| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2021-05-05 | |||
| 09:31:46 | openstack | Launchpad bug 1900800 in OpenStack Compute (nova) "VGPUs is not recreated on host reboot" [Low,Confirmed] - Assigned to Sylvain Bauza (sylvain-bauza) | |
| 09:31:46 | openstack | Launchpad bug 1900800 in OpenStack Compute (nova) "VGPUs is not recreated on host reboot" [Low,Confirmed] - Assigned to Sylvain Bauza (sylvain-bauza) | |
| 09:31:46 | bauzas | sean-k-mooney: when you're around, if you have a moment, I'm blocked with https://bugs.launchpad.net/nova/+bug/1900800 | |
| 09:31:51 | bauzas | and I need a second pair of eyes | |
| 09:31:51 | bauzas | and I need a second pair of eyes | |
| 10:12:44 | sean-k-mooney | sure just looking now | |
| 10:12:44 | sean-k-mooney | sure just looking now | |
| 10:12:54 | sean-k-mooney | oh that | |
| 10:12:54 | sean-k-mooney | oh that | |
| 10:13:05 | sean-k-mooney | bauzas: how can i help | |
| 10:13:05 | sean-k-mooney | bauzas: how can i help | |
| 10:13:23 | bauzas | sean-k-mooney: lemme explain | |
| 10:13:23 | bauzas | sean-k-mooney: lemme explain | |
| 10:13:36 | bauzas | sean-k-mooney: we need to call placement for getting the RP | |
| 10:13:36 | bauzas | sean-k-mooney: we need to call placement for getting the RP | |
| 10:13:47 | bauzas | but, there are two issues | |
| 10:13:47 | bauzas | but, there are two issues | |
| 10:14:05 | bauzas | first, the method looking at the mdevs is called by the libvirt init_host() | |
| 10:14:05 | bauzas | first, the method looking at the mdevs is called by the libvirt init_host() | |
| 10:14:49 | bauzas | which is called *before* the RT update_avalaible_resource() | |
| 10:14:49 | bauzas | which is called *before* the RT update_avalaible_resource() | |
| 10:15:04 | sean-k-mooney | right so you can just use the instance uuid to call https://docs.openstack.org/api-ref/placement/?expanded=list-allocations-detail#list-allocations | |
| 10:15:04 | sean-k-mooney | right so you can just use the instance uuid to call https://docs.openstack.org/api-ref/placement/?expanded=list-allocations-detail#list-allocations | |
| 10:15:34 | sean-k-mooney | that will give you all the allocations by RP | |
| 10:15:34 | sean-k-mooney | that will give you all the allocations by RP | |
| 10:15:45 | sean-k-mooney | for a given consomeer i.e. vm | |
| 10:15:46 | sean-k-mooney | for a given consomeer i.e. vm | |
| 10:15:55 | bauzas | sean-k-mooney: so, in case of a vGPU reshape, the allocation will tell that the RP is the root one | |
| 10:15:55 | bauzas | sean-k-mooney: so, in case of a vGPU reshape, the allocation will tell that the RP is the root one | |
| 10:16:10 | bauzas | (as the reshape is done *after*) | |
| 10:16:10 | bauzas | (as the reshape is done *after*) | |
| 10:16:24 | bauzas | we could tell it's OK | |
| 10:16:24 | bauzas | we could tell it's OK | |
| 10:16:33 | bauzas | but then we have a second concern | |
| 10:16:33 | bauzas | but then we have a second concern | |
| 10:16:55 | bauzas | we don't accept the libvirt driver to call directly Placement API | |
| 10:16:55 | bauzas | we don't accept the libvirt driver to call directly Placement API | |
| 10:17:06 | bauzas | instead, we directly pass allocations to the methods | |
| 10:17:06 | bauzas | instead, we directly pass allocations to the methods | |
| 10:17:40 | bauzas | we could use the provider tree that we have in the driver | |
| 10:17:40 | bauzas | we could use the provider tree that we have in the driver | |
| 10:17:47 | sean-k-mooney | right so either we have to allow the driver to do that or we need to modify init_host to accpet a dict of allocations keyed by instance uuid | |
| 10:17:47 | sean-k-mooney | right so either we have to allow the driver to do that or we need to modify init_host to accpet a dict of allocations keyed by instance uuid | |
| 10:18:11 | bauzas | but, this variable is not yet created by init_host() | |
| 10:18:11 | bauzas | but, this variable is not yet created by init_host() | |
| 10:18:37 | bauzas | sean-k-mooney: I'm not sure we could accept this | |
| 10:18:37 | bauzas | sean-k-mooney: I'm not sure we could accept this | |
| 10:18:50 | bauzas | (allowing the driver to directly call placement) | |
| 10:18:50 | bauzas | (allowing the driver to directly call placement) | |
| 10:19:13 | bauzas | what we could do is maybe the latter | |
| 10:19:13 | bauzas | what we could do is maybe the latter | |
| 10:19:35 | sean-k-mooney | im just pulling up the code in the compute manager now | |
