Re: [RFC v3] memcg: add memcg lru

[Hillf Danton <hdanton@sina.com>]

On Mon, 18 Nov 2019 11:29:50 +0100 Michal Hocko wrote:
> 
> On Sun 17-11-19 19:35:26, Hillf Danton wrote:
> > 
> > Currently soft limit reclaim (slr) is frozen, see
> > Documentation/admin-guide/cgroup-v2.rst for reasons.
> > 
> > This work adds memcg hook into kswapd's logic to bypass slr, paving
> > a brick for its cleanup later.
> > 
> > After b23afb93d317 ("memcg: punt high overage reclaim to
> > return-to-userland path"), high limit breachers (hlb) are reclaimed
> > one after another spiraling up through the memcg hierarchy before
> > returning to userspace.
> > 
> > The current memcg high work helps to add the lru because we get to
> > collect hlb at zero price and in particular without adding changes
> > to the high work's behavior.
> > 
> > Then a fifo list, which is essencially a simple copy of the page lru,
> > is needed to facilitate queuing up hlb and ripping pages off them in
> > round robin once kswapd starts doing its job.
> > 
> > Finally new hook is added with slr's two problems addressed i.e.
> > hierarchy-unaware reclaim and overreclaim.
> > 
> > Thanks to Rong Chen for testing.
> 
Hey Michal

Thanks for your comments, this time and previous.

> You have ignored the previous review feedback again [1]. I have nacked
> the patch on grounds that it is completely missing any real use case
> scenario or any numbers suggesting there is an actual improvement.
> 
You are right though around half.

After another peep at your comment on v2, I think you didn't approve
the change added in high work to defer reclaim until kswapd becomes
active with good reasoning. That defer is cut in v3.

The added lru will take the place of the current slr, so slr's use
cases apply to it with no exception, yes? Please feel free let us
know what use cases else you may have interests in.

And the concerned improvement is just on the table, Sir, if we fail
to see regressions following lru, with no need to say the price we
pay for maintaining hlb is close to zero for instance among others.

Hillf

> Please do not post new versions until you make those things clear.
> 
> [1] http://lkml.kernel.org/r/20191029083730.GC31513@dhcp22.suse.cz
> -- 
> Michal Hocko
> SUSE Labs