On Thu, Dec 08, 2022 at 07:01:57PM +0000, Sean Christopherson wrote:
> On Thu, Dec 08, 2022, Ricardo Koller wrote:
> > On Thu, Dec 08, 2022 at 12:37:23AM +0000, Oliver Upton wrote:
> > > On Thu, Dec 08, 2022 at 12:24:20AM +0000, Sean Christopherson wrote:
> > > > > Even still, that's just a kludge to make ucalls work. We have other
> > > > > MMIO devices (GIC distributor, for example) that work by chance since
> > > > > nothing conflicts with the constant GPAs we've selected in the tests.
> > > > >
> > > > > I'd rather we go down the route of having an address allocator for the
> > > > > for both the VA and PA spaces to provide carveouts at runtime.
> > > >
> > > > Aren't those two separate issues? The PA, a.k.a. memslots space, can be solved
> > > > by allocating a dedicated memslot, i.e. doesn't need a carve. At worst, collisions
> > > > will yield very explicit asserts, which IMO is better than whatever might go wrong
> > > > with a carve out.
> > >
> > > Perhaps the use of the term 'carveout' wasn't right here.
> > >
> > > What I'm suggesting is we cannot rely on KVM memslots alone to act as an
> > > allocator for the PA space. KVM can provide devices to the guest that
> > > aren't represented as memslots. If we're trying to fix PA allocations
> > > anyway, why not make it generic enough to suit the needs of things
> > > beyond ucalls?
> >
> > One extra bit of information: in arm, IO is any access to an address (within
> > bounds) not backed by a memslot. Not the same as x86 where MMIO are writes to
> > read-only memslots. No idea what other arches do.
>
> I don't think that's correct, doesn't this code turn write abort on a RO memslot
> into an io_mem_abort()? Specifically, the "(write_fault && !writable)" check will
> match, and assuming none the the edge cases in the if-statement fire, KVM will
> send the write down io_mem_abort().
You are right. In fact, page_fault_test checks precisely that: writes on
RO memslots are sent to userspace as an mmio exit. I was just referring
to the MMIO done for ucall.
Having said that, we could use ucall as writes on read-only memslots
like what x86 does.
>
> gfn = fault_ipa >> PAGE_SHIFT;
> memslot = gfn_to_memslot(vcpu->kvm, gfn);
> hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
> write_fault = kvm_is_write_fault(vcpu);
> if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
> /*
> * The guest has put either its instructions or its page-tables
> * somewhere it shouldn't have. Userspace won't be able to do
> * anything about this (there's no syndrome for a start), so
> * re-inject the abort back into the guest.
> */
> if (is_iabt) {
> ret = -ENOEXEC;
> goto out;
> }
>
> if (kvm_vcpu_abt_iss1tw(vcpu)) {
> kvm_inject_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
> ret = 1;
> goto out_unlock;
> }
>
> /*
> * Check for a cache maintenance operation. Since we
> * ended-up here, we know it is outside of any memory
> * slot. But we can't find out if that is for a device,
> * or if the guest is just being stupid. The only thing
> * we know for sure is that this range cannot be cached.
> *
> * So let's assume that the guest is just being
> * cautious, and skip the instruction.
> */
> if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
> kvm_incr_pc(vcpu);
> ret = 1;
> goto out_unlock;
> }
>
> /*
> * The IPA is reported as [MAX:12], so we need to
> * complement it with the bottom 12 bits from the
> * faulting VA. This is always 12 bits, irrespective
> * of the page size.
> */
> fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
> ret = io_mem_abort(vcpu, fault_ipa);
> goto out_unlock;
> }
