Re: [PATCH] KVM: set_msr_mce: Permit guests to ignore single-bit ECC errors

[Sean Christopherson <seanjc@google.com>]

"KVM: x86:" for the shortlog scope.

On Sat, May 21, 2022, Lev Kujawski wrote:
> Certain guest operating systems (e.g., UNIXWARE) clear bit 0 of
> MC1_CTL to ignore single-bit ECC data errors.

Not that it really matters, but is this behavior documented anywhere?  I've searched
a variety of SDMs, APMs, and PPRs, and can't find anything that documents this exact
behavior.  I totally believe that some CPUs behave this way, but it'd be nice to
document exactly which generations of whose CPUs allow clearing bit zero.

> Single-bit ECC data errors are always correctable and thus are safe to ignore
> because they are informational in nature rather than signaling a loss of data
> integrity.
> 
> Prior to this patch, these guests would crash upon writing MC1_CTL,
> with resultant error messages like the following:
> 
> error: kvm run failed Operation not permitted
> EAX=fffffffe EBX=fffffffe ECX=00000404 EDX=ffffffff
> ESI=ffffffff EDI=00000001 EBP=fffdaba4 ESP=fffdab20
> EIP=c01333a5 EFL=00000246 [---Z-P-] CPL=0 II=0 A20=1 SMM=0 HLT=0
> ES =0108 00000000 ffffffff 00c09300 DPL=0 DS   [-WA]
> CS =0100 00000000 ffffffff 00c09b00 DPL=0 CS32 [-RA]
> SS =0108 00000000 ffffffff 00c09300 DPL=0 DS   [-WA]
> DS =0108 00000000 ffffffff 00c09300 DPL=0 DS   [-WA]
> FS =0000 00000000 ffffffff 00c00000
> GS =0000 00000000 ffffffff 00c00000
> LDT=0118 c1026390 00000047 00008200 DPL=0 LDT
> TR =0110 ffff5af0 00000067 00008b00 DPL=0 TSS32-busy
> GDT=     ffff5020 000002cf
> IDT=     ffff52f0 000007ff
> CR0=8001003b CR2=00000000 CR3=0100a000 CR4=00000230
> DR0=00000000 DR1=00000000 DR2=00000000 DR3=00000000
> DR6=ffff0ff0 DR7=00000400
> EFER=0000000000000000
> Code=08 89 01 89 51 04 c3 8b 4c 24 08 8b 01 8b 51 04 8b 4c 24 04 <0f>
> 30 c3 f7 05 a4 6d ff ff 10 00 00 00 74 03 0f 31 c3 33 c0 33 d2 c3 8d
> 74 26 00 0f 31 c3
> 
> Signed-off-by: Lev Kujawski <lkujaw@member.fsf.org>
> ---
>  arch/x86/kvm/x86.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 4790f0d7d40b..128dca4e7bb7 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -3215,10 +3215,13 @@ static int set_msr_mce(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>  			/* only 0 or all 1s can be written to IA32_MCi_CTL
>  			 * some Linux kernels though clear bit 10 in bank 4 to
>  			 * workaround a BIOS/GART TBL issue on AMD K8s, ignore
> -			 * this to avoid an uncatched #GP in the guest
> +			 * this to avoid an uncatched #GP in the guest.
> +			 *
> +			 * UNIXWARE clears bit 0 of MC1_CTL to ignore
> +			 * correctable, single-bit ECC data errors.
>  			 */
>  			if ((offset & 0x3) == 0 &&
> -			    data != 0 && (data | (1 << 10)) != ~(u64)0)
> +			    data != 0 && (data | (1 << 10) | 1) != ~(u64)0)
>  				return -1;

If KVM injects a #GP like it's supposed to[*], will UNIXWARE eat the #GP and continue
on, or will it explode?  If it continues on, I'd prefer to avoid more special casing in
KVM.

If it explodes, I think my preference would be to just drop the MCi_CTL checks
entirely.  AFAICT, P4-based and P5-based Intel CPus, and all? AMD CPUs allow
setting/clearing arbitrary bits.  The checks really aren't buying us anything,
and it seems like Intel retroactively defined the "architectural" behavior of
only 0s/1s.

[*] https://lore.kernel.org/all/20220512222716.4112548-2-seanjc@google.com