Re: [PATCH] mm: memcontrol: reclaim when shrinking memory.high below usage

[Vladimir Davydov]

On Fri, Mar 11, 2016 at 03:22:27PM +0100, Michal Hocko wrote:
> On Fri 11-03-16 17:01:46, Vladimir Davydov wrote:
> > On Fri, Mar 11, 2016 at 02:39:36PM +0100, Michal Hocko wrote:
> > > On Fri 11-03-16 14:49:34, Vladimir Davydov wrote:
> > > > On Fri, Mar 11, 2016 at 10:53:09AM +0100, Michal Hocko wrote:
> > > > > > OTOH memory.low and memory.high are perfect to be changed dynamically,
> > > > > > basing on containers' memory demand/pressure. A load manager might want
> > > > > > to reconfigure these knobs say every 5 seconds. Spawning a thread per
> > > > > > each container that often would look unnecessarily overcomplicated IMO.
> > > > > 
> > > > > The question however is whether we want to hide a potentially costly
> > > > > operation and have it unaccounted and hidden in the kworker context.
> > > > 
> > > > There's already mem_cgroup->high_work doing reclaim in an unaccounted
> > > > context quite often if tcp accounting is enabled.
> > > 
> > > I suspect this is done because the charging context cannot do much
> > > better.
> > > 
> > > > And there's kswapd.
> > > > memory.high knob is for the root only so it can't be abused by an
> > > > unprivileged user. Regarding a privileged user, e.g. load manager, it
> > > > can screw things up anyway, e.g. by configuring sum of memory.low to be
> > > > greater than total RAM on the host and hence driving kswapd mad.
> > > 
> > > I am not worried about abuse. It is just weird to move something which
> > > can be perfectly sync to an async mode.
> > >  
> > > > > I mean fork() + write() doesn't sound terribly complicated to me to have
> > > > > a rather subtle behavior in the kernel.
> > > > 
> > > > It'd be just a dubious API IMHO. With memory.max everything's clear: it
> > > > tries to reclaim memory hard, may stall for several seconds, may invoke
> > > > OOM, but if it finishes successfully we have memory.current less than
> > > > memory.max. With this patch memory.high knob behaves rather strangely:
> > > > it might stall, but there's no guarantee you'll have memory.current less
> > > > than memory.high; moreover, according to the documentation it's OK to
> > > > have memory.current greater than memory.high, so what's the point in
> > > > calling synchronous reclaim blocking the caller?
> > > 
> > > Even if the reclaim is best effort it doesn't mean we should hide it
> > > into an async context. There is simply no reason to do so. We do the
> > > some for other knobs which are performing a potentially expensive
> > > operation and do not guarantee the result.
> > 
> > IMO it depends on what a knob is used for. If it's for testing or
> > debugging or recovering the system (e.g. manual oom, compact,
> > drop_caches), this must be synchronous, but memory.high is going to be
> > tweaked at runtime during normal system operation every several seconds
> > or so,
> 
> Is this really going to happen in the real life?

I don't think anyone can say for sure now, but I think it's possible.

> And if yes is it really probable that such an adjustment would cause
> such a large disruption?

Dunno.

> 
> > at least in my understanding. I understand your concern, and may
> > be you're right in the end, but think about userspace that will probably
> > have to spawn thousands threads every 5 seconds or so just to write to a
> > file. It's painful IMO.
> > 
> > Are there any hidden non-obvious implications of handing over reclaim to
> > a kernel worker on adjusting memory.high? May be, I'm just missing
> > something obvious, and it can be really dangerous or sub-optimal.
> 
> I am just thinking about what would happen if workers just start
> stacking up because they couldn't be processed and then would race with
> each other.

We can use one work per memcg (high_work?) and make it requeue itself if
it finds that memory usage is above memory.high and there are
reclaimable pages, so no stacking should happen.

> I mean this all would be fixable but I really fail to see
> how that makes sense from the very beginning.

This is a user API we're defining right now. Changing it in future will
be troublesome.

Thanks,
Vladimir