Hi,

On Thu, Apr 23, 2026 at 02:24:58PM -0400, Dave Voutila wrote:
> >     VM.  Now we have a race:
> >
> >      - the vmd process queues an imsg with that descriptor in config_setvm()
> >        (for the launching VM)
> >      - the vmd process calls in vm_stop() close() on that descriptor
> >        (for the terminating VM)
> >      - when the vmd process calls proc_dispatch() imsgbuf_send() for
> >        imsg queued in config_setvm(), sendmsg() will return EBADF (the
> >        descriptor in the control message is invalid)
> >
> >     By dupping kernfd we can avoid this race.
> 
> This makes sense. Do we need to do this to any others?

I think the others are fine.  However, I'm seeing some other issue:
When rebooting 32 VMs at once, sometimes I see issues with vio(4)
interfaces (vio0: device needs reset) on a few VMs.  Might be a
similar race.  I will investigate.

> In general this is ok dv@

thanks!

Take care.
HJ.

Reply via email to