Set the arp_notify sysctl on new interfaces, and set all sysctls before the interface...

[invirt/packages/invirt-xen-config.git] / vif-invirtroute

diff --git a/vif-invirtroute b/vif-invirtroute
index f484f87..ffa7c73 100755 (executable)
--- a/vif-invirtroute
+++ b/vif-invirtroute
@@ -23,7 +23,6 @@
 # type read to determine if the device is ioemu
 
 #============================================================================
-
 dir=$(dirname "$0")
 . "$dir/vif-common.sh"
 
@@ -32,10 +31,11 @@ dev=${dev:-${vif}}
 
 case "$command" in
     online|add)
-        ifconfig ${dev} ${main_ip} netmask 255.255.255.255 up
         echo 1 >/proc/sys/net/ipv4/conf/${dev}/proxy_arp
-       echo 1 >/proc/sys/net/ipv4/conf/${dev}/rp_filter 
-       xenstore-write "$XENBUS_PATH/feature-gso-tcpv4" 0
+        echo 1 >/proc/sys/net/ipv4/conf/${dev}/arp_notify
+        echo 1 >/proc/sys/net/ipv4/conf/${dev}/rp_filter
+        ifconfig ${dev} ${main_ip} netmask 255.255.255.255 up
+        xenstore-write "$XENBUS_PATH/feature-gso-tcpv4" 0
         if [ x${qemu_online} != xyes ]; then
           ethtool -K ${dev} tx off
         fi
@@ -55,7 +55,38 @@ 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
-           ${cmdprefix} ip route ${ipcmd} ${addr} dev ${dev} src ${main_ip}
+           # 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*)
+                       metric="metric 2"
+                       ;;
+                   tap*)
+                       metric="metric 1"
+                       ;;
+               esac
+           fi
+           ${cmdprefix} ip route ${ipcmd} ${addr} dev ${dev} src ${main_ip} $metric
            case "$command" in
                online|add)
                    arpspoof -i $(invirt-getconf xen.iface) -t 18.181.0.1 ${addr}&