Re: [BUG] net_cls: Panic occured when net_cls subsystem use

[Jarek Poplawski <jarkao2@gmail.com>]

On Sun, May 31, 2009 at 09:24:53AM -0400, jamal wrote:
> On Sun, 2009-05-31 at 10:12 +0200, Jarek Poplawski wrote:
> 
> > If somebody runs tc add filter with a new priority but existing handle
> > a newly created (and not linked now) tc would be handled with goto
> > errout and would leak, I guess.
> 
> This would imply the classifier is buggy. I will stare at the different
> classifier - and if any exhibits such traits it needs to be fixed
> 
> > The change (in the two spots) is:
> > > +               if (n->nlmsg_type == RTM_NEWTFILTER &&
> > > +                   (n->nlmsg_flags&(NLM_F_CREATE|NLM_F_EXCL))) {
> > > 
> > 
> > Sorry, but I don't think this change is enough; tc filter replace
> > with only this (n->nlmsg_type == RTM_NEWTFILTER &&
> > (n->nlmsg_flags&(NLM_F_CREATE))) can get here with an "old" tp
> > and will relink it or destroy depending on the ->change() return.
> > 
> 
> Excellent point - there could be buggy user space apps that will do
> that. Minoru change the check to:
> 
> +               if (n->nlmsg_type == RTM_NEWTFILTER &&
> +                   (n->nlmsg_flags&NLM_F_CREATE &&
> +                    n->nlmsg_flags&NLM_F_EXCL)) {


But then, there could be "tc filter replace" with only this
(n->nlmsg_type == RTM_NEWTFILTER && (n->nlmsg_flags&(NLM_F_CREATE)))
which can't get here with a newly created tp, I guess.

Cheers,
Jarek P.