Re: [PATCH] mm: introduce oom_kill_disable sysctl knob

[Minchan Kim]

On Mon, Nov 09, 2020 at 05:06:18PM +0100, Michal Hocko wrote:
> On Mon 09-11-20 07:39:33, Minchan Kim wrote:
> > On Mon, Nov 09, 2020 at 08:37:06AM +0100, Michal Hocko wrote:
> > > On Fri 06-11-20 12:32:38, Minchan Kim wrote:
> > > > It's hard to have some tests to be supposed to work under heavy
> > > > memory pressure(e.g., injecting some memory hogger) because
> > > > out-of-memory killer easily kicks out one of processes so system
> > > > is broken or system loses the memory pressure state since it has
> > > > plenty of free memory soon so.
> > > 
> > > I do not follow the reasoning here. So you want to test for a close to
> > > no memory available situation and the oom killer stands in the way
> > > because it puts a relief?
> > 
> > Yub, technically, I'd like to have consistent memory pressure to cause
> > direct reclaims on proesses on the system and swapping in/out.
>  
> > > 
> > > > Even though we could mark existing process's oom_adj to -1000,
> > > > it couldn't cover upcoming processes to be forked for the job.
> > > 
> > > Why?
> > 
> > Thing is the system has out-of-control processes created on demand.
> > so only option to prevent OOM is echo -1000 > `pidof the process`
> > since they are forked. However, I have no idea when they are forked
> > so should race with OOM with /proc polling and OOM is frequently
> > ahead of me.
> 
> I am still confused. Why would you want all/most processes to be hidden
> from the oom killer?

If one of processes in the system is killed, the memory pressure
disappear.

>  
> > > > This knob is handy to keep system memory pressure.
> > > 
> > > This sounds like a very dubious reason to introduce a knob to cripple
> > > the system.
> > > 
> > > I can see some reason to control the oom handling policy because the
> > > effect of the oom killer is really disruptive but a global on/off switch
> > > sounds like a too coarse interface. Really what kind of production
> > > environment would ever go with oom killer disabled completely?
> > 
> > I don't think shipping production system will use it. It would be
> > just testing only option.
> 
> Then it doesn't really belong to the kernel IMHO.
> 
> > My intention uses such heavy memory load to see various system behaviors
> > before the production launching because it usually happens in real workload
> > once we shipped but hard to generate such a corner case without artificial
> > memory pressure.
> 
> But changing the oom behavior will result in a completely different
> system behavior. So you would be testing something that doesn't really
> happen in any production system.

Since OOM is not instantly reacted, it still provides a good chance how
the system reacts on such memory pressure until someeone releases the
memory. For example, Android already gives lots of system processes to 
-1000 which could effectively disable the OOM at certain point.

> 
> > Any suggestion?
> 
> Not really because I still do not understand your objective. You can
> generate memory pressure and tune it up for specific testing scenario.
> Sure there will be a some interference from the background noise (kernel
> subsystems reacting to external events, processes created etc.) but why
> that is a problem? This is normal to any running system.

Putting more pressure from background processes is okay for the goal but
not okay for relieving th memory pressure since we lost the testing
environment.