lkml.org 
[lkml]   [2019]   [Apr]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC 0/2] opportunistic memory reclaim of a killed process
On Thu 11-04-19 15:14:30, Joel Fernandes wrote:
> On Thu, Apr 11, 2019 at 08:12:43PM +0200, Michal Hocko wrote:
> > On Thu 11-04-19 12:18:33, Joel Fernandes wrote:
> > > On Thu, Apr 11, 2019 at 6:51 AM Michal Hocko <mhocko@kernel.org> wrote:
> > > >
> > > > On Wed 10-04-19 18:43:51, Suren Baghdasaryan wrote:
> > > > [...]
> > > > > Proposed solution uses existing oom-reaper thread to increase memory
> > > > > reclaim rate of a killed process and to make this rate more deterministic.
> > > > > By no means the proposed solution is considered the best and was chosen
> > > > > because it was simple to implement and allowed for test data collection.
> > > > > The downside of this solution is that it requires additional “expedite”
> > > > > hint for something which has to be fast in all cases. Would be great to
> > > > > find a way that does not require additional hints.
> > > >
> > > > I have to say I do not like this much. It is abusing an implementation
> > > > detail of the OOM implementation and makes it an official API. Also
> > > > there are some non trivial assumptions to be fullfilled to use the
> > > > current oom_reaper. First of all all the process groups that share the
> > > > address space have to be killed. How do you want to guarantee/implement
> > > > that with a simply kill to a thread/process group?
> > >
> > > Will task_will_free_mem() not bail out in such cases because of
> > > process_shares_mm() returning true?
> >
> > I am not really sure I understand your question. task_will_free_mem is
> > just a shortcut to not kill anything if the current process or a victim
> > is already dying and likely to free memory without killing or spamming
> > the log. My concern is that this patch allows to invoke the reaper
>
> Got it.
>
> > without guaranteeing the same. So it can only be an optimistic attempt
> > and then I am wondering how reasonable of an interface this really is.
> > Userspace send the signal and has no way to find out whether the async
> > reaping has been scheduled or not.
>
> Could you clarify more what you're asking to guarantee? I cannot picture it.
> If you mean guaranteeing that "a task is dying anyway and will free its
> memory on its own", we are calling task_will_free_mem() to check that before
> invoking the oom reaper.

No, I am talking about the API aspect. Say you kall kill with the flag
to make the async address space tear down. Now you cannot really
guarantee that this is safe to do because the target task might
clone(CLONE_VM) at any time. So this will be known only once the signal
is sent, but the calling process has no way to find out. So the caller
has no way to know what is the actual result of the requested operation.
That is a poor API in my book.

> Could you clarify what is the draback if OOM reaper is invoked in parallel to
> an exiting task which will free its memory soon? It looks like the OOM reaper
> is taking all the locks necessary (mmap_sem) in particular and is unmapping
> pages. It seemed to me to be safe, but I am missing what are the main draw
> backs of this - other than the intereference with core dump. One could be
> presumably scalability since the since OOM reaper could be bottlenecked by
> freeing memory on behalf of potentially several dying tasks.

oom_reaper or any other kernel thread doing the same is a mere
implementation detail I think. The oom killer doesn't really need the
oom_reaper to act swiftly because it is there to act as a last resort if
the oom victim cannot terminate on its own. If you want to offer an
user space API then you can assume users will like to use it and expect
a certain behavior but what that is? E.g. what if there are thousands of
tasks killed this way? Do we care that some of them will not get the
async treatment? If yes why do we need an API to control that at all?

Am I more clear now?

--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2019-04-11 22:13    [W:0.078 / U:2.840 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site