On Mon, Sep 18, 2023 at 07:52:23PM +0400, Nikolay
Edigaryev wrote:
Hello Stefano, I will try to clarify:
I have a single host machine, a dedicated amd64 server, capable of
running multiple Cloud Hypervisor virtual machines backed by /dev/kvm.
I also have a daemon-less CLI software that can provision as many VM
instances as the user wants, e.g. by running "mycli create --kernel
... --disk ... ubuntu".
To run a VM, the user types "mycli run ubuntu", which results in the
creation of two TAP interfaces: one is for passt, one is for Cloud
Hypervisor
"mycli run" then creates a bridge(8) interface, assigns a free IP from
/29 network to it (for example, 10.0.0.3/29), and adds both the TAP
interfaces to that bridge forming up a virtual switch, which allows
passt <-> VM and host <-> communication.
Ok. So, to check my understanding: the VM only has a single virtual
NIC, which connects to this bridge, then you're connecting the bridge
to the outside world using passt. Is that correct?
"mycli run ubuntu" also invokes the
passt with the following arguments:
>passt --foreground --address 10.0.0.2 --netmask 255.255.255.248 --gateway 10.0.0.1
--mac-addr 52:f1:18:34:28:0b -4 --mtu 1500 --tap-fd 3
What owns the address 10.0.0.1 here? I'm assuming that's an address
of the host, but is it on an external interface, or on this special
bridge? Or somewhere else?
[Btw, clamping the passt mtu to 1500 is probably going to be pretty
bad for TCP throughput]
Now to the issue: if the user wants to access the
VM, for provisioning
purposes, e.g. by running "ssh 10.0.0.2", there's a race between the
real ARP reply from that VM and an ARP reply from passt due to the
code fixed in the patch above.
And even if we add a static ARP entry for that VM on the host, there's
still exist a race on the VM's side.
Here the VM looks up the host's ethernet address and receives one
reply from host (ba:46:4e:27:8b:93) and another from passt
(52:f1:18:34:28:0b):
17:26:42.685718 5a:b7:e3:dc:bb:9f > ba:46:4e:27:8b:93, ethertype ARP
(0x0806), length 42: Request who-has 10.0.0.3 tell 10.0.0.2, length 28
17:26:42.685744 ba:46:4e:27:8b:93 > 5a:b7:e3:dc:bb:9f, ethertype ARP
(0x0806), length 42: Reply 10.0.0.3 is-at ba:46:4e:27:8b:93, length 28
17:26:42.685908 52:f1:18:34:28:0b > 5a:b7:e3:dc:bb:9f, ethertype ARP
(0x0806), length 42: Reply 10.0.0.3 is-at 52:f1:18:34:28:0b, length 28
Right.
Ok, so Stefano mentioned that this change will break the case of a
guest not using the gateway it's supposed to. That's true, but
there's certainly a pretty strong case that no-one has any right to
expect that case to work anyway, so we need not consider it.
I believe there's some other rare but legitimate cases it can also
break though. For now I think these can only occur with pasta, not
passt, but they'd still be affected:
* Although it's not common, it's possible to have a default route
with an interface, but no gateway (this can occur if the host has
connectivity over a point to point link like a VPN). With pasta
--config-net we'll copy that gateway-less default route to the
namespace, and it will then ARP for *everything*. That will work
now, because we'll answer all those arps, but would not if we only
arp the gateway address.
* A lesser version of the same same thing: even if we have a normal
default gateway, we may also have specific subnet routes on the
host which override it. With pasta --config-net again we will copy
those routes to the namespace, and so packets routed that way will
induce ARPs for something other than the default gateway (either
for the destination address or for the route specific gateway).
Apart from the ARP issues, I think there's at least one other
fragility in the setup you've described. This is what I was thinking
about when I mentioned elsewhere that I don't think ARP will be the
only issue with having a non-trivial broadcast domain on the guest
side of passt:
If from the host you to send packets on the bridge addressed to
passt's address, rather than the host, I believe that would cause
passt to update its 'addr_seen' to that of the host. That could then
cause packets which should be going to the guest to be sent to the
host instead. That could have a variety of effects from just a brief
interruption to essentially breaking connectivity.