[PATCH] drm/msm/a6xx: Skip gpu secure fw load in EL2 mode

[Mark Rutland]

On Thu, Dec 12, 2024 at 08:50:12AM +0000, Marc Zyngier wrote:
> On Thu, 12 Dec 2024 05:31:00 +0000,
> Pavan Kondeti <quic_pkondeti at quicinc.com> wrote:
> > 
> > On Wed, Dec 11, 2024 at 10:40:02AM +0000, Marc Zyngier wrote:
> > > On Wed, 11 Dec 2024 00:37:34 +0000,
> > > Pavan Kondeti <quic_pkondeti at quicinc.com> wrote:
> > > > 
> > > > On Tue, Dec 10, 2024 at 09:24:03PM +0000, Marc Zyngier wrote:
> > > > > > +static int a6xx_switch_secure_mode(struct msm_gpu *gpu)
> > > > > > +{
> > > > > > +	int ret;
> > > > > > +
> > > > > > +#ifdef CONFIG_ARM64
> > > > > > +	/*
> > > > > > +	 * We can access SECVID_TRUST_CNTL register when kernel is booted in EL2 mode. So, use it
> > > > > > +	 * to switch the secure mode to avoid the dependency on zap shader.
> > > > > > +	 */
> > > > > > +	if (is_kernel_in_hyp_mode())
> > > > > > +		goto direct_switch;
> > > > > 
> > > > > No, please. To check whether you are *booted* at EL2, you need to
> > > > > check for is_hyp_available(). Whether the kernel runs at EL1 or EL2 is
> > > > > none of the driver's business, really. This is still absolutely
> > > > > disgusting from an abstraction perspective, but I guess we don't have
> > > > > much choice here.
> > > > > 
> > > > 
> > > > Thanks Marc. Any suggestions on how we can make is_hyp_mode_available()
> > > > available for modules? Do you prefer exporting
> > > > kvm_protected_mode_initialized and __boot_cpu_mode symbols directly or
> > > > try something like [1]?
> > > 
> > > Ideally, neither. These were bad ideas nine years ago, and they still
> > > are. The least ugly hack I can come up with is the patch below, and
> > > you'd write something like:
> > > 
> > > 	if (cpus_have_cap(ARM64_HAS_EL2_OWNERSHIP))
> > > 		blah();
> > > 
> > > This is obviously completely untested.
> > > 
> > 
> > I have tested your patch. It works as intended. Thanks Marc.
> 
> Note that you will probably get some push-back from the arm64
> maintainers on this front, because this is a fairly incomplete (and
> fragile) solution.
> 
> It would be much better if the discriminant came from the device tree.
> After all, the hypervisor is fscking-up^W^Wchanging the programming
> model of the GPU, and that should be reflected in the DT. Because for
> all intent and purposes, this is not the same hardware anymore.

FWIW I agree 100%, this should be described in DT.

The cpucap doesn't describe the actual property we care about, and it
cannot in general (e.g. for nested virt). I would strongly prefer to not
have that as it's setting ourselves up for failure.

> The GPU isn't the only device that needs fixing in that way: the
> SMMUv3 needs to be exposed to the OS, and the PCIe ports need to be
> linked to it and the ITS. So at the end of the day, detecting EL2 only
> serves a limited purpose. You need to handle these cases, and might as
> well put the GPU in the same bag.
> 
> Which means that you'd either have a pair of static DTs (one that
> exposes the brokenness of the firmware, and one that doesn't), or you
> go the dtbhack route to compose the DT at boot time.

Liekwise, agreed on all of this.

Mark.