From 9e985cbf2942a1bb8fcef9adc2a17d90fd7ca8ee Mon Sep 17 00:00:00 2001 From: Sean Christopherson Date: Wed, 6 Mar 2024 16:58:33 -0800 Subject: [PATCH] KVM: x86/pmu: Disable support for adaptive PEBS Drop support for virtualizing adaptive PEBS, as KVM's implementation is architecturally broken without an obvious/easy path forward, and because exposing adaptive PEBS can leak host LBRs to the guest, i.e. can leak host kernel addresses to the guest. Bug #1 is that KVM doesn't account for the upper 32 bits of IA32_FIXED_CTR_CTRL when (re)programming fixed counters, e.g fixed_ctrl_field() drops the upper bits, reprogram_fixed_counters() stores local variables as u8s and truncates the upper bits too, etc. Bug #2 is that, because KVM _always_ sets precise_ip to a non-zero value for PEBS events, perf will _always_ generate an adaptive record, even if the guest requested a basic record. Note, KVM will also enable adaptive PEBS in individual *counter*, even if adaptive PEBS isn't exposed to the guest, but this is benign as MSR_PEBS_DATA_CFG is guaranteed to be zero, i.e. the guest will only ever see Basic records. Bug #3 is in perf. intel_pmu_disable_fixed() doesn't clear the upper bits either, i.e. leaves ICL_FIXED_0_ADAPTIVE set, and intel_pmu_enable_fixed() effectively doesn't clear ICL_FIXED_0_ADAPTIVE either. I.e. perf _always_ enables ADAPTIVE counters, regardless of what KVM requests. Bug #4 is that adaptive PEBS *might* effectively bypass event filters set by the host, as "Updated Memory Access Info Group" records information that might be disallowed by userspace via KVM_SET_PMU_EVENT_FILTER. Bug #5 is that KVM doesn't ensure LBR MSRs hold guest values (or at least zeros) when entering a vCPU with adaptive PEBS, which allows the guest to read host LBRs, i.e. host RIPs/addresses, by enabling "LBR Entries" records. Disable adaptive PEBS support as an immediate fix due to the severity of the LBR leak in particular, and because fixing all of the bugs will be non-trivial, e.g. not suitable for backporting to stable kernels. Note! This will break live migration, but trying to make KVM play nice with live migration would be quite complicated, wouldn't be guaranteed to work (i.e. KVM might still kill/confuse the guest), and it's not clear that there are any publicly available VMMs that support adaptive PEBS, let alone live migrate VMs that support adaptive PEBS, e.g. QEMU doesn't support PEBS in any capacity. Link: https://lore.kernel.org/all/20240306230153.786365-1-seanjc@google.com Link: https://lore.kernel.org/all/ZeepGjHCeSfadANM@google.com Fixes: c59a1f106f5c ("KVM: x86/pmu: Add IA32_PEBS_ENABLE MSR emulation for extended PEBS") Cc: stable@vger.kernel.org Cc: Like Xu Cc: Mingwei Zhang Cc: Zhenyu Wang Cc: Zhang Xiong Cc: Lv Zhiyuan Cc: Dapeng Mi Cc: Jim Mattson Acked-by: Like Xu Link: https://lore.kernel.org/r/20240307005833.827147-1-seanjc@google.com Signed-off-by: Sean Christopherson --- arch/x86/kvm/vmx/vmx.c | 24 ++++++++++++++++++++++-- 1 file changed, 22 insertions(+), 2 deletions(-) diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c index c37a89eda90f..e3315bda6237 100644 --- a/arch/x86/kvm/vmx/vmx.c +++ b/arch/x86/kvm/vmx/vmx.c @@ -7882,8 +7882,28 @@ static u64 vmx_get_perf_capabilities(void) if (vmx_pebs_supported()) { perf_cap |= host_perf_cap & PERF_CAP_PEBS_MASK; - if ((perf_cap & PERF_CAP_PEBS_FORMAT) < 4) - perf_cap &= ~PERF_CAP_PEBS_BASELINE; + + /* + * Disallow adaptive PEBS as it is functionally broken, can be + * used by the guest to read *host* LBRs, and can be used to + * bypass userspace event filters. To correctly and safely + * support adaptive PEBS, KVM needs to: + * + * 1. Account for the ADAPTIVE flag when (re)programming fixed + * counters. + * + * 2. Gain support from perf (or take direct control of counter + * programming) to support events without adaptive PEBS + * enabled for the hardware counter. + * + * 3. Ensure LBR MSRs cannot hold host data on VM-Entry with + * adaptive PEBS enabled and MSR_PEBS_DATA_CFG.LBRS=1. + * + * 4. Document which PMU events are effectively exposed to the + * guest via adaptive PEBS, and make adaptive PEBS mutually + * exclusive with KVM_SET_PMU_EVENT_FILTER if necessary. + */ + perf_cap &= ~PERF_CAP_PEBS_BASELINE; } return perf_cap;