[fedora-virt] vfio-vga-reset branches build problems

[Alex Williamson]

On Mon, 2013-07-08 at 17:20 +0200, Mario Ceresa wrote:
> Dear all,
> I'm trying to compile Alex's vfio-vga-reset branches.
> 
> However, I have the following problems:
> 
> * I cannot apply the changes on top of fedora kernel 3.9.9-301.fc19. Doing
> on 3.9.8-300.fc19 is fine though.
> 
> * I cannot compile qemu-vfio using fedora's spec (qemu-1.4.2-4.fc19 and
> newer). I get:
> 
> qmp.c: In function 'qmp_cpu_add':
> qmp.c:114:24: error: 'QEMUMachine' has no member named 'hot_add_cpu'
>      if (current_machine->hot_add_cpu) {
> 
> Anybody has any suggestions?

This is really not surprising at all, these branches are just a
development snapshot.  The qemu changes are based on something much
closer to qemu 1.5 and the kernel changes touch core PCI code, so are
bound to get broken as the kernel progresses.

FWIW, I'm still stuck on how to reliably do the PCI secondary bus reset.
In most cases it works fine so long as the guest is properly shut down
or rebooted.  However if I issue a system_reset to the guest while it's
running, the PCI bus never comes back from reset.  The link status of
the root port reports that it's stuck in training mode.  A host reboot
is required to recover the bus.  I assume it has to be related to
ongoing DMA across the bridge during reset, but disabling the devices
and the bridge via the command register and even stopping the guest for
reset isn't able to prevent the bus dying.  Suggestions or reports
welcome.  I can reliably reproduce by simply calling system_reset on an
idle Win7/x64 guest with HD5450 assigned and Catalyst drivers loaded
(AMD/990FX host).  Thanks,

Alex