| 10:19:35 | sean-k-mooney | im just pulling up the code in the compute manager now | |
| 10:20:09 | sean-k-mooney | https://github.com/openstack/nova/blob/master/nova/compute/manager.py#L1441-L1443 | |
| 10:20:09 | sean-k-mooney | https://github.com/openstack/nova/blob/master/nova/compute/manager.py#L1441-L1443 | |
| 10:20:19 | sean-k-mooney | so this is where we end up callint ingo the dirver eventurally right | |
| 10:20:19 | sean-k-mooney | so this is where we end up callint ingo the dirver eventurally right | |
| 10:20:26 | sean-k-mooney | from _init_instanc | |
| 10:20:26 | sean-k-mooney | from _init_instanc | |
| 10:21:11 | sean-k-mooney | for example thsi is where we plug vifs https://github.com/openstack/nova/blob/7953c0197d1a4466cb5b78070d47626c92f9db6e/nova/compute/manager.py#L1137 | |
| 10:21:11 | sean-k-mooney | for example thsi is where we plug vifs https://github.com/openstack/nova/blob/7953c0197d1a4466cb5b78070d47626c92f9db6e/nova/compute/manager.py#L1137 | |
| 10:21:13 | bauzas | not really | |
| 10:21:13 | bauzas | not really | |
| 10:21:24 | bauzas | rather here https://github.com/openstack/nova/blob/master/nova/compute/manager.py#L1413 | |
| 10:21:24 | bauzas | rather here https://github.com/openstack/nova/blob/master/nova/compute/manager.py#L1413 | |
| 10:21:58 | sean-k-mooney | right but we do not create the instance until https://github.com/openstack/nova/blob/7953c0197d1a4466cb5b78070d47626c92f9db6e/nova/compute/manager.py#L1204 | |
| 10:21:58 | sean-k-mooney | right but we do not create the instance until https://github.com/openstack/nova/blob/7953c0197d1a4466cb5b78070d47626c92f9db6e/nova/compute/manager.py#L1204 | |
| 10:22:24 | bauzas | hmmm, I see | |
| 10:22:24 | bauzas | hmmm, I see | |
| 10:22:39 | sean-k-mooney | sorry a few lines above https://github.com/openstack/nova/blob/7953c0197d1a4466cb5b78070d47626c92f9db6e/nova/compute/manager.py#L1195 | |
| 10:22:39 | sean-k-mooney | sorry a few lines above https://github.com/openstack/nova/blob/7953c0197d1a4466cb5b78070d47626c92f9db6e/nova/compute/manager.py#L1195 | |
| 10:22:51 | sean-k-mooney | so what im thinking is we can extend this function to create the mdevs | |
| 10:22:51 | sean-k-mooney | so what im thinking is we can extend this function to create the mdevs | |
| 10:23:02 | sean-k-mooney | this happesn after the driver init_host is run | |
| 10:23:02 | sean-k-mooney | this happesn after the driver init_host is run | |
| 10:23:18 | sean-k-mooney | so the reshapes would already have run | |
| 10:23:18 | sean-k-mooney | so the reshapes would already have run | |
| 10:23:23 | bauzas | instead of having init_host() calling https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L789 | |
| 10:23:23 | bauzas | instead of having init_host() calling https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L789 | |
| 10:23:36 | bauzas | we would use init_instance() | |
| 10:23:36 | bauzas | we would use init_instance() | |
| 10:23:40 | sean-k-mooney | yep | |
| 10:23:43 | bauzas | this could work | |
| 10:23:43 | bauzas | this could work | |
| 10:24:00 | bauzas | * bauzas needs to look at what we have in init_instance() | |
| 10:24:17 | bauzas | but then, we could ask its allocations | |
| 10:24:17 | bauzas | but then, we could ask its allocations | |
| 10:24:27 | bauzas | as it's within the compute manager | |
| 10:24:27 | bauzas | as it's within the compute manager | |
| 10:24:32 | sean-k-mooney | it basicaly just plugs vifs and cleans up deleted instance ectra | |
| 10:24:32 | sean-k-mooney | it basicaly just plugs vifs and cleans up deleted instance ectra | |
| 10:24:36 | sean-k-mooney | yep | |
| 10:24:56 | sean-k-mooney | its in the compute manager so we can have it get the allocations | |
| 10:24:56 | sean-k-mooney | its in the compute manager so we can have it get the allocations | |
| 10:24:59 | bauzas | and then, pass the allocation to the virt driver which would use the related RP name for knowing which type | |
| 10:24:59 | bauzas | and then, pass the allocation to the virt driver which would use the related RP name for knowing which type | |
| 10:25:09 | sean-k-mooney | we might need to extend the dirver interface and add an optional funciton | |
| 10:25:09 | sean-k-mooney | we might need to extend the dirver interface and add an optional funciton | |
| 10:25:25 | sean-k-mooney | which we woudl call that take the instnace and allcoations | |
| 10:25:25 | sean-k-mooney | which we woudl call that take the instnace and allcoations | |