ifaddr refcount problem

[Gleb Smirnoff]

On Mon, Jun 23, 2014 at 10:44:58AM -0600, Alan Somers wrote:
A> > On Fri, Jun 20, 2014 at 12:15:21PM -0700, Navdeep Parhar wrote:
A> > N> Revision 264905 and 266860 that followed it seem to leak ifaddr
A> > N> references.  ifa_ifwithdstaddr and ifa_ifwithnet both install a
A> > N> reference on the ifaddr returned to the caller but ip_output does not
A> > N> release it, eventually leading to a panic when the refcount wraps over
A> > N> to 0 and the ifaddr is freed while it is still on various lists.
A> > N>
A> > N> I'm using this patch for now.  Thoughts?
A> > N>
A> > N> Regards,
A> > N> Navdeep
A> > N>
A> > N>
A> > N> diff -r 6dfcecd314af sys/netinet/ip_output.c
A> > N> --- a/sys/netinet/ip_output.c        Fri Jun 20 10:33:22 2014 -0700
A> > N> +++ b/sys/netinet/ip_output.c        Fri Jun 20 12:07:12 2014 -0700
A> > N> @@ -243,6 +243,7 @@ again:
A> > N>              ifp = ia->ia_ifp;
A> > N>              ip->ip_ttl = 1;
A> > N>              isbroadcast = 1;
A> > N> +            ifa_free((void *)ia);
A> > N>      } else if (flags & IP_ROUTETOIF) {
A> > N>              if ((ia = ifatoia(ifa_ifwithdstaddr(sintosa(dst)))) == NULL &&
A> > N>                  (ia = ifatoia(ifa_ifwithnet(sintosa(dst), 0))) == NULL) {
A> > N> @@ -253,6 +254,7 @@ again:
A> > N>              ifp = ia->ia_ifp;
A> > N>              ip->ip_ttl = 1;
A> > N>              isbroadcast = in_broadcast(dst->sin_addr, ifp);
A> > N> +            ifa_free((void *)ia);
A> > N>      } else if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) &&
A> > N>          imo != NULL && imo->imo_multicast_ifp != NULL) {
A> > N>              /*
A> >
A> > The patch shouldn't use void * casts, but instead specify explicit member:
A> >
A> >         ifa_free(&ia->ia_ifa);
A> >
A> > Apart from that it, the patch looks entirely correct and plugging a leak.
A> > Thanks!
A> 
A> I still don't see how this patch would work without breaking stuff
A> like the statistics collection at line 673 of ip_output.c.  If we call
A> ifa_free immediately after choosing our ifp, then ia won't be
A> available at lines 630 or 673, and ip_output will never record
A> statistics, right?

You are right, thanks.

In case of IP_SENDONES/IP_ROUTETOIF we should hold the reference to ia
throughout the function and free it at the end.

Suggested patch, not tested.

-- 
Totus tuus, Glebius.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ip_output.diff
Type: text/x-diff
Size: 910 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-net/attachments/20140624/cb81dcec/attachment.diff>