Using virt-manager "hypervisor default" type:
<interface type='bridge'>
<mac address='00:16:3e:5e:09:9d'/>
<source bridge='br0'/>
<script path='vif-bridge'/>
</interface>
This causes the qemu call to have "-net none" which removes PXE boot
abilities. A linux kernel has network through the xen pv-driver.
Changing in virt-manager to "e1000" type:
<interface type='bridge'>
<mac address='00:16:3e:5e:09:9d'/>
<source bridge='br0'/>
<script path='vif-bridge'/>
<model type='e1000'/>
</interface>
This currently (Xen 4.3.1 + libvirt 1.2.0) fails and the qemu arguments
in the log look suspicious:
-device e1000,id=nic-1,netdev=net-1,mac=00:16:3e:5e:09:9d
-netdev type=tap,id=net-1,ifname=vif1.-1-emu,script=no,downscript=no
Looking through git I found ba64b97134a6129a48684f22f31be92c3b6eef96
libxl: Allow libxl to set NIC devid
which removes a line that sets the devid of new NICs. I assume the comment
says (sorry I always seem to get confused reading it) that devid should
be auto-set by the libxl library provided by Xen. But I could not find
where that would be done. Even the xl command implementation sets it
explicitly. libxl_device_nic_init sets it to -1 which would explain the
wrong qemu arguments above.
For testing I re-added the following after the libxlMakeNic call:
--- libvirt-1.2.0.orig/src/libxl/libxl_conf.c 2013-12-11 17:04:17.000000000 +0
+++ libvirt-1.2.0/src/libxl/libxl_conf.c 2013-12-16 19:08:46.830016646 +0
@@ -907,6 +907,8 @@ libxlMakeNicList(virDomainDefPtr def, l
for (i = 0; i < nnics; i++) {
if (libxlMakeNic(l_nics[i], &x_nics[i]))
goto error;
+ if (x_nics[i].devid < 0)
+ x_nics[i].devid = i;
}
d_config->nics = x_nics;
And with that I get a working PXE boot when I set the device type
in the config. (Side note that I think not using a type defaults to
rtl8139 in the old xend driver. Maybe this should be the same for the
xl driver).
I am not sure how the right fix for it should look like as I am unsure
which part was supposed to set the devid. Just right now it does not
seem to be done.
-Stefan