Re: [PATCH] pci: Refuse to hotplug PCI Devices when the Guest OS is not ready

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Oct 28, 2020 at 04:39:45PM +0100, Igor Mammedov wrote:
> On Wed, 28 Oct 2020 14:31:35 +1100
> David Gibson <dgibson@redhat.com> wrote:
> 
> > On Tue, 27 Oct 2020 13:54:26 +0100
> > Igor Mammedov <imammedo@redhat.com> wrote:
> > 
> > > On Tue, 27 Oct 2020 07:26:44 -0400
> > > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > 
> > >  [...]  
> > >  [...]  
> > >  [...]  
> > > [...]
> > >  [...]    
> > > > > 
> > > > > It certainly shouldn't wait an unbounded time.  But a wait with timeout
> > > > > seems worth investigating to me.      
> > > racy, timeout is bound to break once it's in overcommited env.  
> > 
> > Hm.  That's no less true at the management layer than it is at the qemu
> > layer.
> true, but it's user policy which is defined by user not by QEMU.
> 
> > 
> > > > If it's helpful, I'd add a query to check state
> > > > so management can figure out why doesn't guest see device yet.    
> > > that means mgmt would have to poll it and forward it to user
> > > somehow.  
> > 
> > If that even makes sense.  In the case of Kata, it's supposed to be
> > autonomously creating the VM, so there's nothing meaningful it can
> > forward to the user other than "failed to create the container because
> > of some hotplug problem that means nothing to you".
> > 
> > >  [...]  
> > > I have more questions wrt the suggestion/workflow:
> > > * at what place would you suggest buffering it?
> > > * what would be the request in this case, i.e. create PCI device anyways
> > > and try to signal hotplug event later?
> > > * what would baremethal do in such case?
> > > * what to do in case guest is never ready, what user should do in such case?
> > > * can be such device be removed?
> > > 
> > > not sure that all of this is worth of the effort and added complexity.
> > > 
> > > alternatively:
> > > maybe ports can send QMP events about it's state changes, which end user would
> > > be able to see + error like in this patch.
> > > 
> > > On top of it, mgmt could build a better UIx, like retry/notify logic if
> > > that's what user really wishes for and configures (it would be up to user to
> > > define behaviour).  
> > 
> > That kind of makes sense if the user is explicitly requesting hotplugs,
> > but that's not necessarily the case.
> user doesn't have to be a human, it could be some mgmt layer that would
> automate retry logic, depending on what actually user needs for particular task
> (i.e. fail immediately, retry N time then fail, retry with time out - then fail,
> don't care - succeed, ...). The point is for QEMU to provide means for mgmt to
> implement whatever policy user would need.

We are not coming up with new APIs here. Let's make existing ones
work reliably first. We can talk about a flag where it fails
instead of deferring hotplug, separately.

> PS:
> but then, I know close to nothing about PCI, so all of above might be nonsense.