KVM: arm: Refuse to enable KVM on systems with SME but not FGT
cover KVM: arm: Refuse to enable KVM on systems with SME but not FGT
The architecture requires that any system which implements SME also has fine grained traps since SME is a v9.2 feature, meaning that v8.7 must be implemented, and FGT is mandatory from v8.6. SME relies on fine grained traps to control access to SMPRI_EL1 and in nVHE mode to TPIDR2_EL0, without traps SMPRI_EL1.Priority and TPIDR2_EL0 can be used as side channels.
This series adds support for detecting FGT and refuses to allow KVM to be used in architecturally invalid configurations which have SME but not FGT, without detection the issue presents as faults due to EL2 attempting to access the FGT registers which isn't obvious to users. Currently fine grained traps are only used in nVHE but but a series "arm64/sme: Fix SMPRI_EL1 traps for KVM guests" sent along with this will add usage for VHE mode too making the issue more pressing.
Mark Brown (2):
arm64/cpufeature: Add feature detection for fine grained traps
KVM: arm: Refuse to enable KVM on systems with FEAT_SME but not
FEAT_FGT
arch/arm64/kernel/cpufeature.c | 11 +++++++++++ arch/arm64/kvm/arm.c | 15 +++++++++++++++ arch/arm64/tools/cpucaps | 1 + 3 files changed, 27 insertions(+)
base-commit: 9abf2313adc1ca1b6180c508c25f22f9395cc780
Signature · 1 line
2.30.2
PATCH 1/2 arm64/cpufeature: Add feature detection for fine grained traps
Virtualisation support for SME depends on FEAT_FGT since it uses fine grained traps to control access to SMPRI_EL1 and (for nVHE) TPIDR2_EL0 but systems have been encountered which implement SME without VHE. In order to work around such systems we need to detect the absence of FEAT_FGT, add detection in the cpufeature code.
--- arch/arm64/kernel/cpufeature.c | 11 +++++++++++ arch/arm64/tools/cpucaps | 1 + 2 files changed, 12 insertions(+)
arch/arm64/kernel/cpufeature.c +11 -0
@@ -2590,6 +2590,17 @@ static const struct arm64_cpu_capabilities arm64_features[] = { .matches = has_cpuid_feature, .min_field_value = 1, }, + { + .desc = "Fine Grained Traps", + .type = ARM64_CPUCAP_SYSTEM_FEATURE, + .capability = ARM64_HAS_FGT, + .sys_reg = SYS_ID_AA64MMFR0_EL1, + .sign = FTR_UNSIGNED, + .field_pos = ID_AA64MMFR0_EL1_FGT_SHIFT, + .field_width = ID_AA64MMFR0_EL1_FGT_WIDTH, + .min_field_value = ID_AA64MMFR0_EL1_FGT_IMP, + .matches = has_cpuid_feature, + }, #ifdef CONFIG_ARM64_SME { .desc = "Scalable Matrix Extension",
arch/arm64/tools/cpucaps +1 -0
@@ -23,6 +23,7 @@ HAS_DCPOP HAS_E0PD HAS_ECV HAS_EPAN +HAS_FGT HAS_GENERIC_AUTH HAS_GENERIC_AUTH_ARCH_QARMA3 HAS_GENERIC_AUTH_ARCH_QARMA5
Signature · 1 line
2.30.2
PATCH 2/2 KVM: arm: Refuse to enable KVM on systems with FEAT_SME but not FEAT_FGT
In reply to Mark BrownThe architecture requires that any system which implements SME also has fine grained traps since SME is a v9.2 feature, meaning that v8.7 must be implemented, and FGT is mandatory from v8.6. Virtualisation support for SME relies on fine grained traps to control access to SMPRI_EL1 and in nVHE mode to TPIDR2_EL0, without traps SMPRI_EL1.Priority and TPIDR2_EL0 can be used as side channels even if SME support is not exposed to the guest.
--- arch/arm64/kvm/arm.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)
arch/arm64/kvm/arm.c +15 -0
@@ -2183,6 +2183,21 @@ int kvm_arch_init(void *opaque)return -ENODEV ; } +/* +* SME without fine grained traps is an architecturally +* invalid configuration since SME is a v9.2 feature and FGT +* is required from v8.6 but virtual platforms have been +* encountered which don't respect this. Without FGT we can't +* trap access to TPIDR2_EL0 in nVHE mode or SMPRI_EL1 in any +* mode, making this conditional in the code would lead to +* side channels on these out of spec systems. +*/ +if (cpus_have_final_cap (ARM64_SME ) && + !cpus_have_final_cap (ARM64_HAS_FGT )) { +kvm_err ("KVM disabled since system has SME without FGT\n" ); +return -ENODEV ; + } +if (kvm_get_mode () ==KVM_MODE_NONE ) {kvm_info ("KVM disabled from command line\n" );return -ENODEV ;
Signature · 1 line
2.30.2
Quoted 7 lines · Mark Brown
The architecture requires that any system which implements SME also has fine grained traps since SME is a v9.2 feature, meaning that v8.7 must be implemented, and FGT is mandatory from v8.6. SME relies on fine grained traps to control access to
SMPRI_EL1and in nVHE mode toTPIDR2_EL0, without traps SMPRI_EL1.Priority andTPIDR2_EL0can be used as side channels.
