On Tue, May 24, 2011 at 03:34:33PM +0200, Jan Kiszka wrote:
> On 2011-05-24 15:01, Gleb Natapov wrote:
> > On Tue, May 24, 2011 at 02:42:55PM +0200, Jan Kiszka wrote:
> >> On 2011-05-24 14:37, Gleb Natapov wrote:
> >>> On Mon, May 23, 2011 at 04:48:16PM +0200, Jan Kiszka wrote:
> >>>> This aligns the code to what the documentation claims: Allow everything
> >>>> but requests that would have to be routed outside of the virtual LAN.
> >>>>
> >>>> So we need to drop the unneeded IP-level filter, allow TFTP requests,
> >>>> and add the missing protocol-level filter to ICMP.
> >>>>
> >>> May be I am missing something, but how do you disallow requests by
> >>> removing code that actually does filtering.
> >>
> >> All we need to filter are the per-IP-protocol parts that do the
> >> forwarding via the host IP stack. That does not need to happen at IP level.
> >>
> >> Moreover, the existing code contained some practically dead bits anyway:
> >>
> >> if ((ip->ip_dst.s_addr & slirp->vnetwork_mask.s_addr) ==
> >> slirp->vnetwork_addr.s_addr) {
> >> if (ip->ip_dst.s_addr == 0xffffffff && ip->ip_p !=
> >> IPPROTO_UDP)
> >> goto bad;
> >>
> >> This could only trigger if vnetwork_mask.s_addr was 0 (the same applied
> >> to the original code before my refactoring in 2009).
> >>
> > Not sure what do you mean by that. This checks that the ip_dst.s_addr is in
> > the vnetwork range. It does this by comparing net mask bits of
> > ip_dst.s_addr with
> > vnetwork_addr.s_addr. Grep for vnetwork_mask.s_addr. This idiom is used
> > many times throughout the code.
>
> Ok, it's a bit more tricky, and I contributed some buglet. Let
>
> ip_dst.s_addr = 255.255.255.255
> vnetwork_mask.s_addr = 0.255.255.255
Isn't it 255.0.0.0?
> vnetwork_addr.s_addr = 10.0.2.0
> (QEMU's strange defaults)
>
> then dst & vnetwork_mask != vnetwork_addr, so the second condition to
> exclude network broadcasts can't trigger.
>
> Your original code matched the first three bytes of dst against the
> first three of vnetwork_addr, mine inverted the mask. However, both
> variants fail to let DHCP broadcasts pass.
The original code used memcmp which return 0 when equal. When you
changed it to use variable length mask you also inverted if() condition.
It should be != not == !
The code worked (for some value of 'worked') back then :)
>
> In short, this was always wrong and unneeded as we can (and partly
> already did) check for restricted mode in the various IP protocols.
>
Checking of exec_list here and again in tcp_input looks suspiciously
similar. I have to admit I do not remember much about slirp code though.
--
Gleb.
