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

[Michal Hocko]

On Tue 30-01-18 10:57:39, Michal Hocko wrote:
> On Tue 30-01-18 10:02:34, Dmitry Vyukov wrote:
> > On Tue, Jan 30, 2018 at 9:28 AM, Kirill A. Shutemov
> > <kirill@xxxxxxxxxxxxx> wrote:
> > > On Tue, Jan 30, 2018 at 09:11:27AM +0100, Florian Westphal wrote:
> > >> Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > >> > On Mon 29-01-18 23:35:22, Florian Westphal wrote:
> > >> > > Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
> > >> > [...]
> > >> > > > I hate what I'm saying, but I guess we need some tunable here.
> > >> > > > Not sure what exactly.
> > >> > >
> > >> > > Would memcg help?
> > >> >
> > >> > That really depends. I would have to check whether vmalloc path obeys
> > >> > __GFP_ACCOUNT (I suspect it does except for page tables allocations but
> > >> > that shouldn't be a big deal). But then the other potential problem is
> > >> > the life time of the xt_table_info (or other potentially large) data
> > >> > structures. Are they bound to any process life time.
> > >>
> > >> No.
> > >
> > > Well, IIUC they bound to net namespace life time, so killing all
> > > proccesses in the namespace would help to get memory back. :)
> > 
> > ... unless the namespace is mounted into file system.
> > 
> > Let's start with NOWARN as that's what kernel generally uses for
> > allocations with user-controllable size. ENOMEM is roughly as
> > informative as the WARNING message in this case.
> 
> You want __GFP_NORETRY but that is not _fully_ supported by kvmalloc
> right now. More specifically kvmalloc doesn't guanratee that the request
> will not trigger the OOM killer (like regular __GFP_NORETRY). This is
> because of internal vmalloc restrictions. If you are however OK to
> simply bail out in most cases then __GFP_NORETRY should work reasonably
> fine.
> 
> > I think we also need to consider setting up memory cgroup for
> > syzkaller test processes (we do RLIMIT_AS, but that's weak).
> 
> Well, this is not about syzkaller, it merely pointed out a potential
> DoS... And that has to be addressed somehow.

So how about this?
---