On Fri, 2020-07-17 at 19:44 -0700, John M. Harris Jr wrote:
On Friday, July 17, 2020 10:06:53 AM MST Benjamin Berg wrote:
> What we achieve by killing a process is that we give the kernel more
> flexibility in how it manages the available memory. It really doesn't
> matter what you kill, all that matters is that some memory is free'ed
> up again.
It does matter what you kill, because you're wiping out users' data and
stopping software the user intended to have run. The kernel is already more
than capable of freeing memory for itself, that's not what this Change is for
either. This Change is to abuse the OOM killer to run in non-OOM scenarios
using a userspace daemon.
No, an OOM scenario from a kernel point of view means, that it has no
other choice than to kill a process.
You *really* need to accept that the kernel OOM killer is insufficient
in many scenarios. It is only the last line of defence, that is solely
concerned about whether the system is *technically* capable of running.
But thrashing scenarios are exactly that, *technically* running but
*practically* dead.
I think it only makes sense to continue a discussion if you acknowledge
the existence and really understand the scenario described here:
https://lkml.org/lkml/2019/8/4/15
Benjamin