[RFC/RFT PATCH 0/3] arm64: KVM: work around incoherency with uncached guest mappings

[Catalin Marinas]

On Wed, Mar 04, 2015 at 03:12:12PM +0100, Andrew Jones wrote:
> On Wed, Mar 04, 2015 at 01:43:02PM +0100, Ard Biesheuvel wrote:
> > On 4 March 2015 at 13:29, Catalin Marinas <catalin.marinas at arm.com> wrote:
> > > On Wed, Mar 04, 2015 at 12:50:57PM +0100, Ard Biesheuvel wrote:
> > >> I think we have established that the performance hit is not the
> > >> problem but the correctness is.
> > >
> > > I haven't looked at the performance figures but has anyone assessed the
> > > hit caused by doing cache maintenance in Qemu vs cacheable guest
> > > accesses (and no maintenance)?
> 
> I'm working on a PoC of a QEMU/KVM cache maintenance approach now.
> Hopefully I'll send it out this evening. Tomorrow at the latest.
> Getting numbers of that approach vs. a guest's use of cached memory
> for devices would take a decent amount of additional work, so won't
> be part of that post.

OK.

> I'm actually not sure we should care about
> the numbers for a guest using normal mem attributes for device
> memory - other than out of curiosity. For correctness this issue
> really needs to be solved 100% host-side. We can't rely on a
> guest to do different/weird things, just because it's a guest.
> Ideally guests don't even know that they're guests. (Even if we
> describe the memory as cache-able to the guest, I don't think we
> can rely on the guest believing us.)

I disagree it is 100% a host-side issue. It is a host-side issue _if_
the host tells the guest that the (virtual) device is non-coherent (or,
more precisely, it does not explicitly tell the guest that the device is
coherent). If the guest thinks the (virtual) device is non-coherent
because of information passed by the host, I fully agree that the host
needs to manage the cache coherency.

However, the host could also pass a "dma-coherent" property in the DT
given to the guest and avoid any form of cache maintenance. If the guest
does not honour such coherency property, it's a guest problem and it
needs fixing in the guest. This isn't any different from a real physical
device behaviour.

(there are counter arguments for the latter as well like emulating old
platforms that never had coherency but from a performance/production
perspective, I strongly recommend that guests are passed the
"dma-coherent" property for such virtual devices)

-- 
Catalin