Re: ovmf / PCI passthrough impaired due to very limiting PCI64 aperture

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Daniel P. Berrangé (berrange@redhat.com) wrote:
> On Wed, Jun 17, 2020 at 05:41:41PM +0100, Dr. David Alan Gilbert wrote:
> > * Eduardo Habkost (ehabkost@redhat.com) wrote:
> > > On Wed, Jun 17, 2020 at 05:17:17PM +0100, Daniel P. Berrangé wrote:
> > > > On Wed, Jun 17, 2020 at 05:04:12PM +0100, Dr. David Alan Gilbert wrote:
> > > > > * Eduardo Habkost (ehabkost@redhat.com) wrote:
> > > > > > On Wed, Jun 17, 2020 at 02:46:52PM +0100, Dr. David Alan Gilbert wrote:
> > > > > > > * Laszlo Ersek (lersek@redhat.com) wrote:
> > > > > > > > On 06/16/20 19:14, Guilherme Piccoli wrote:
> > > > > > > > > Thanks Gerd, Dave and Eduardo for the prompt responses!
> > > > > > > > > 
> > > > > > > > > So, I understand that when we use "-host-physical-bits", we are
> > > > > > > > > passing the *real* number for the guest, correct? So, in this case we
> > > > > > > > > can trust that the guest physbits matches the true host physbits.
> > > > > > > > > 
> > > > > > > > > What if then we have OVMF relying in the physbits *iff*
> > > > > > > > > "-host-phys-bits" is used (which is the default in RH and a possible
> > > > > > > > > machine configuration on libvirt XML in Ubuntu), and we have OVMF
> > > > > > > > > fallbacks to 36-bit otherwise?
> > > > > > > > 
> > > > > > > > I've now read the commit message on QEMU commit 258fe08bd341d, and the
> > > > > > > > complexity is simply stunning.
> > > > > > > > 
> > > > > > > > Right now, OVMF calculates the guest physical address space size from
> > > > > > > > various range sizes (such as hotplug memory area end, default or
> > > > > > > > user-configured PCI64 MMIO aperture), and derives the minimum suitable
> > > > > > > > guest-phys address width from that address space size. This width is
> > > > > > > > then exposed to the rest of the firmware with the CPU HOB (hand-off
> > > > > > > > block), which in turn controls how the GCD (global coherency domain)
> > > > > > > > memory space map is sized. Etc.
> > > > > > > > 
> > > > > > > > If QEMU can provide a *reliable* GPA width, in some info channel (CPUID
> > > > > > > > or even fw_cfg), then the above calculation could be reversed in OVMF.
> > > > > > > > We could take the width as a given (-> produce the CPU HOB directly),
> > > > > > > > plus calculate the *remaining* address space between the GPA space size
> > > > > > > > given by the width, and the end of the memory hotplug area end. If the
> > > > > > > > "remaining size" were negative, then obviously QEMU would have been
> > > > > > > > misconfigured, so we'd halt the boot. Otherwise, the remaining area
> > > > > > > > could be used as PCI64 MMIO aperture (PEI memory footprint of DXE page
> > > > > > > > tables be darned).
> > > > > > > > 
> > > > > > > > > Now, regarding the problem "to trust or not" in the guests' physbits,
> > > > > > > > > I think it's an orthogonal discussion to some extent. It'd be nice to
> > > > > > > > > have that check, and as Eduardo said, prevent migration in such cases.
> > > > > > > > > But it's not really preventing OVMF big PCI64 aperture if we only
> > > > > > > > > increase the aperture _when  "-host-physical-bits" is used_.
> > > > > > > > 
> > > > > > > > I don't know what exactly those flags do, but I doubt they are clearly
> > > > > > > > visible to OVMF in any particular way.
> > > > > > > 
> > > > > > > The firmware should trust whatever it reads from the cpuid and thus gets
> > > > > > > told from qemu; if qemu is doing the wrong thing there then that's our
> > > > > > > problem and we need to fix it in qemu.
> > > > > > 
> > > > > > It is impossible to provide a MAXPHYADDR that the guest can trust
> > > > > > unconditionally and allow live migration to hosts with different
> > > > > > sizes at the same time.
> > > > > 
> > > > > It would be nice to get to a point where we could say that the reported
> > > > > size is no bigger than the physical hardware.
> > > > > The gotcha here is that (upstream) qemu is still reporting 40 by default
> > > > > when even modern Intel desktop chips are 39.
> > > > > 
> > > > > > Unless we want to drop support live migration to hosts with
> > > > > > different sizes entirely, we need additional bits to tell the
> > > > > > guest how much it can trust MAXPHYADDR.
> > > > > 
> > > > > Could we go with host-phys-bits=true by default, that at least means the
> > > > > normal behaviour is correct; if people want to migrate between different
> > > > > hosts with different sizes they should set phys-bits (or
> > > > > host-phys-limit) to the lowest in their set of hardware.
> > > > 
> > > > Is there any sense in picking the default value based on -cpu selection ?
> > > > 
> > > > If user has asked for -cpu host, there's no downside to host-phys-bits=true,
> > > > as the user has intentionally traded off live migration portability already.
> > > 
> > > Setting host-phys-bits=true when using -cpu host makes a lot of
> > > sense, and we could start doing that immediately.
> > > 
> > > > 
> > > > If the user askes for -cpu $MODEL, then could we set phys-bits=NNN for some
> > > > NNN that is the lowest value for CPUs that are capable of running $MODEL ?
> > > > Or will that get too complicated with the wide range of SKU variants, in
> > > > particular server vs desktop CPUs.
> > > 
> > > This makes sense too.  We need some help from CPU vendors to get
> > > us this data added to our CPU model table.  I'm CCing some Intel
> > > and AMD people that could help us.
> > 
> > That bit worries me because I think I agree it's SKU dependent and has
> > been for a long time (on Intel at least) and we don't even have CPU
> > models for all Intel devices. (My laptop for example is a Kaby Lake, 39
> > bits physical).  Maybe it works on the more modern ones where we have
> > 'Icelake-Client' and 'Icelake-Server'.
> 
> Yeah, I think introducing the Client/Server variants was a good idea
> and long overdue.
> 
> For older CPUs, we have a choice of picking a lowest common denominator
> which would likely be a regression for people running on Xeon hosts,
> or we could introduce Client/Server variants for old models which currently
> lack them too.
> 
> I'd tend towards the latter - the Xeon vs non-Xeon SKUs for Intel have
> always been different in various ways, and we've just been lazy using
> one model for both.
> 
> With CPU versioning, that would let us "do the right thing".
> eg Broadwell-Client (physbits=36) / Broadwell-Server (physbits=46)
> Then "-cpu Broadwell" would automatically resolve to whichever of
> Broadwell-Client/Server were supported on the current host taking into
> account its phys-bits too. This would give us live migration safety and
> not limit us to the lowest common denominator.

That does assume there's any consistency within the 'server' range.
Note also that the low end Xeons used to actually be closer to the
Clients.

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK