Re: [netfilter-core] kernel panic: Out of memory and no killable processes... (2)

[Florian Westphal]

Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
> On Mon, Jan 29, 2018 at 05:57:22PM +0100, Florian Westphal wrote:
> > Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
> > > On Mon, Jan 29, 2018 at 08:23:57AM +0100, Florian Westphal wrote:
> > > > > vmalloc() once became killable by commit 5d17a73a2ebeb8d1 ("vmalloc: back
> > > > > off when the current task is killed") but then became unkillable by commit
> > > > > b8c8a338f75e052d ("Revert "vmalloc: back off when the current task is
> > > > > killed""). Therefore, we can't handle this problem from MM side.
> > > > > Please consider adding some limit from networking side.
> > > > 
> > > > I don't know what "some limit" would be.  I would prefer if there was
> > > > a way to supress OOM Killer in first place so we can just -ENOMEM user.
> > > 
> > > Just supressing OOM kill is a bad idea. We still leave a way to allocate
> > > arbitrary large buffer in kernel.
> > 
> > Isn't that what we do everywhere in network stack?
> > 
> > I think we should try to allocate whatever amount of memory is needed
> > for the given xtables ruleset, given that is what admin requested us to do.
> 
> Is it correct that "admin" in this case is root in random container?

Yes.

> I mean, can we get access to it with CLONE_NEWUSER|CLONE_NEWNET?

Yes.

> This can be fun.

Do we prevent "admin in random container" to insert 2m ipv6 routes
(alternatively: ipsec tunnels, interfaces etc etc)?

> > I also would not know what limit is sane -- I've seen setups with as much
> > as 100k iptables rules, and that was 5 years ago.
> > 
> > And even if we add a "Xk rules" limit, it might be too much for
> > low-memory systems, or not enough for whatever other use case there
> > might be.
> 
> I hate what I'm saying, but I guess we need some tunable here.
> Not sure what exactly.

Would memcg help?

(I don't buy the "run untrusted binaries on linux is safe" thing, so
 I would not know).