Re: [RFC PATCH] binder: Don't require the binder lock when killed in binder_thread_read()

[Michal Hocko]

On Tue 04-04-17 08:10:03, Doug Anderson wrote:
> Hi,
> 
> On Tue, Apr 4, 2017 at 12:40 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
[...]
> > So the primary question is what are you trying to achieve?
> 
> Ideally it would be nice to bring back patches to make the OOM
> handling lockup free.  However, presuming that's not feasible then at
> least it would be nice if we could get some minimal set of patches
> that:
> 
> - Isn't too scary to backport.
> - Handles the low hanging fruit.
> - Is fairly self contained.

oom_reaper as introduced by aac4536355496 and 4 follow up patches should
be a good yet not too scary to start.

> 1. It would be nice if other users of linuxstable could benefit.

well most of the oom fixes are rather subtle and it took quite some time
to do the properly so I am not really sure we want to take risk. I would
rather handle those issues and backport specific fixes for the issue.
 
> 2. I know the above patches are not as ideal as the work that has
> happened upstream, so of course I'd prefer to get the upstream
> solution.
> 
> 3. I always appreciate being closer to the upstream solution which
> means we get more people looking at the code and more people testing
> the code.

I can hardly help you more than offer to use the upstream code. I fully
realize that this is hard and I am facing similar issues with enterprise
kernel @Suse but so far I have seen OOMs behaving mostly OK except for
extreme cases which usually do not happen enough to take the risk and
backport non-trivial changes to the stable trees.
-- 
Michal Hocko
SUSE Labs