[pve-devel] [PATCH pve-common] network: disable unicast flooding on tap|veth|fwln ports

[Josef Johansson]

On 1/14/22 12:23, Josef Johansson wrote:
> On 1/14/22 11:51, Wolfgang Bumiller wrote:
>> On Thu, Sep 16, 2021 at 11:48:15PM +0200, alexandre derumier wrote:
>>> Le mercredi 15 septembre 2021 à 19:09 +0200, Thomas Lamprecht a écrit :
>>>> On 15.09.21 17:33, alexandre derumier wrote:
>>>>> I have looked at other hypervisors implementations (as it don't see
>>>>> to
>>>>> have problem with hetzner),
>>>>>
>>>>>
>>>>> https://listman.redhat.com/archives/libvir-list/2014-December/msg00173.html
>>>>>
>>>>>
>>>>> https://docs.vmware.com/en/VMware-NSX-T-Data-Center/3.1/administration/GUID-C5752084-A582-4AEA-BD5D-03FE5DBC746E.html
>>>>>
>>>>>
>>>>> Both vmware && libvirt have a mode to manually manage fdb entries
>>>>> in
>>>>> bridge mac table.
>>>>>
>>>>> This will work if only 1mac is behind 1 nic, so it should be an
>>>>> option
>>>>> (nested hypervisor for examples).
>>>>>
>>>>> but for classic vm , it could allow to disable unicast_flood &&
>>>>> learning for the tap interface, but also promisc mode on tap
>>>>> interface!
>>>>>
>>>>> I was think about add an option on vmbrX  or vnetX directly to
>>>>> enable/disable.
>>>> As this would be on the VM tap devices it would sound somewhat
>>>> reasonable to
>>>> have it as per vNIC setting, but naturally it would then be a bit
>>>> annoying to
>>>> change for all; a tradeoff could be to allow setting the default
>>>> value per
>>>> bridge, node or datacenter (I'd do only one of those).
>>>>
>>>> What do you think?
>>>>
>>> I have done test, I think the best way is to enable it on the bridge
>>>  or vnet for sdn.
>>> because ovs don't support it for example, or its not needed for routed
>>> setup or vxlan.
>>> I don't known too much where add this option for classic vmbr ? in
>>> /etc/network/interfaces ?.
>>> I can't reuse bridge-unicast-flood off, bridge-learning off on vmbr
>>> with ifupdown, because it's apply on all ports (ethX), and we don't
>>> want that.
>>> I could add a custom attribute, but I need to parse
>>> /etc/network/interfaces in this case  for the tap_plug sub. 
>> As far as I can tell, ifupdown2 only applies this to the ports it knows
>> about, so in theory we *could* start to honor this for the interfaces we
>> crate for VMs as a default, and have an on/off/auto value on VM network
>> interfaces (override or use whatever /e/n/interfaces says).
>>
>> Or do you mean you typically want this to be on for VMs but off
>> specifically for the physical port? Then /e/n/interfaces won't fit.
>>
>> Although it *does* allow listing ports and doesn't seem to mind if a
>> port does not exist, so we *may* get away with saying we expect
>> something like this:
>>
>>     bridge-unicast-flood eth0=on _pve=off
>>
>> Either way, it's a port setting, so I wonder a by-vm-interface optional
>> override probably makes sense, not sure (but would be easy enough to
>> do).
>>
>>
>> _______________________________________________
>> pve-devel mailing list
>> pve-devel at lists.proxmox.com
>> https://lists.proxmox.com/cgi-bin/mailman/listinfo/pve-devel
>>
> Maybe something along the lines of
>
> --- Network.pm	2021-05-25 16:35:27.000000000 +0200
> +++ Network.pm.new	2022-01-14 12:20:48.181632198 +0100
> @@ -309,6 +309,7 @@ sub veth_create {
>      disable_ipv6($vethpeer);
>      &$activate_interface($veth);
>      &$activate_interface($vethpeer);
> +    PVE::Tools::run_command(['/sbin/bridge', 'link', 'set', 'dev', $vethpeer, 'flood', 'off']);
>  }
>  
>  sub veth_delete {
>
>
> This is basically what I do right to solve this problem. I leave everything else with unicast_flood on.
> I do not use ovs yet, so basic bridging.
>
> If I enable it on fwln some ARP-functionality stops working.
>
> Regards
> Josef
>
I'm landing on these commits, I've tested this one and it seems to work
properly.
--- a/usr/share/perl5/PVE/Network.pm    2022-01-28 04:20:22.704806437 +0100
+++ b/usr/share/perl5/PVE/Network.pm.new    2022-01-28
04:20:10.256959699 +0100
@@ -335,6 +335,13 @@ my $create_firewall_bridge_linux = sub {
     &$bridge_add_interface($fwbr, $vethfw);
     &$bridge_add_interface($bridge, $vethfwpeer, $tag, $trunks);

+    # When the server does not know where to send traffic, it will
broadcast
+    # the traffic onto every port on a bridge. There's two cases where
this is
+    # not wanted. Servers can siphone traffic not intended for it but also
+    # cause the first request (between correct servers) to fail. Leave
$vethpeer
+    # since otherwise ARP between HV and VM will not work.
+    PVE::Tools::run_command(['/sbin/bridge', 'link', 'set', 'dev',
$vethfwpeer, 'flood', 'off']);
+
     &$bridge_add_interface($fwbr, $iface);
 };

This one is not tested, but should in theory be the same. One thing that
can happen is that outbound traffic from this server is not working
properly. I will do some testing with this soon.

--- /usr/share/perl5/PVE/Network.pm    2022-01-28 04:21:10.524217677 +0100
+++ /usr/share/perl5/PVE/Network.pm.new    2022-01-28 04:35:23.921695488
+0100
@@ -411,6 +411,9 @@ sub tap_plug {
         &$create_firewall_bridge_linux($iface, $bridge, $tag, $trunks);
     } else {
         &$bridge_add_interface($bridge, $iface, $tag, $trunks);
+
+        # Do not allow VMs to siphone traffic not destined for them
+        PVE::Tools::run_command(['/sbin/bridge', 'link', 'set', 'dev',
$iface, 'flood', 'off']);
     }
 
     } else {


Let me know what you think! And I'll make proper patches.
Regards
- Josef