Re: PCI pass-through vs PoD

["Roger Pau Monné" <roger.pau@citrix.com>]

On Wed, Nov 17, 2021 at 09:39:17AM +0100, Jan Beulich wrote:
> On 13.09.2021 11:02, Jan Beulich wrote:
> > libxl__domain_config_setdefault() checks whether PoD is going to be
> > enabled and fails domain creation if at the same time devices would get
> > assigned. Nevertheless setting up of IOMMU page tables is allowed.

I'm unsure whether allowing enabling the IOMMU with PoD is the right
thing to do, at least for our toolstack.

> > However, when later assigning a device to a domain which has IOMMU page
> > tables, libxl__device_pci_add() does not appear to be concerned of PoD:
> > - xc_test_assign_device() / XEN_DOMCTL_test_assign_device only check the
> >   device for being available to assign,
> > - libxl__device_pci_setdefault() is only concerned about the RDM policy,
> > - other functions called look to not be related to such checking at all.
> 
> I've now verified this to be the case. In fact creating the guest and
> assigning it a device while the guest still sits in the boot loader
> allowed the (oldish) Linux guest I've been using to recognize the device
> (and hence load its driver) even without any hotplug driver. Obviously
> while still in the boot loader ...
> 
> > IMO assignment should fail if pod.count != pod.entry_count,
> 
> ... this holds, and hence assignment should have failed.
> 
> IOW this approach currently is a simple "workaround" to avoid the "PCI
> device assignment for HVM guest failed due to PoD enabled" error upon
> domain creation.
> 
> I'll see if I can find a reasonable place to add the missing check; I'm
> less certain about ...
> 
> > and all PoD
> > entries should be resolved otherwise (whether explicitly by the
> > hypervisor or through some suitable existing hypercalls - didn't check
> > yet whether there are any reasonable candidates - by the tool stack is
> > secondary).
> 
> ... the approach to take here.

I think forcing all entries to be resolved would be unexpected when
assigning a device.

I would rather print a message saying that either the guest must
balloon down to the requested amount of memory, or that all entries
should be resolved (ie: using mem-set to match the mem-max value).

Thanks, Roger.