Several times over the last few days, I've lost network connectivity 
from one of my guests. This has happened only during interactive 
sessions in which I take an action resulting in a large screen update. I 
have tried flood pinging (only with the default, small packet size), and 
not been able to reproduce under those circumstances.

This guest is running a current RHEL5/CentOS kernel (2.6.18-53.1.13.el5) 
with clocksource=acpi_pm on the command line. The host side of the 
network is a tap device, which is joined to a bridge. Other VMs on the 
bridge still have working bidirectional networking, and dmesg on the 
host shows the relevant port on the bridge in forwarding state.

Rebooting the guest without shutting down the kvm instance does not 
resolve the issue. Powering down the VM and starting a new kvm instance 
*does* resolve the issue.

Within the guest, tcpdump sees both incoming and outgoing traffic; 
however, on the host, only traffic going *to* the guest is visible; 
traffic the guest attempts to send is not visible.

When attempting to send, none of the counters (RX packets/TX 
packets/etc) increase on the emulated e1000 device within the guest; RX 
bytes and TX bytes are both 0, and all the error counters are likewise 
zeroed. The e1000 module can be reloaded without any visible errors.

Where should I start in attempting to debug this?