update VM phase by checking domain status and reason #149
Conversation
|
Can one of the admins verify this patch? |
|
I could think of a privileged pod which runs on the host pid namespace and just greps for the qemu processes for the vm (something like ps aux | grep qemu | grep and sends kill -9 to that process. We can then just launch that pod on the node of the VM. That would also be useful for later, for random kill, to verify everything is working ... You can have a look at the libvirt daemon set to see how to run a privileged pod with host pidnamespace, |
|
@rmohr Cool idea, like the future extensibility. Looking into it. |
|
@mpolednik looks good so far. Could you rebase it? @admiyo changed the method calls on the controllers a little bit. Should not be hard to update. |
|
@rmohr noticed when working on it today :( I'd love to add the "crash" test in the same PR, update coming today at night or someday tomorrow. |
pkg/virt-handler/domain.go
Outdated
| } | ||
| obj, vmExists, err := d.vmStore.GetByKey(key.(string)) | ||
| if err != nil { | ||
| queue.AddRateLimited(key) | ||
| } else if !vmExists || obj.(*v1.VM).GetObjectMeta().GetUID() != domain.GetObjectMeta().GetUID() { | ||
| // The VM is not in the vm cache, or is a VM with a differend uuid, tell the VM controller to investigate it | ||
| d.vmQueue.Add(key) | ||
| } else { | ||
| setVmPhaseForStatusReason(domain, obj.(*v1.VM), d.restClient) |
There was a problem hiding this comment.
You also have to check if the VM has the same uid like the Domain, otherwise you are setting a potentially new posted VM to stopped, although it was not even scheduled ...
There was a problem hiding this comment.
Oh and one more thing, you have to check for the error. If an error is returned, you have to do a queue.AddRateLimited(key).
There was a problem hiding this comment.
About the first comment: shouldn't that be handled by else-if clause above?
Can help debugging stuff later on.
Uses new pod to kill VMs.
|
ok to test |
add log verbosity
Second attempt since the event changes. The only way to really see the phase change is to actually crash the domain - there is no phase change when VM TPR is deleted (so is the domain, but at that point there is nothing to update).
@rmohr Any idea how to nicely test this? Currently experimenting with pkill -9 qemu, which may be a bit strong. Pretty much need anything to send kill -9 or equivalent.
Closes #75