ifaddr refcount problem

[Alan Somers]

On Fri, Jun 20, 2014 at 1:15 PM, Navdeep Parhar <navdeep at chelsio.com> wrote:
> Revision 264905 and 266860 that followed it seem to leak ifaddr
> references.  ifa_ifwithdstaddr and ifa_ifwithnet both install a
> reference on the ifaddr returned to the caller but ip_output does not
> release it, eventually leading to a panic when the refcount wraps over
> to 0 and the ifaddr is freed while it is still on various lists.

Are you referring to the ifa_ref() calls at lines 1674, 1745, 1756,
and 1795 of sys/net/if.c?  Those long predate 264905.    Why do you
think that 264905 introduced the bug?  If anything, I guess that it
was introduced by change 262747, which removed ifa_free() from the
done: block of ip_output().

>
> I'm using this patch for now.  Thoughts?
>
> Regards,
> Navdeep
>
>
> diff -r 6dfcecd314af sys/netinet/ip_output.c
> --- a/sys/netinet/ip_output.c   Fri Jun 20 10:33:22 2014 -0700
> +++ b/sys/netinet/ip_output.c   Fri Jun 20 12:07:12 2014 -0700
> @@ -243,6 +243,7 @@ again:
>                 ifp = ia->ia_ifp;
>                 ip->ip_ttl = 1;
>                 isbroadcast = 1;
> +               ifa_free((void *)ia);
>         } else if (flags & IP_ROUTETOIF) {
>                 if ((ia = ifatoia(ifa_ifwithdstaddr(sintosa(dst)))) == NULL &&
>                     (ia = ifatoia(ifa_ifwithnet(sintosa(dst), 0))) == NULL) {
> @@ -253,6 +254,7 @@ again:
>                 ifp = ia->ia_ifp;
>                 ip->ip_ttl = 1;
>                 isbroadcast = in_broadcast(dst->sin_addr, ifp);
> +               ifa_free((void *)ia);
>         } else if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) &&
>             imo != NULL && imo->imo_multicast_ifp != NULL) {
>                 /*

I don't think this is quite right.  ip_output() references ia at lines
366, 433, 630-637, 674, and 675.  All of those references have NULL
checks, so your patch won't cause a use-after-free bug, but I think
that the patch might cause statistics collection to fail, and it might
cause the IP source address to not be set.

Do you have a test case that can reproduce the panic?

-Alan