Re: [PATCH] netfilter: per netns nf_conntrack_cachep

[Jon Masters <jonathan@jonmasters.org>]

On Thu, 2010-02-04 at 14:04 +0100, Patrick McHardy wrote:
> Alexey Dobriyan wrote:
> > On Thu, Feb 4, 2010 at 2:30 PM, Patrick McHardy <kaber@trash.net> wrote:
> >> Alexey Dobriyan wrote:
> >>> On Thu, Feb 4, 2010 at 2:25 PM, Patrick McHardy <kaber@trash.net> wrote:
> >>>> Jon Masters wrote:
> >>>>> On Wed, 2010-02-03 at 21:09 +0200, Alexey Dobriyan wrote:
> >>>>>> On Wed, Feb 03, 2010 at 01:38:09PM -0500, Jon Masters wrote:
> >>>>>>> *). Per namespace cacheing allocation (the cachep bits). We know it's
> >>>>>>> still possible for weirdness to happen in the SLAB cache here.
> >>>>>> Tiny race, needs reproducer.
> >>>>> Maybe. I think it's worth fixing anyway.
> >>>> Absolutely, I'll also apply Eric's patch with the %p fix for the
> >>>> slab name.
> >>> This would show kernel pointers in userspace ;-)
> >>> So, net->id is required.
> >> I don't see the problem. But yes, it would be nicer to have an ID.
> > 
> > This is done (or rather, not done) to not show attackers
> > where data structures are.
> 
> That's news to me, my /proc is full of kernel space pointers,
> including data.

And anyway, you can guess it in many cases. I don't think there's a huge
problem, but you could of course just stash a global atomic somewhere to
keep track of how many of these you've made if that's easier for now.

> In any case, we need a fix for this suitable for 2.6.33. If
> you don't like using the pointer, please send a patch to add
> an id to the network namespaces.

Right. I think the quick solution is fine for 2.6.33. So that makes the
hashtable non-resize patch, the crash fix, and the cachep bits. I will
try to get involved and help you out with the per-ns hashtable clean
rather than just being a whiner :)

Thanks a bunch! Fedora kernels have already been built with this fix,
since it will allow us to close a fair number of "KVM goes boom" bugs.

Jon.