No one objected pcie support itself. But patch quality is another issue. Someone commented that the merge might be slow because pcie support is not urgent.
Some people are interested in multi pci buses instead of pcie itself because just 32 (- some) slots is too few. The irusage model is a VM as file server with high end storage. They want many LUs to be attached to host and export them to guest VM.
He talked about SPICE. The main part is QXL,virtual display device, and its device driver.
It supports multi display and he demoed it.
Interesting future TODO is network tunneling for local printing and USB forwarding.USB forwarding is for sharing files on local usb memory. This soundsvery useful.
Developing tests for the KVM auto test
Lucas Meneghel Rodrigues
KVM auto test is fork of autotest. It runs 1 VM and runs tests in the VM. At the moment, it isn't able to create multiple VMs and orchestrate them. Someone suggested to run KVM auto test in KVM. That sound good
for nested virtualization.
This session should be about threading qemu, but the main theme was splitting maintainer ship.
KVM vcpu part is really threaded, but device emulation is single threaded. However the arbitration between KVM(virtualization) and TCG(embedded)is difficult. So he claimed to split out qemu maintainer ship and repository into libqemu-*.so.
QEMU's new device model qdev
He explained qemu new device model, qdev. I think it's been a while since it was introduced so that it is already well understood. Nothing special.
Migration How to hop from machine to machine without losing state
This session is about new machine save/load framework called vmstate. It's much cleaner than old ad-hoc save/load handlers. Live migration is difficult with device/ram hot plug.
Sheepdog: distributed storage system for QEMU
New part from the last presentation at LCA is benchmark with 256VM. It scaled well with 64 servers. Someone asked about disk scrub.
A Quest Against Time
Time is difficult issue to handle. Especially old OSes. The eventual solution is KVM clock(paravirtualize clock).
Paravirtualized DMA using IOMMU emulation
This is about to minimize pinning guest pages. The current IOMMU can't recover from DMA page fault, so that all guest pages must pinned. To overcome this restriction, paravirtualized guest DMA API.