On Mon, Jun 17, 2024 at 11:27:31AM +0100, Peter Maydell wrote:
> On Thu, 13 Jun 2024 at 11:50, Jean-Philippe Brucker
> <jean-philippe@xxxxxxxxxx> wrote:
> >
> > On Wed, Jun 12, 2024 at 11:59:22AM +0100, Suzuki K Poulose wrote:
> > > On 12/06/2024 11:40, Jean-Philippe Brucker wrote:
> > > > On Wed, Jun 05, 2024 at 10:29:54AM +0100, Steven Price wrote:
> > > > > From: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > > > >
> > > > > Detect that the VM is a realm guest by the presence of the RSI
> > > > > interface.
> > > > >
> > > > > If in a realm then all memory needs to be marked as RIPAS RAM initially,
> > > > > the loader may or may not have done this for us. To be sure iterate over
> > > > > all RAM and mark it as such. Any failure is fatal as that implies the
> > > > > RAM regions passed to Linux are incorrect - which would mean failing
> > > > > later when attempting to access non-existent RAM.
> > > > >
> > > > > Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > > > > Co-developed-by: Steven Price <steven.price@xxxxxxx>
> > > > > Signed-off-by: Steven Price <steven.price@xxxxxxx>
> > > >
> > > > > +static bool rsi_version_matches(void)
> > > > > +{
> > > > > + unsigned long ver_lower, ver_higher;
> > > > > + unsigned long ret = rsi_request_version(RSI_ABI_VERSION,
> > > > > + &ver_lower,
> > > > > + &ver_higher);
> > > >
> > > > There is a regression on QEMU TCG (in emulation mode, not running under KVM):
> > > >
> > > > qemu-system-aarch64 -M virt -cpu max -kernel Image -nographic
> > > >
> > > > This doesn't implement EL3 or EL2, so SMC is UNDEFINED (DDI0487J.a R_HMXQS),
> > > > and we end up with an undef instruction exception. So this patch would
> > > > also break hardware that only implements EL1 (I don't know if it exists).
> > >
> > > Thanks for the report, Could we not check ID_AA64PFR0_EL1.EL3 >= 0 ? I
> > > think we do this for kvm-unit-tests, we need the same here.
> >
> > Good point, it also fixes this case and is simpler. It assumes RMM doesn't
> > hide this field, but I can't think of a reason it would.
> >
> > This command won't work anymore:
> >
> > qemu-system-aarch64 -M virt,secure=on -cpu max -kernel Image -nographic
> >
> > implements EL3 and SMC still treated as undef. QEMU has a special case for
> > starting at EL2 in this case, but I couldn't find what this is for.
>
> That's a bit of an odd config, because it says "emulate EL3 but
> never use it". QEMU's boot loader starts the kernel at EL2 because
> the kernel boot protocol requires that (this is more relevant on
> boards other than virt where EL3 is not command-line disableable).
> I have a feeling we've occasionally found that somebody's had some
> corner case reason to use it, though. (eg
> https://gitlab.com/qemu-project/qemu/-/issues/1899
> is from somebody who says they use this when booting Windows 11 because
> it asserts at boot time that EL3 is present and won't boot otherwise.)
Thanks for the pointer. In this case it looks like Linux was used as
reproducer and not the main use-case, though I wonder if some CIs
regularly boot this particular configuration.
>
> Your underlying problem here seems to be that you don't have
> a way for the firmware to say "hey, SMC works, you can use it" ?
We do: SMCCC recommends to look at the PSCI conduit declared in DT/ACPI.
Given that RMM mandates using the SMC conduit for both PSCI and RSI calls,
we could use this discovery mechanism here. The problem is that we have to
discover it very early at boot, before the DT infrastructure is ready,
so the implementation is a little awkward. I did post one earlier in this
thread but it doesn't yet account for ACPI-only boot, which will need
something similar. Testing ID_AA64PFR0_EL1.EL3 would be much simpler, but
it would break this config.
Thanks,
Jean
