Re: [patch for-5.3 0/4] revert immediate fallback to remote hugepages

[Michal Hocko]

On Thu 03-10-19 10:00:08, Vlastimil Babka wrote:
> On 10/3/19 12:32 AM, David Rientjes wrote:
> > On Wed, 2 Oct 2019, Michal Hocko wrote:
> > 
> >>>> If 
> >>>> hugetlb wants to stress this to the fullest extent possible, it already 
> >>>> appropriately uses __GFP_RETRY_MAYFAIL.
> >>>
> >>> Which doesn't work anymore right now, and should again after this patch.
> >>
> >> I didn't get to fully digest the patch Vlastimil is proposing. (Ab)using
> >> __GFP_NORETRY is quite subtle but it is already in place with some
> >> explanation and a reference to THPs. So while I am not really happy it
> >> is at least something you can reason about.
> >>
> > 
> > It's a no-op:
> > 
> >         /* Do not loop if specifically requested */
> >         if (gfp_mask & __GFP_NORETRY)
> >                 goto nopage;
> > 
> >         /*
> >          * Do not retry costly high order allocations unless they are
> >          * __GFP_RETRY_MAYFAIL
> >          */
> >         if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
> >                 goto nopage;
> > 
> > So I'm not sure we should spend too much time discussing a hunk of a patch 
> > that doesn't do anything.
> 
> I believe Michal was talking about my (ab)use of __GFP_NORETRY, where it
> controls the earlier 'goto nopage' condition.

That is correct. From a maintainability point of view it would be better
to have only a single bailout of an optimistic compaction attempt. If we
go with [1] then we have two different criterion to bail out and
that is really messy and error prone. While sticking __GFP_RETRY_MAYFAIL
as suggest in [1] fixes up the immediate regression in the simplest way
this all really begs for a proper analysis and a _real_ fix. Can we move
that direction finally, please?

I would really love to conduct further testing but I haven't really
heard anything to results presented so far. I have no idea whether
that is even remotely resembling anything David needs for his claimed
regression.

[1] http://lkml.kernel.org/r/alpine.DEB.2.21.1910021556270.187014@xxxxxxxxxxxxxxxxxxxxxxxxx

-- 
Michal Hocko
SUSE Labs