I hit a performance issue playing with Mesa running in a kvm guest (MeeGo 1.0, as it happens).  TCP connections to the X server in the outer host were showing very slow performance, and seemed to be highly fill limited on the VM where they were not on the host (obviously transport bandwidth shouldn't change fill rate for the same geometry), even on applications like glxgears which should be almost identically performant.

Hacking at the glxgears source to set the "direct" parameter to FALSE restored the performance characteristics I expected.

My guess as to what is happening here is that the presence of DRI drivers in the both the guest OS and the target X server is fooling Mesa into interpreting the "direct" parameter as possible, even when the server is on the other side of a TCP socket and "direct" rendering must take a software path.