Re: Skylake early panic after GDS mark

[Dave Hansen]

... adding some mailing lists

On 10/11/23 08:48, Mike Pagano wrote:
> Hello, Dave,
> 
> I get a very early kernel panic with commit:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/arch/x86/kernel/cpu/common.c?id=c9f4c45c8ec3f07f4f083f9750032a1ec3eab6b2
> 
> I reverted this and the system boots
> 
> $ dmesg | grep -i microcode
> [    0.000000] microcode: updated early: 0xc2 -> 0xf2, date = 2023-01-02
> [    0.528415] microcode: Microcode Update Driver: v2.2.

You've probably got a bug in your init binary.  The instruction that
puked shows up as:

	vmovd  xmm2,esi

... and it's in userspace.  You don't have a microcode mitigation for
GDS, so you're probably set to use GDS_MITIGATION_FORCE, which means the
kernel is disabling AVX:

>         /* No microcode */
>         if (!(x86_read_arch_cap_msr() & ARCH_CAP_GDS_CTRL)) {
>                 if (gds_mitigation == GDS_MITIGATION_FORCE) {
>                         /*
>                          * This only needs to be done on the boot CPU so do it
>                          * here rather than in update_gds_msr()
>                          */
>                         setup_clear_cpu_cap(X86_FEATURE_AVX);
>                         pr_warn("Microcode update needed! Disabling AVX as mitigation.\n");
>                 } else {
>                         gds_mitigation = GDS_MITIGATION_UCODE_NEEDED;
>                 }
>                 goto out;
>         }

One, you should file a bug on your init process since it's probably not
doing AVX enumeration properly.  It probably needs to have a fallback
path for when the CPU doesn't support AVX.

Second, you can work around this with:

	gather_data_sampling=off

That's not great because you'll still be exposed to the vulnerability.
But you can at least boot.