 20 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 73f60ad6315f..e575a4f38555 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -2056,7 +2056,8 @@ static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
> > >   * intense memory pressure but failed atomic allocations should be easier
> > >   * to recover from than an OOM.
> > >   */
> > > -static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
> > > +static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
> > > +		bool force)
> > >  {
> > >  	struct zonelist *zonelist = ac->zonelist;
> > >  	unsigned long flags;
> > > @@ -2067,8 +2068,14 @@ static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
> > >  
> > >  	for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
> > >  								ac->nodemask) {
> > > -		/* Preserve at least one pageblock */
> > > -		if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
> > > +		if (!zone->nr_reserved_highatomic)
> > > +			continue;
> > > +
> > > +		/*
> > > +		 * Preserve at least one pageblock unless we are really running
> > > +		 * out of memory
> > > +		 */
> > > +		if (!force && zone->nr_reserved_highatomic <= pageblock_nr_pages)
> > >  			continue;
> > >  
> > >  		spin_lock_irqsave(&zone->lock, flags);
> > > @@ -2102,10 +2109,12 @@ static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
> > >  			set_pageblock_migratetype(page, ac->migratetype);
> > >  			move_freepages_block(zone, page, ac->migratetype);
> > >  			spin_unlock_irqrestore(&zone->lock, flags);
> > > -			return;
> > > +			return true;
> > 
> > Such cut-off makes reserved pageblock remained before the OOM.
> > We call it as premature OOM kill.
> 
> Not sure I understand. The above should get rid of all atomic reserves
> before we go OOM. We can do it all at once but that sounds too

The problem is there is race between page freeing path and unreserve
logic so that some pages could be in highatomic free list even though
zone->nr_reserved_highatomic is already zero.
So, at least, it would be better to have a draining step at some point
where was (no_progress_loops == MAX_RECLAIM RETRIES) in my patch.

Also, your patch makes retry loop greater than MAX_RECLAIM_RETRIES
if unreserve_highatomic_pageblock returns true. Theoretically,
it would make live lock. You might argue it's *really really* rare
but I don't want to add such subtle thing.
Maybe, we could drain when no_progress_loops == MAX_RECLAIM_RETRIES.

> aggressive to me. If we just do one at the time we have a chance to
> keep some reserves if the OOM situation is really ephemeral.
> 
> Does this patch work in your usecase?

I didn't test but I guess it works but it has problems I mentioned
above.