Re: [PATCH v2 2/3] KVM: x86: introduce KVM_MEM_PCI_HOLE memory

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Thu, Aug 20, 2020 at 09:46:25PM -0400, Michael S. Tsirkin wrote:
> On Mon, Aug 17, 2020 at 09:32:07AM -0700, Sean Christopherson wrote:
> > On Fri, Aug 14, 2020 at 10:30:14AM -0400, Michael S. Tsirkin wrote:
> > > On Thu, Aug 13, 2020 at 07:31:39PM -0700, Sean Christopherson wrote:
> > > > > @@ -2318,6 +2338,11 @@ static int __kvm_read_guest_page(struct kvm_memory_slot *slot, gfn_t gfn,
> > > > >  	int r;
> > > > >  	unsigned long addr;
> > > > >  
> > > > > +	if (unlikely(slot && (slot->flags & KVM_MEM_PCI_HOLE))) {
> > > > > +		memset(data, 0xff, len);
> > > > > +		return 0;
> > > > > +	}
> > > > 
> > > > This feels wrong, shouldn't we be treating PCI_HOLE as MMIO?  Given that
> > > > this is performance oriented, I would think we'd want to leverage the
> > > > GPA from the VMCS instead of doing a full translation.
> > > > 
> > > > That brings up a potential alternative to adding a memslot flag.  What if
> > > > we instead add a KVM_MMIO_BUS device similar to coalesced MMIO?  I think
> > > > it'd be about the same amount of KVM code, and it would provide userspace
> > > > with more flexibility, e.g. I assume it would allow handling even writes
> > > > wholly within the kernel for certain ranges and/or use cases, and it'd
> > > > allow stuffing a value other than 0xff (though I have no idea if there is
> > > > a use case for this).
> > > 
> > > I still think down the road the way to go is to map
> > > valid RO page full of 0xff to avoid exit on read.
> > > I don't think a KVM_MMIO_BUS device will allow this, will it?
> > 
> > No, it would not, but adding KVM_MEM_PCI_HOLE doesn't get us any closer to
> > solving that problem either.
> 
> I'm not sure why. Care to elaborate?

The bulk of the code in this series would get thrown away if KVM_MEM_PCI_HOLE
were reworked to be backed by a physical page.  If we really want a physical
page, then let's use a physical page from the get-go.

I realize I suggested the specialized MMIO idea, but that's when I thought the
primary motivation was memory, not performance.

> > What if we add a flag to allow routing all GFNs in a memslot to a single
> > HVA?
> 
> An issue here would be this breaks attempts to use a hugepage for this.

What are the performance numbers of hugepage vs. aggressively prefetching
SPTEs?  Note, the unbounded prefetching from the original RFC won't fly,
but prefetching 2mb ranges might be reasonable.

Reraising an earlier unanswered question, is enlightening the guest an
option for this use case?