Re: [PATCH] mm,hwpoison: non-current task should be checked early_kill for force_early

["HORIGUCHI NAOYA(堀口 直也)" <naoya.horiguchi@nec.com>]

Hi Aili,

On Fri, Jan 15, 2021 at 05:26:22PM +0800, Aili Yao wrote:
> On Fri, 15 Jan 2021 09:49:24 +0100
> Oscar Salvador <osalvador@suse.de> wrote:
> 
> > I am having a hard time trying to grasp what are you trying to achieve here.
> > Could you elaborate some more? Ideally stating what is the problem you are
> > fixing here.
> > 
> Sorry for confusion, example: there are four process A,B,C,D,which map the same file into
> there process space, which set there PF_MCE_KILL_EARLY flag to TRUE, if process A trigger one
> UE with  MF_ACTION_REQUIRED set, in current code, only process A will be killed, B,C,D remain
> alive, but for the PF_MCE_KILL_EARLY we set, we want B,C,D also be killed.

This behavior seems not to me what PF_MCE_KILL_EARLY intends.  This flag
controls whether memory error handler kills processes immediately or not,
and it only affects action optional cases (i.e. called without
MF_ACTION_REQUIRED).  In MF_ACTION_REQUIRED case, we have no such choice
and affected processes should be always killed immediately.

We may also need to consider the difference in context of these two cases.
Action optional case is called asynchronously by background process like
memory scrubbing, so all processes mapping the error memory are the affected
ones.  Action required event is more synchronous, and is called when a
process experiences memory access errors on data load and instruction fetch
instructions.  So the affected process in this case is only the process.
So I still think the this background justifies the current behavior.

But my knowledge might be old, if you have newer hardwares which define
other type of memory error and that doesn't fit with current implementation,
I'd like to extend code to support the new cases, so please let me know.

Thanks,
Naoya Horiguchi