Open Stack

On Thu Oct 14 2021 11:37:43 GMT-0700 (Pacific Daylight Time), Gustavo 
Faganello Santos <gustavofaganello.santos at windriver.com> wrote:
> Hello, everyone!
> 
> I'm working on a solution for Nova to reattach previously used mediated
> devices (vGPU instances, in my case) to VMs coming back from suspension,
> which seems to have been left on hold in the past [1] because of an old
> libvirt limitation, and I'm having a bit of a hard time doing so, since
> I'm not too familiar with the repo.
> 
> I have tried creating a function that does the opposite of the mdev
> detach function, but the get_all_devices method seems to return an empty
> list when looking for mdevs at the moment of resuming the VM. Looking at
> the instance's XML file, I noticed that the mdev property remains while
> the VM is suspended, but it disappears AFTER the whole resume function
> is executed. I'm failing to understand why the mdev list returns empty,
> even though the mdev property exists in the instance's XML, and also why
> the mdev is removed from the XML after the resume function is executed.
> 
> With that in mind, does anyone know if there's been any attempt to solve
> this issue since it was left on hold? If not, is there anything I should
> know while I attempt to do so?

I'm not sure whether this will be helpful but there is similar (or 
adjacent?) work currently in progress to handle the case of recreating 
mediated devices after a compute host reboot [2][3]. The launchpad bug 
contains some info on workarounds for this case and the proposed patch 
pulls allocation information from the placement service to recreate the 
mdevs.

-melanie

[2] https://bugs.launchpad.net/nova/+bug/1900800
[3] https://review.opendev.org/c/openstack/nova/+/810220

> Thanks in advance.
> Gustavo
> 
> [1]
> https://opendev.org/openstack/nova/src/branch/master/nova/virt/libvirt/driver.py#L8007
>