On 2020-08-25 11:54, Catalin Marinas wrote:
On Tue, Aug 25, 2020 at 09:53:16AM +0100, Marc Zyngier wrote:
On 2020-08-24 19:27, Catalin Marinas wrote:
> diff --git a/arch/arm64/include/asm/kvm_arm.h
> b/arch/arm64/include/asm/kvm_arm.h
> index 8a1cbfd544d6..6c3b2fc922bb 100644
> --- a/arch/arm64/include/asm/kvm_arm.h
> +++ b/arch/arm64/include/asm/kvm_arm.h
> @@ -78,7 +78,7 @@
> HCR_AMO | HCR_SWIO | HCR_TIDCP | HCR_RW | HCR_TLOR | \
> HCR_FMO | HCR_IMO)
> #define HCR_VIRT_EXCP_MASK (HCR_VSE | HCR_VI | HCR_VF)
> -#define HCR_HOST_NVHE_FLAGS (HCR_RW | HCR_API | HCR_APK)
> +#define HCR_HOST_NVHE_FLAGS (HCR_RW | HCR_API | HCR_APK | HCR_ATA)
> #define HCR_HOST_VHE_FLAGS (HCR_RW | HCR_TGE | HCR_E2H)
Why is HCR_ATA only set for nVHE? HCR_EL2.ATA seems to apply to both,
doesn't it?
We need HCR_EL2.ATA to be set when !VHE so that the host kernel can use
MTE. That said, I think we need to turn it off when running a guest.
Even if we hide the ID register, the guest may still attempt to enable
tags on some memory that doesn't support it, leading to unpredictable
behaviour (well, only if we expose device memory to guests directly;
Steve's patches will deal with this but for now we just disable MTE in
guests).
With VHE, HCR_EL2.ATA only affects the guests, so it can stay off. The
host's use of tags is controlled by SCTLR_EL1/EL2.ATA (i.e. HCR_EL2.ATA
has no effect if E2H and TGE are both 1; qemu has a bug here which I
discovered yesterday).
Ah, I missed that too.
> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
> index 077293b5115f..59b91f58efec 100644
> --- a/arch/arm64/kvm/sys_regs.c
> +++ b/arch/arm64/kvm/sys_regs.c
> @@ -1131,6 +1131,8 @@ static u64 read_id_reg(const struct kvm_vcpu
> *vcpu,
> if (!vcpu_has_sve(vcpu))
> val &= ~(0xfUL << ID_AA64PFR0_SVE_SHIFT);
> val &= ~(0xfUL << ID_AA64PFR0_AMU_SHIFT);
> + } else if (id == SYS_ID_AA64PFR1_EL1) {
> + val &= ~(0xfUL << ID_AA64PFR1_MTE_SHIFT);
Hiding the capability is fine, but where is the handling of trapping
instructions done? They should result in an UNDEF being injected.
They are a few new MTE-specific MSR/MRS which are trapped at EL2 but
since KVM doesn't understand them yet, shouldn't it already inject
undef back at EL1? That would be safer regardless of MTE support.
An UNDEF will be injected, but not without spitting a nastygram in
the kernel log (look at emulate_sys_reg()).
The best course of action is to have an entry in the sysreg table
that would explicitly do the handling.
Thanks,
M.
--
Jazz is not dead. It just smells funny...
