Re: [PATCH v4 05/15] x86/sev: Use kernel provided SVSM Calling Areas

From: Tom Lendacky
Date: Wed May 08 2024 - 16:09:22 EST


On 5/8/24 14:58, Borislav Petkov wrote:
On Wed, May 08, 2024 at 02:13:17PM -0500, Tom Lendacky wrote:
ok, maybe __perform_svsm_msr_protocol or such.

We'll bikeshed it in the coming weeks.

:)


There's quite a bit of interaction so I'll make sure to prefix everything.

Ack.

The paravirt versions of local_irq_save and local_irq_restore can't be used
as early as this routine is called.

tglx says you should do native_local_irq_save()/.._restore() helpers
just like the arch_local_irq_save()/..._restore() ones but use only
native_ functions without the paravirt gunk.

In a prepatch pls.

Will do.


+ struct svsm_call call = {};
+ int ret;
+ u64 pa;
+
+ /*
+ * Record the SVSM Calling Area address (CAA) if the guest is not
+ * running at VMPL0. The CA will be used to communicate with the
+ * SVSM to perform the SVSM services.
+ */
+ setup_svsm_ca(cc_info);
+
+ /* Nothing to do if not running under an SVSM. */
+ if (!vmpl)
+ return;

You set up stuff above and now you bail out?

setup_svsm_ca() is what sets the vmpl variable. So nothing will have been
setup if the VMPL is zero, in which case we don't continue on.

You still assign

/*
* The CA is identity mapped when this routine is called, both by the
* decompressor code and the early kernel code.
*/
boot_svsm_caa = (struct svsm_ca *)caa;
boot_svsm_caa_pa = caa;

regardless of vmpl.

If we're not running at VMPL0 (based on the RMPADJUST check) and if the SVSM doesn't advertise a non-zero VMPL value, we will self-terminate. So those values are only set if we are not running at VMPL0 and the SVSM has provided a non-zero value to us.

I'm going to turn the function into a bool function so that the call becomes:

if (!svsm_setup_caa(cc_info))
return;


I think you should assign those only when vmpl != 0.

I do. I think you're missing the RMPADJUST check that causes the function to return early if we're running at VMPL0.


Otherwise the code is confusing.

Judging by setup_svsm_ca() you don't really need that vmpl var but you
can check

if (!boot_svsm_caa)
return;

to determine whether a SVSM was detected.

Yes, but the vmpl var will be used for attestation requests, sysfs, etc.

I guess that comes later in the patchset...

Huh, setup_svsm_ca() already assigned those...

setup_svsm_ca() assigned the ones from the secrets page. The kernel now
switches to using its own CA.

Comment pls.

There's a block comment above it all, but maybe it isn't clear enough. I'll rework it.

Thanks,
Tom


Thx.