X-Git-Url: http://xvm.mit.edu/gitweb/invirt/packages/invirt-xen-config.git/blobdiff_plain/ba2cbfd952ad6ef034730d81f7c79c71c7d674d9..ae33cf776d263f7027adc9ec7ff3004fae72c9ff:/vif-invirtroute

diff --git a/vif-invirtroute b/vif-invirtroute
index 49a7e4c..4872758 100755
--- a/vif-invirtroute
+++ b/vif-invirtroute
@@ -54,6 +54,27 @@ if [  ${vif_type} != "ioemu"  -o  x${qemu_online} = xyes ] ; then
     # If we've been given a list of IP addresses, then add routes from dom0 to
     # the guest using those addresses.
 	for addr in ${ip} ; do
+	    # When PVHVM is enabled, Xen plugs two interfaces into
+	    # HVMs - an emulated tap device and a paravirt vif device.
+	    # vif-invirtroute (and vif-route, for that matter!) will
+	    # fail when the second one is brought up, because the
+	    # second invocation of 'ip route add' is identical to the
+	    # first (same source and destination IPs) and the kernel
+	    # rejects the new route.
+	    #
+	    # We work around this by adding the routes with different metrics.
+	    # This should work because:
+	    #
+	    # 1) In the case of a pv-aware guest, the kernel will
+	    # unplug the tap interface, which will bring down the tap
+	    # interface's route, leaving only the one via the vif (and
+	    # so the metric shouldn't matter, because it's the only
+	    # route)
+	    #
+	    # 2) In the case of a non-pv-aware guest, the tap route
+	    # (with metric 1) should take precedence over the vif
+	    # route and carry all the traffic.
+
 	    if [ $ipcmd == "add" ]; then
 		case $dev in
 		    vif*)