Re: [PATCH] make vfio and DAX cache work together

[Dr. David Alan Gilbert]

* Alex Williamson (alex.williamson@redhat.com) wrote:
> On Wed, 28 Apr 2021 20:17:23 +0100
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> wrote:
> 
> > * Dev Audsin (dev.devaqemu@gmail.com) wrote:
> > > Thanks Dave for your explanation.
> > > Any suggestions on how to make VFIO not attempt to map into the
> > > unaccessible and unallocated RAM.  
> > 
> > I'm not sure;:
> > 
> > static bool vfio_listener_skipped_section(MemoryRegionSection *section)
> > {
> >     return (!memory_region_is_ram(section->mr) &&
> >             !memory_region_is_iommu(section->mr)) ||
> >            section->offset_within_address_space & (1ULL << 63);
> > }
> > 
> > I'm declaring that region with memory_region_init_ram_ptr;  should I be?
> > it's not quite like RAM.
> > But then I *do* want a kvm slot for it, and I do want it to be accessed
> > by mapping rather htan calling IO functions; that makes me think mr->ram
> > has to be true.
> > But then do we need to add another flag to memory-regions; if we do,
> > what is it;
> >    a) We don't want an 'is_virtio_fs' - it needs to be more generic
> >    b) 'no_vfio' also feels wrong
> > 
> > Is perhaps 'not_lockable' the right thing to call it?
> 
> This reasoning just seems to lead back to "it doesn't work, therefore
> don't do it" rather than identifying the property of the region that
> makes it safe not to map it for device DMA (assuming that's actually
> the case). 

Yes, I'm struggling to get to what that generic form of that property
is, possibly because I've not got an example of another case to compare
it with.

> It's clearly "RAM" as far as QEMU is concerned given how
> it's created, but does it actually appear in the VM as generic physical
> RAM that the guest OS could program to the device as a DMA target?  If
> not, what property makes that so, create a flag for that.  Thanks,

The guest sees it as a PCI-bar; so it knows it's not 'generic physical
RAM' - but can a guest set other BARs (like frame buffers or pmem) as
DMA targets?  If so, how do I distinguish our bar?

Dave

> Alex
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK