Re: [Qemu-devel] using -net dump with tap networking

[Laszlo Ersek]

On 02/19/13 18:50, Markus Armbruster wrote:

> Your example uses -net to connect additional backends to the hubport.
> If I understand you correctly, you can't use -netdev to do that.
> Ignorant question: fundamental reason or just not implemented?

(I'll let Stefan correct me, but I should be able to understand it as
well, since I reviewed his series...)

                   N       1
   NetClientState  --------- NICConf
         ^                      ^
         |                      |
         |                      |
  user/tap/socket/       e1000/virtio-net-pci
  vde/bridge/hubport        -device
     -netdev

-netdev describes an object (a backend) derived from NetClientState.
NetClientState has a single peer.

-device describes a frontend (a guest device). Some of those are network
interfaces, derived from NICConf. (They also inherit a bunch of other
stuff). NICConf has an array of peers.

When a backend is created with -netdev, qemu_new_net_client() is called
in such a way that the peer to set is NULL, no links are created.

The linkage is established from the NIC side,
(a) -net nic,model=...,netdev=...; in qemu_new_nic(),
(b) -device e1000,netdev=...; in parse_netdev() I think.

All 2*N links are created from the NIC side (two per each guest queue I
think), ie. when the frontend is added, that sets the peer link inside
all N referencedbackends as well.

NetHubPort is a special backend (host device) that forwards to all other
NetHubPort backends on the same NetHub. This is internal to NetHubPort.

"-net" is super-confusing.

(1) -net nic,model=...,netdev=... corresponds to case (a) above, it adds
a frontend and "links back" all referenced, pre-existent backends to it.

(2) -net nic,model=...[,vlan=...] creates a NetHubPort backend on the
specified NetHub (vlan defaults to 0), creates the frontend as well, and
links them together.

(3) -net BACKEND[,vlan=...], including BACKEND==dump, creates a
NetHubPort backend on the specified NetHub, also creates an instance of
the requested backend type, and links them together via each one's peer
link.

(
Regarding the (non-hub) backend created in (3), its peer link is already
occupied, hence you cannot set different linkage for it from the
frontend side, as in (a) or (b). In (a) it triggers an assert(), in (b)
it returns -EEXIST I think.

... I can't actually trigger the assert in (a) -- looks like (b) is
always reached first --, but I can trigger the -EEXIST in (b):

# qemu -net dump,name=dump0 -net nic,model=e1000,netdev=dump0

qemu-system-x86_64: Property 'e1000.netdev' can't take value 'dump0',
it's in use
Aborted

# qemu -net dump,name=dump0 -device e1000,netdev=dump0

qemu-system-x86_64: -device e1000,netdev=dump0: Property 'e1000.netdev'
can't take value 'dump0', it's in use

(no abort, exit status 1)
)

In order to take a stab at your question at last:
- the two (parallel, opposing) links created in (3) connect two backends,
- "-netdev" does not set the peer link in the backend, so you cannot use
that,
- "-device" *would* set the links in both directions, but it's for
creating frontends.

IOW (3) is unique because it links backend to backend.

(It could tempt us to create an infinite loop between two NetHubPorts,
ie. "-net hubport,hubid=0"; one path via the NetHub, the other through
the direct peer link. net_init_hubport() is smart enough to prevent
that; you can only add a hubport with -netdev, which doesn't set the
peer link for the backend, and then you can only set it with a -device
or -net nic from the other side.)

Laszlo