Re: [PATCH net] net: mctp: Don't access ifa_index when missing

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 07 May 2025 09:24:29 +0800 Matt Johnston wrote:
> On Tue, 2025-05-06 at 18:06 -0700, Jakub Kicinski wrote:
> > On Mon, 05 May 2025 17:05:12 +0800 Matt Johnston wrote:  
> > > +		/* Userspace programs providing AF_MCTP must be expecting ifa_index filter
> > > +		 * behaviour, as will those setting strict_check.
> > > +		 */
> > > +		if (hdr->ifa_family == AF_MCTP || cb->strict_check)
> > > +			ifindex = hdr->ifa_index;  
> > 
> > The use of cb->strict_check is a bit strange here. I could be wrong but
> > I though cb->strict_check should only impact validation. Not be used
> > for changing behavior.  
> 
> It was following behaviour of inet_dump_addr()/inet6_dump_addr() where
> filtering is applied if strict check is set.
> I don't have strong opinion whether strict_check makes sense for MCTP though
> - it depends on
> whether userspace expects strict_check to apply to all families, or just
> inet4/inet6.

I see your point. And existing user space may expect filtering
even if !cb->strict_check but family is set to AF_MCTP?

> > If you have a reason to believe all user space passes a valid header -
> > how about we just return an error if message is too short?
> > IPv4 and IPv6 seem to return an error if message is short and
> > cb->strict_check, so they are more strict. MCTP doesn't have a ton of
> > legacy user space, we don't have to be lenient at all. My intuition
> > would be to always act like IP acts under cb->strict_check  
> 
> The problem is that programs will pass ifa_family=AF_UNSPEC with a short
> header, no strict_check 
> (eg busybox "ip addr show").
> An AF_UNSPEC request will reach mctp_dump_addrinfo(), so we don't want that
> returning an error.
> Maybe mctp_dump_addrinfo() should ignore AF_UNSPEC requests entirely, and
> only populate
> a response when ifa_family=AF_MCTP. That would be OK for the existing mctp
> userspace programs 
> I know about, though there may be other users that are calling with AF_UNSPEC
> but filtering 
> userspace-side for AF_MCTP addresses.

Right, and looks like IP filters with strict_check regardless of family.
So we'd be even further away from that behavior if we never filtered
with AF_UNSPEC.