Re: [PATCH 1/2] gfp: add __GFP_NOACCOUNT

[Michal Hocko <mhocko-AlSwsSmVLrQ@public.gmane.org>]

On Wed 06-05-15 15:24:31, Vladimir Davydov wrote:
> On Wed, May 06, 2015 at 01:59:41PM +0200, Michal Hocko wrote:
> > On Tue 05-05-15 12:45:42, Vladimir Davydov wrote:
> > > Not all kmem allocations should be accounted to memcg. The following
> > > patch gives an example when accounting of a certain type of allocations
> > > to memcg can effectively result in a memory leak.
> > 
> > > This patch adds the __GFP_NOACCOUNT flag which if passed to kmalloc
> > > and friends will force the allocation to go through the root
> > > cgroup. It will be used by the next patch.
> > 
> > The name of the flag is way too generic. It is not clear that the
> > accounting is KMEMCG related. __GFP_NO_KMEMCG sounds better?
> > 
> > I was going to suggest doing per-cache rather than gfp flag and that
> > would actually work just fine for the kmemleak as it uses its own cache
> > already. But the ida_simple_get would be trickier because it doesn't use
> > any special cache and more over only one user seem to have a problem so
> > this doesn't sound like a good fit.
> 
> I don't think making this flag per-cache is an option either, but for
> another reason - it would not be possible to merge such a kmem cache
> with caches without this flag set. As a result, total memory pressure
> would increase, even for setups without kmem-active memory cgroups,
> which does not sound acceptable to me.

I am not sure I see the performance implications here because kmem
accounted memcgs would have their copy of the cache anyway, no?
Anyway, I guess it would be good to document these reasons in the
changelog.
 
> > So I do not object to opt-out for kmemcg accounting but I really think
> > the name should be changed.
> 
> I named it __GFP_NOACCOUNT to match with __GFP_NOTRACK, which is a very
> specific flag too (kmemcheck),  nevertheless it has a rather generic
> name.

__GFP_NOTRACK is a bad name IMHO as well. One has to go and check the
comment to see this is kmemleak related.

> Anyways, what else apart from memcg can account kmem so that we have to
> mention KMEMCG in the flag name explicitly?

NOACCOUNT doesn't imply kmem at all so it is not clear who is in charge
of the accounting. I do not insist on __GFP_NO_KMEMCG of course but
it sounds quite specific about its meaning and scope.
-- 
Michal Hocko
SUSE Labs