Hi, Recently, while trying to understand why the pmu_counters_test selftest sometimes fails when run nested I stumbled upon a very interesting and old bug: It turns out that KVM caches guest segment state, but this cache doesn't have any protection against concurrent use. This usually works because the cache is per vcpu, and should only be accessed by vCPU thread, however there is an exception: If the full preemption is enabled in the host kernel, it is possible that vCPU thread will be preempted, for example during the vmx_vcpu_reset. vmx_vcpu_reset resets the segment cache bitmask and then initializes the segments in the vmcs, however if the vcpus is preempted in the middle of this code, the kvm_arch_vcpu_put is called which reads SS's AR bytes to determine if the vCPU is in the kernel mode, which caches the old value. Later vmx_vcpu_reset will set the SS's AR field to the correct value in vmcs but the cache still contains an invalid value which can later for example leak via KVM_GET_SREGS and such. In particular, kvm selftests will do KVM_GET_SREGS, and then KVM_SET_SREGS, with a broken SS's AR field passed as is, which will lead to vm entry failure. This issue is not a nested issue, and actually I was able to reproduce it on bare metal, but due to timing it happens much more often nested. The only requirement for this to happen is to have full preemption enabled in the kernel which runs the selftest. pmu_counters_test reproduces this issue well, because it creates lots of short lived VMs, but the issue as was noted about is not related to pmu. To fix this issue, I wrapped the places which write the segment fields with preempt_disable/enable. It's not an ideal fix, other options are possible. Please tell me if you prefer these: 1. Getting rid of the segment cache. I am not sure how much it helps these days - this code is very old. 2. Using a read/write lock - IMHO the cleanest solution but might also affect performance. 3. Making the kvm_arch_vcpu_in_kernel not touch the cache and instead do a vmread directly. This is a shorter solution but probably less future proof. Best regards, Maxim Levitsky Maxim Levitsky (2): KVM: nVMX: use vmx_segment_cache_clear KVM: VMX: disable preemption when writing guest segment state arch/x86/kvm/vmx/nested.c | 7 ++++++- arch/x86/kvm/vmx/vmx.c | 22 ++++++++++++++++++---- arch/x86/kvm/vmx/vmx.h | 5 +++++ 3 files changed, 29 insertions(+), 5 deletions(-) -- 2.26.3