Quoted 7 lines
This series adds support for detecting FGT and refuses to allow KVM to be used in architecturally invalid configurations which have SME but not FGT, without detection the issue presents as faults due to EL2 attempting to access the FGT registers which isn't obvious to users. Currently fine grained traps are only used in nVHE but but a series "arm64/sme: Fix
SMPRI_EL1traps for KVM guests" sent along with this will add usage for VHE mode too making the issue more pressing.
Series:
r~
Marc Zyngier
maintainerQuoted 8 lines · Mark Brown
The architecture requires that any system which implements SME also has fine grained traps since SME is a v9.2 feature, meaning that v8.7 must be implemented, and FGT is mandatory from v8.6. SME relies on fine grained traps to control access to
SMPRI_EL1and in nVHE mode toTPIDR2_EL0, without traps SMPRI_EL1.Priority andTPIDR2_EL0can be used as side channels.
Quoted 7 lines
This series adds support for detecting FGT and refuses to allow KVM to be used in architecturally invalid configurations which have SME but not FGT, without detection the issue presents as faults due to EL2 attempting to access the FGT registers which isn't obvious to users. Currently fine grained traps are only used in nVHE but but a series "arm64/sme: Fix
SMPRI_EL1traps for KVM guests" sent along with this will add usage for VHE mode too making the issue more pressing.
I think this goes the wrong way around. SME without FGT is invalid, and yet you keep SME around and device to kill virtualisation support.
I'd rather it is SME that gets disabled when the kernel boots at EL2. Furthermore, this is only working around a QEMU issue which can be fixed (as opposed to HW that is forever baked).
To me, it looks like the most reasonable course of action is a mention in the QEMU release notes that virtualization and SME are currently incompatible, and that the user needs to chose one or the other.
We had similar issues in the past where QEMU would ignore certain trap bits (HCR_EL2.TID{1,2,3}), leading to KVM misbehaving. Did we disable KVM? No, we fixed QEMU instead. This case isn't different.
Thanks,
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Catalin Marinas
maintainerQuoted 18 lines · Marc Zyngier
Quoted 8 lines · Mark Brown
The architecture requires that any system which implements SME also has fine grained traps since SME is a v9.2 feature, meaning that v8.7 must be implemented, and FGT is mandatory from v8.6. SME relies on fine grained traps to control access to
SMPRI_EL1and in nVHE mode toTPIDR2_EL0, without traps SMPRI_EL1.Priority andTPIDR2_EL0can be used as side channels.Quoted 7 lines
This series adds support for detecting FGT and refuses to allow KVM to be used in architecturally invalid configurations which have SME but not FGT, without detection the issue presents as faults due to EL2 attempting to access the FGT registers which isn't obvious to users. Currently fine grained traps are only used in nVHE but but a series "arm64/sme: Fix
SMPRI_EL1traps for KVM guests" sent along with this will add usage for VHE mode too making the issue more pressing.
Quoted 4 lines
I think this goes the wrong way around. SME without FGT is invalid, and yet you keep SME around and device to kill virtualisation support.
I'd rather it is SME that gets disabled when the kernel boots at EL2.
I agree, I rather we did it the other way around.
Signature · 1 line
Catalin
Quoted 6 lines · Marc Zyngier
I think this goes the wrong way around. SME without FGT is invalid, and yet you keep SME around and device to kill virtualisation support.
I'd rather it is SME that gets disabled when the kernel boots at EL2. Furthermore, this is only working around a QEMU issue which can be fixed (as opposed to HW that is forever baked).
I do agree that it would be better if we were able to do this the other way around however access to SMPRI_EL1 at EL1 is not controlled by the SME enable control for EL2, it is controlled by either the fine grained trap or the EL3 SME enable trap. This was the primary driver here rather than the issues with writing to the FGT registers.
Quoted 3 lines
To me, it looks like the most reasonable course of action is a mention in the QEMU release notes that virtualization and SME are currently incompatible, and that the user needs to chose one or the other.
Trying to resolve it at the qemu level would certainly be fine, though note that qemu has already released versions with problems so there's a bit of stable door/horse issue and it'd just be a note on the web site and any point releases.
Quoted 3 lines
We had similar issues in the past where QEMU would ignore certain trap bits (HCR_EL2.TID{1,2,3}), leading to KVM misbehaving. Did we disable KVM? No, we fixed QEMU instead. This case isn't different.
OK.