Until now, KVM used to assume that CS.RPL could always be used as the CPL value when KVM_SET_SREGS is called. Unfortunately this is not the case. If userspace decides to call KVM_GET_SREGS/KVM_SET_SREGS exactly after CR0.PE has been set to 1, but before the long jump that reloads CS, the CPL will be reset to bits 0-1 of CS (aka CS.RPL). This can work or not, depending on the placement of the code that transitions to protected mode. If CS.RPL != 0 the emulator will see CS.RPL != CS.DPL (the DPL will always be zero) and fail to fetch the next instruction of the transition code. The same bug exists with SVM, where you don't have the emulator but the guest will triple fault. Strangely, it doesn't occur with Intel's unrestricted guest mode. To trigger this using QEMU, it is enough to send "info cpus" continuously while running iPXE (which places its code for real->protected mode in the EBDA). iPXE does a lot of transitions, and the guest will crash very quickly. Avi or Gleb, this is a bit tricky. Can you review it please? Paolo Bonzini (5): KVM: x86: support KVM_ENABLE_CAP on vm file descriptors KVM: x86: get CPL in KVM_GET_SREGS, prepare to be able to set it KVM: vmx: always compute the CPL on KVM_SET_SREGS KVM: vmx: force CPL=0 on real->protected mode transition KVM: x86: add capability to get/set CPL arch/x86/include/asm/kvm_host.h | 4 ++- arch/x86/include/uapi/asm/kvm.h | 5 +++- arch/x86/kvm/svm.c | 13 +++++++-- arch/x86/kvm/vmx.c | 62 ++++++++++++++++++++++++++--------------- arch/x86/kvm/x86.c | 53 +++++++++++++++++++++++++++-------- include/uapi/linux/kvm.h | 1 + 6 files changed, 98 insertions(+), 40 deletions(-) -- 1.8.3.1 -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
