Re: [REGRESSION][BISECTED] x86: kernel 6.12 crashes during get_xsave_addr_user on Apple Virtualization

[Dave Hansen <dave.hansen@intel.com>]

On 3/13/25 23:22, Akihiro Suda wrote:
>> The warning here seems to be because the kernel thinks protections keys
>> is enabled (X86_FEATURE_OSPKE) but also thinks XSAVE management of
>> protection keys is _dis_abled. That's a valid hardware configuration,
>> but it's a configuration the kernel never expects to see.
> Right.
> The VM seems missing XSAVE feature 0x200: 'Protection Keys User
> registers', while it has ospke in cpuinfo.
> This configuration seems still weird, as both features seem missing on
> the macOS host.
> (See `physical-macos-host.txt`)

Yeah, so the CPU is enumerating support for protection keys but missing
XSAVE support for managing PKRU. We've envisioned the _reverse_
situation before, but not this specific scenario.

One option is to do nothing. Throw up our hands and declare the VMM so
broken that it's silly to go chasing a fix.

Another way to fix it is to throw this hunk in somewhere:

	if (!fpu_kernel_cfg.max_features & XFEATURE_PKRU) {
		setup_clear_cpu_cap(X86_FEATURE_PKU);
		setup_clear_cpu_cap(X86_FEATURE_OSPKE);
	}

but that's PKU-specific.

There are other features that would have similar issues if the VMM did
something similar for them, like:

        [XFEATURE_PASID]                        = X86_FEATURE_ENQCMD,
        [XFEATURE_CET_USER]                     = X86_FEATURE_SHSTK,

which kinda calls for a more generic solution like:

diff --git a/arch/x86/kernel/fpu/xstate.c b/arch/x86/kernel/fpu/xstate.c
index 0bab497c94369..a358d45c781ec 100644
--- a/arch/x86/kernel/fpu/xstate.c
+++ b/arch/x86/kernel/fpu/xstate.c
@@ -800,6 +800,10 @@ void __init fpu__init_system_xstate(unsigned int
legacy_size)
 		/* Careful: X86_FEATURE_FPU is 0! */
 		if ((i != XFEATURE_FP && !cid) || !boot_cpu_has(cid))
 			fpu_kernel_cfg.max_features &= ~BIT_ULL(i);
+
+		if (!(fpu_kernel_cfg.max_features & BIT_ULL(i)) &&
+		    boot_cpu_has(cid))
+			setup_clear_cpu_cap(cid);
 	}

 	if (!cpu_feature_enabled(X86_FEATURE_XFD))

Which leads me back to thinking that doing nothing may be the best
approach here.

Let me noodle on it a bit more.