http://wiki.libvirt.org/page/Guest_can_reach_outside_network,_but_can%27t_reach_host_%28macvtap%29
macvtap interfaces (type='direct' - see the libvirt documentation on the topic) can be useful even when not connecting to a VEPA or VNLINK capable switch - setting the mode of such an interface to 'bridge' will allow the guest to be directly connected to the physical network in a very simple manner without the setup hassles (or NetworkManager incompatibility) that accompany use of a traditional host bridge device.
However, once a guest has been configured to use a "type='direct'" network interface (a.k.a. macvtap), users will commonly be surprised that the guest is able to communicate with other guests, and also with other external hosts on the network, but cannot communicate with the virt host on which the guest in question lives.
This is not a bug, it is the defined behavior of macvtap - due to the way that the host's physical ethernet is attached to the macvtap bridge, traffic into that bridge from the guests that is forwarded to the physical interface cannot be bounced back up to the host's IP stack (and also, traffic from the host's IP stack that is sent to the physical interface cannot be bounced back up to the macvtap bridge for forwarding to the guests.)
One possible method of eliminating this problem would be to create a separate macvtap interface for host use, and give it the IP configuration previously on the physical ethernet (see this page for an example of how to manually configure an interface on the physical host to use macvtap, and this page for a script) - in this way, the host would be an equal peer attached to the macvlap bridge, and thus guest and host could communicate directly.
However, this solution has two problems - 1) it reintroduces just as much complexity to the configuration as would setting up a traditional Linux host bridge and 2) Just as NetworkManager currently doesn't understand bridge devices, it also doesn't understand macvtap devices, so NetworkManager would be unable to monitor the online state of the macvtap interface, and would give erroneous reports about the online status of the host. In other words, it's really no better than just using a traditional host bridge (with the added problem that even the traditional methods of network configuration (e.g. initscripts on Fedora and RHEL) don't support configuration of a macvtap device).
There is an alternate solution which preserves NetworkManager compatibility while allowing guest and host to directly communicate. In short, the solution is use libvirt to create an isolated network, and give each guest a second interface that is connected to this network; host<-->guest communication will then take place over the isolated network.
1) Save the following XML to /tmp/isolated.xml:
<network> <name>isolated</name> <ip address='192.168.254.1' netmask='255.255.255.0'> <dhcp> <range start='192.168.254.2' end='192.168.254.254' /> </dhcp> </ip> </network>
(if the 192.168.254.0/24 network is already in use elsewhere on your network, you can choose a different network).
2) Create the network, set it to autostart, and start it:
virsh net-define /tmp/isolated.xml virsh net-autostart isolated virsh net-start isolated
3) Edit (using "virsh edit $guestname") the configuration of each guest that uses direct (macvtap) for its network connection and add a new <interface> in the <devices> section similar to the following:
<interface type='network'> <source network='isolated'/> <model type='virtio'/> <-- This line is optional. </interface>
4) shutdown, then restart each of these guests.
The guests will now be able to reach the host at the address 192.168.254.1, and the host will be able to reach the guests at whatever IP address they acquired from DHCP (alternately you can manually configure them). Since this new network is isolated to only the host and guests, all other communication from the guests will use the macvtap interface.