Re: [patch] mm, thp: always direct reclaim for MADV_HUGEPAGE even when deferred

[Michal Hocko]

On Wed 28-12-16 13:33:49, David Rientjes wrote:
> On Wed, 28 Dec 2016, Michal Hocko wrote:
> 
> > I do care more about _users_ and their _experience_ than what
> > application _writers_ think is the best. This is the whole point
> > of giving the defrag tunable. madvise(MADV_HUGEPAGE) is just a hint to
> > the system that using transparent hugepages is _preferable_, not
> > mandatory. We have an option to allow stalls for those vmas to increase
> > the allocation success rate. We also have tunable to completely ignore
> > it. And we should also have an option to not stall.
> > 
> 
> The application developer who uses madvise(MADV_HUGEPAGE) is doing so for 
> a reason.

and nobody questions that... But the application developer can hardly
forsee the environment where the application runs. And what might
look as a reasonable cost/benefit balance in one setup can turn out
completely wrong in a different one - just consider the fragmentation
which is the primary contributor to stalls. It is hardly predictable
and vary between different workloads/setups a lot. While we have a way
(policty if you will) to tell that madvise should be honored as much
as possible (defrag=madvise) we do not have a way to tell that even
madvised vmas are not worth stalling over because the benefit would not
offset the cost.

> We lack the ability to defragment in the background for all users who 
> don't want to block while allowing madvise(MADV_HUGEPAGE) users to block, 
> as the changelog for this patch clearly indicates.

And I agree that this is something to be addressed. I just disagree that
this patch is the way how to achieve that.
-- 
Michal Hocko
SUSE Labs