Re: [PATCH 5/8] mm: memcontrol: account socket memory on unified hierarchy

[Michal Hocko]

On Tue 27-10-15 11:41:38, Johannes Weiner wrote:
> On Tue, Oct 27, 2015 at 01:26:47PM +0100, Michal Hocko wrote:
> > On Mon 26-10-15 12:56:19, Johannes Weiner wrote:
> > [...]
> > > Now you could argue that there might exist specialized workloads that
> > > need to account anonymous pages and page cache, but not socket memory
> > > buffers.
> > 
> > Exactly, and there are loads doing this. Memcg groups are also created to
> > limit anon/page cache consumers to not affect the others running on
> > the system (basically in the root memcg context from memcg POV) which
> > don't care about tracking and they definitely do not want to pay for an
> > additional overhead. We should definitely be able to offer a global
> > disable knob for them. The same applies to kmem accounting in general.
> 
> I don't see how you make such a clear distinction between, say, page
> cache and the dentry cache, and call one user memory and the other
> kernel memory.

Because the kernel memory footprint would be so small that it simply
doesn't change the picture at all. While the page cache or anonymous
memory consumption might be so large it might be disruptive.  I am
talking about loads where good enough is better than "perfect" and
ephemeral global memory pressure when kmem goes over expectations is
better than a permanent cpu overhead. Whatever we do it will always
be non-zero.

Also kmem accounting will make the load more non-deterministic because
many of the resources are shared between tasks in separate cgroups
unless they are explicitly configured. E.g. [id]cache will be shared
and first to touch gets charged so you would end up with more false
sharing.

Nevertheless, I do not want to shift the discussion from the topic. I
just think that one-fits-all simply won't work.

> That just doesn't make sense to me. They're both kernel
> memory allocated on behalf of the user, the only difference being that
> one is tracked on the page level and the other on the slab level, and
> we started accounting one before the other.
> 
> IMO that's an implementation detail and a historical artifact that
> should not be exposed to the user. And that's the thing I hate about
> the current opt-out knob.
> 
> > > I don't think there is a compelling case for an elaborate interface
> > > to make individual memory consumers configurable inside the memory
> > > controller.
> > 
> > I do not think we need an elaborate interface. We just want to have
> > a global boot time knob to overwrite the default behavior. This is
> > few lines of code and it should give the sufficient flexibility.
> 
> Okay, then let's add this for the socket memory to start with. I'll
> have to think more about how to distinguish the slab-based consumers.
> Or maybe you have an idea.

Isn't that as simple as enabling the jump label during the
initialization depending on the knob value? All the charging paths
should be disabled by default already.

> For now, something like this as a boot commandline?
> 
> cgroup.memory=nosocket

That would work for me. I would even see a place to have
CONFIG_MEMCG_TCP_KMEM_ENABLED config option for the default and
[no]socket as a kernel parameter to override the configuratioin default.
This would allow distributions to define their policy without enforcing
it hard and those who compile the kernel to define their own policy.

-- 
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/