Re: [PATCH RFC v9 08/51] x86/speculation: Do not enable Automatic IBRS if SEV SNP is enabled

[Dave Hansen]

On 7/20/23 12:11, Kim Phillips wrote:
> Hopefully the commit text in this version will help answer all your
> questions?:

To be honest, it didn't really.  I kinda feel like I was having the APM
contents tossed casually in my direction rather than being provided a
fully considered explanation.

Here's what I came up with instead:

Host-side Automatic IBRS has different behavior based on whether SEV-SNP
is enabled.

Without SEV-SNP, Automatic IBRS protects only the kernel.  But when
SEV-SNP is enabled, the Automatic IBRS protection umbrella widens to all
host-side code, including userspace.  This protection comes at a cost:
reduced userspace indirect branch performance.

To avoid this performance loss, nix using Automatic IBRS on SEV-SNP
hosts.  Fall back to retpolines instead.

=====

Is that about right?

I don't think any chit-chat about the guest side is even relevant.

This also absolutely needs a comment.  Perhaps just pull the code up to
the top level of the function and do this:

	/*
	 * Automatic IBRS imposes unacceptable overhead on host
	 * userspace for SEV-SNP systems.  Zap it instead.
	 */
	if (cpu_feature_enabled(X86_FEATURE_SEV_SNP))
		setup_clear_cpu_cap(X86_FEATURE_AUTOIBRS);

BTW, I assume you've grumbled to folks about this.  It's an awful shame
the hardware (or ucode) was built this was.  It's just throwing
Automatic IBRS out the window because it's not architected in a nice way.

Is there any plan to improve this?