[PATCH v4 17/19] arm64: KVM: add SGI generation register emulation

[Christoffer Dall]

On Wed, Dec 03, 2014 at 05:50:51PM +0000, Andre Przywara wrote:
> On 30/11/14 08:45, Christoffer Dall wrote:
> > On Fri, Nov 28, 2014 at 03:40:12PM +0000, Andre Przywara wrote:
> >> Hej Christoffer,
> >>
> >> On 25/11/14 11:03, Christoffer Dall wrote:
> >>> Hi Andre,
> >>>
> >>> On Mon, Nov 24, 2014 at 04:37:58PM +0000, Andre Przywara wrote:
> >>>> Hi,
> >>>>
> >>>> On 23/11/14 15:08, Christoffer Dall wrote:
> >>>>> On Fri, Nov 14, 2014 at 10:08:01AM +0000, Andre Przywara wrote:
> >>>>>> While the generation of a (virtual) inter-processor interrupt (SGI)
> >>>>>> on a GICv2 works by writing to a MMIO register, GICv3 uses the system
> >>>>>> register ICC_SGI1R_EL1 to trigger them.
> >>>>>> Trap that register on ARM64 hosts and handle it in a new handler
> >>>>>> function in the GICv3 emulation code.
> >>>>>
> >>>>> Did you reorder something or does my previous comment still apply that
> >>>>> you're not enabling trapping yet, you're just adding the handler - those
> >>>>> are two different things.
> >>>>
> >>>> Yes, I can fix the wording.
> >>>>
> >>>>> You sort of left my question about access_gic_sgi() not checking if the
> >>>>> gicv3 is presetn hanging from the last thread, but I think I'm
> >>>>> understanding properly now, that as long as you're not setting the
> >>>>> ICC_SRE_EL2.Enable = 1, then we'll never get here, right?
> >>>>
> >>>> Right, that is the idea. Just to make sure that I got this right from
> >>>> the discussion the other day: We will not trap to EL2 as long as
> >>>> ICC_SRE_EL2.Enable is 0 - which it should still be at this point, right?
> >>>
> >>> No, when ICC_SRE_EL2.Enable is 0, then Non-secure EL1 access to
> >>> ICC_SRE_EL1 trap to EL2 (See Section 5.7.39 in the spec), which means
> >>> that accesses to the ICC_SGIx registers will cause an undefined
> >>> exception in the guest because we set ICC_SRE_EL1.SRE to 0 for the
> >>> guest and the guest cannot change this.
> >>>
> >>> Now, when we set ICC_SRE_EL2.Enable to 1, then the guest can set
> >>> ICC_SRE_EL1.SRE to 1 (and we also happen to reset it to 1), and we will
> >>> indeed trap on guest access to the ICC_SGIx registers, because all
> >>> virtual accesses to these registers trap.
> >>>
> >>> (Going back and checking where 'virtual accesses' is defined in the spec
> >>> left me somewhere without any results, but I am guessing that because we
> >>> set the ICH_HCR_EL2.En to 1, all accesses will be deemed virtual
> >>> accesses, maybe the spec should be clarfied on this matter?).
> >>>
> >>> Anyhow, to get back to my original question, getting here requires
> >>> a situation where the guest copy of the ICC_SRE_EL1.SRE is 1, which we
> >>> only allow when we have properly initialized the GICv3 data structures.
> >>
> >> So to summarize (and check) this: There is no real issue at this point?
> >> And the code is totally fine after 19/19?
> > 
> > There is no issue at this point, no.
> > 
> >>
> >> Would this kind of problem actually matter _inside_ a patch series? To
> >> trigger an issue, we would need a bogus guest and bogus userland
> >> (because at this point neither of them would see/inject a GICv3 FDT
> >> node). I'd assume that running a kernel at this point is just for
> >> debugging/bisecting? Where you wouldn't care about every corner case of
> >> execution?
> > 
> > The argument about bogus guests / fdts should *never* be considered in
> > the context of these discussions.  If we have code that looks like the
> > guest can kill the host, or do a NULL pointer dereference, then we need
> > to address it.
> > 
> > Your point about it being inside a patch series, sure, it's unlikely
> > that people will run this, but I'm reviewing this patch right now, and
> > honestly not considering how this changes in the subsequent patch.  For
> > this sort of thing, if we were leaving a gaping hole open, that would at
> > least require a clear note in the commit message on why we're doing it.
> 
> I see, makes sense.
> So I thought about adding a line like this to the very beginning of
> vgic_v3_dispatch_sgi(). This would cover all cases of spurious traps.
> Does that sound useful as a security precaution (though unneeded as
> described)?
> Shall there be a warning before the return?
> 
> +	/* only valid for an initialized VGICv3 */
> +	if (!vgic_initialized(kvm)  ||
> +	    kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V3)
> +		return;
> 

If anything you should have a BUG_ON(), but especially when you've
tested this, it won't happen.

> > Hopefully you understood and agreed with my deduction about the various
> > SRE settings above though?
> 
> Yes, I got this. We are safe as long as ICC_SRE_EL1.SRE is 0, which is
> true until patch 19/19 allows userland to request a GICv3 guest, which
> will force it to 1.
> I also tested this explicitly, starting with patch 17/19 (for the host
> kernel) and going over the remaining two as well. Starting a guest with
> GICv2 and accessing ICC_SRE_EL1 and ICC_SGI1R_EL1 from a custom module
> inside the guest will always keep ICC_SRE_EL1.SRE to 0 (thanks to your
> recent trap patch), accesses to ICC_SGI1R_EL1 provoke an #UNDEF
> exception in the guest. The host was never bothered.
> Creating a guest with a GICv3 was only successful after patch 19/19, and
> ICC_SRE_EL1.SRE couldn't be cleared.
> 
> So I consider this topic done.
> 
Indeed!
-Christoffer