Re: [PATCH 12/14] mm, oom: protect !costly allocations some more

[Michal Hocko]

On Wed 04-05-16 23:57:50, Joonsoo Kim wrote:
> 2016-05-04 17:56 GMT+09:00 Michal Hocko <mhocko@xxxxxxxxxx>:
> > On Wed 04-05-16 15:31:12, Joonsoo Kim wrote:
> >> On Wed, May 04, 2016 at 03:01:24PM +0900, Joonsoo Kim wrote:
> >> > On Wed, Apr 20, 2016 at 03:47:25PM -0400, Michal Hocko wrote:
> > [...]
> >> > > @@ -3408,6 +3456,17 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> >> > >                            no_progress_loops))
> >> > >           goto retry;
> >> > >
> >> > > + /*
> >> > > +  * It doesn't make any sense to retry for the compaction if the order-0
> >> > > +  * reclaim is not able to make any progress because the current
> >> > > +  * implementation of the compaction depends on the sufficient amount
> >> > > +  * of free memory (see __compaction_suitable)
> >> > > +  */
> >> > > + if (did_some_progress > 0 &&
> >> > > +                 should_compact_retry(order, compact_result,
> >> > > +                         &migration_mode, compaction_retries))
> >> >
> >> > Checking did_some_progress on each round have subtle corner case. Think
> >> > about following situation.
> >> >
> >> > round, compaction, did_some_progress, compaction
> >> > 0, defer, 1
> >> > 0, defer, 1
> >> > 0, defer, 1
> >> > 0, defer, 1
> >> > 0, defer, 0
> >>
> >> Oops...Example should be below one.
> >>
> >> 0, defer, 1
> >> 1, defer, 1
> >> 2, defer, 1
> >> 3, defer, 1
> >> 4, defer, 0
> >
> > I am not sure I understand. The point of the check is that if the
> > reclaim doesn't make _any_ progress then checking the result of the
> > compaction after it didn't lead to a successful allocation just doesn't
> > make any sense.
> 
> Even if this round (#4) doesn't reclaim any pages, previous rounds
> (#0, #1, #2, #3) would reclaim enough pages to succeed future
> compaction attempt.

Then the compaction shouldn't back off and I would consider it a
compaction bug. I haven't see this happening though. Vlastimil is
already working on patches which would simply guarantee that really
important allocations will not defer.

So unless I can see an example of a real issue with this I think it is
just a theoretical issue which shouldn't block the patch as is.
-- 
Michal Hocko
SUSE Labs