yes, the description of KVM is more or less accurate. One thing is that KVM supports a couple of optimizations to the usual behavior of MMIO. One is “ioeventfd”, where writing a value to a memory address or I/O port will trigger an eventfd in qemu-kvm. This is faster than a normal MMIO cycle, and very similar to a Xen event channel. The second is a special paravirtualized end-of-interrupt command that is only a single instruction and reduces the number of exits for an interrupt from 2 to 1. So it is somewhat between PV-on-HVM and PVHVM.

In general, the philosophy of KVM is that “if it is slow, hardware vendors will make it fast” and it is not worthwhile to paravirtualize it. It happened with hardware-assisted paging, with the APIC TPR (which made Windows XP very slow under Xen or KVM with the first implementations of VMX), and it is now happening with the rest of the APIC.