[nova][dev] Reattaching mediated devices to instance coming back from suspended state
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? Thanks in advance. Gustavo [1] https://opendev.org/openstack/nova/src/branch/master/nova/virt/libvirt/drive...
On Thu Oct 14 2021 11:37:43 GMT-0700 (Pacific Daylight Time), Gustavo Faganello Santos <gustavofaganello.santos@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/drive...
On 14/10/2021 16:02, melanie witt wrote:
[Please note: This e-mail is from an EXTERNAL e-mail address]
On Thu Oct 14 2021 11:37:43 GMT-0700 (Pacific Daylight Time), Gustavo Faganello Santos <gustavofaganello.santos@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.
Thank you for your reply! I'm aware of that work, but I'm afraid that it unfortunately does not relate too much to what I'm going for.
-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/drive...
participants (2)
-
Gustavo Faganello Santos
-
melanie witt