Re: [RFC] memory cgroup: weak points of kmem accounting design

[Vladimir Davydov <vdavydov-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org>]

Hi Greg,

On Wed, Sep 17, 2014 at 09:04:00PM -0700, Greg Thelen wrote:
> I've found per memcg per cache type stats useful in answering "why is my
> container oom?"  While these are kernel allocations, it is common for
> user space operations to cause these allocations (e.g. lots of open file
> descriptors).  So I don't specifically need per memcg slabinfo formatted
> data, but at the least a per memcg per cache type active object count
> would be very useful.  Thus I imagine each memcg would have an array of
> slab cache types each with per-cpu active object counters.  Per-cpu is
> used to avoid trashing those counters between cpus as objects are
> allocated and freed.

Hmm, that sounds sane. One more argument for the current design.

> As you say only memcg shrinkable cache types would need list heads.  I
> assume these per memcg shrinkable object list heads would be per cache
> type per cpu list heads for cache performance.  Allocation of a dentry
> today uses the normal slab management structures.  In this proposal I
> suspect the dentry would be dual indexed: once in the global slab/slub
> dentry lru and once in the per memcg dentry list.  If true, this might
> be a hot path regression allocation speed regression.
> 
> Do you have a shrinker design in mind?  I suspect this new design would
> involve a per memcg dcache shrinker which grabs a big per-memcg dcache
> lock while walking the dentry list.  The classic per superblock
> shrinkers would not used for memcg shrinking.

To be honest, I hadn't elaborated that in my mind when I sent this
e-mail, but now I realize that it doesn't look as if there's an easy way
to implement shrinkers in such a setup efficiently. I thought we could
keep each dentry/inode simultaneously in two list, global and memcg.
However, apart from resulting in memory wastes this, as you pointed out,
would result in a regression in operating on the lrus, which is
unacceptable.

That said, I admit my idea sounds crazy. I think sticking to Glauber's
design and trying to make it work is the best we can do now.

Thanks,
Vladimir