On Thu, May 16, 2013 at 12:27:52PM +0100, Nicholas Thomas wrote:

On Thu, 2013-05-16 at 11:40 +0300, Michael S. Tsirkin wrote:

quoted

On Thu, May 16, 2013 at 09:20:55AM +0100, Nicholas Thomas wrote:

quoted

Hi,

On Thu, 2013-05-16 at 09:27 +0300, Michael S. Tsirkin wrote:

quoted

On Thu, May 16, 2013 at 09:24:05AM +0300, Michael S. Tsirkin wrote:

quoted

Is this with or without vhost-net in host?

never mind, I see it's without.
Try to enable vhost-net (you'll have to switch to -netdev syntax
for that to work) and see if this help.
If it does it's likely a qemu bug if not probably a guest bug.

Switching to -netdev is non-trivial for me, unfortunately.

Interesting. Why is that?

Our setup is bond0 <-> vlanX <-> bridgeX <-> [ tap devices ] and we do
all that outside of qemu at the moment, specifying -net tap,ifname=... -
we also run some processes on the TAP interface and insert a bunch of
ebtables rules between creating it and starting qemu. Duplicating that
with -net bridge seemed close to impossible, and -netdev tap was
throwing EBUSY from /dev/net/tun. I guess our external magic should be
using ,fd= instead.

I'm not sure what's wrong with -netdev tap.
You don't have to use fd=, you can specify ifname= with netdev as well.
Here's what I use:

-net nic,model=virtio,netdev=foo
-netdev tap,id=foo,ifname=msttap0,script=/home/mst/ifup,downscript=no,vhost=on

the netdev/id pair above is almost the same as vlan=20
in your example, except there's always exactly one frontend
and one backend in the netdev case, vlans let you connect
more than 2 devices.

quoted

quoted

Anyway, it's
definitely a qemu bug - it happens on kernels 3.2 and 3.9 with 1.4.1,
but doesn't happen with qemu 0.15.0 or 1.5.0rc1.

I'll have a dig through git to see if I can identify the patch that
resolves it. It feels-like qemu sometimes stops reading from the tap
file descriptor between ipxe exiting and the linux kernel bringing up
the network interface, and never recovers from that.

/Nick

You can try to bisect, yes.

Work have decided to accept 1.5.0 when it arrives instead, so I'm afraid
I won't be working on this after all. 

/Nick