[RFC] Per file OOM badness

[Michal Hocko]

On Fri 19-01-18 17:54:36, Christian König wrote:
> Am 19.01.2018 um 13:20 schrieb Michal Hocko:
> > On Fri 19-01-18 13:13:51, Michal Hocko wrote:
> > > On Fri 19-01-18 12:37:51, Christian König wrote:
> > > [...]
> > > > The per file descriptor badness is/was just the much easier approach to
> > > > solve the issue, because the drivers already knew which client is currently
> > > > using which buffer objects.
> > > > 
> > > > I of course agree that file descriptors can be shared between processes and
> > > > are by themselves not killable. But at least for our graphics driven use
> > > > case I don't see much of a problem killing all processes when a file
> > > > descriptor is used by more than one at the same time.
> > > Ohh, I absolutely see why you have chosen this way for your particular
> > > usecase. I am just arguing that this would rather be more generic to be
> > > merged. If there is absolutely no other way around we can consider it
> > > but right now I do not see that all other options have been considered
> > > properly. Especially when the fd based approach is basically wrong for
> > > almost anybody else.
> > And more importantly. Iterating over _all_ fd which is what is your
> > approach is based on AFAIU is not acceptable for the OOM path. Even
> > though oom_badness is not a hot path we do not really want it to take a
> > lot of time either. Even the current iteration over all processes is
> > quite time consuming. Now you want to add the number of opened files and
> > that might be quite many per process.
> 
> Mhm, crap that is a really good argument.
> 
> How about adding a linked list of callbacks to check for the OOM killer to
> check for each process?
> 
> This way we can avoid finding the process where we need to account things on
> when memory is allocated and still allow the OOM killer to only check the
> specific callbacks it needs to determine the score of a process?

I might be oversimplifying but there really have to be a boundary when
you have the target user context, no? Then do the accounting when you
get data to the user.
-- 
Michal Hocko
SUSE Labs