Re: [PATCH v2]mm/oom-kill: direct hardware access processes should get bonus

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

> On Mon, 15 Nov 2010, KOSAKI Motohiro wrote:
> 
> > > I think in cases of heuristics like this where we obviously want to give 
> > > some bonus to CAP_SYS_ADMIN that there is consistency with other bonuses 
> > > given elsewhere in the kernel.
> > 
> > Keep comparision apple to apple. vm_enough_memory() account _virtual_ memory.
> > oom-killer try to free _physical_ memory. It's unrelated.
> > 
> 
> It's not unrelated, the LSM function gives an arbitrary 3% bonus to 
> CAP_SYS_ADMIN.  

Unrelated. LSM _is_ security module. and It only account virtual memory.


> Such threads should also be preferred in the oom killer 
> over other threads since they tend to be more important but not an overly 
> drastic bias such that they don't get killed when using an egregious 
> amount of memory.  So in selecting a small percentage of memory that tends 
> to be a significant bias but not overwhelming, I went with the 3% found 
> elsewhere in the kernel.  __vm_enough_memory() doesn't have that 
> preference for any scientifically calculated reason, it's a heuristic just 
> like oom_badness().

__vm_enough_memory() only gurard to memory overcommiting. And it doesn't
have any recover way. We expect admin should recover their HAND. In the
other hand, oom-killer _is_ automatic recover way. It's no need admin's 
hand. That's the reason why CAP_ADMIN is important or not.




> > > > CAP_SYS_RAWIO mean the process has a direct hardware access privilege
> > > > (eg X.org, RDB). and then, killing it might makes system crash.
> > > > 
> > > 
> > > Then you would want to explicitly filter these tasks from oom kill just as 
> > > OOM_SCORE_ADJ_MIN works rather than giving them a memory quantity bonus.
> > 
> > No. Why does userland recover your mistake?
> > 
> 
> You just said killing any CAP_SYS_RAWIO task may make the system crash, so 
> presuming that you don't want the system to crash, you are suggesting we 
> should make these threads completely immune?  That's never been the case 
> (and isn't for oom_kill_allocating_task, either), so there's no history 
> you can draw from to support your argument.

No. I only require YOU have to investigate userland usecase BEFORE making
change.