A flaw was found in the KVM's AMD code for supporting SVM nested
virtualization. The flaw occurs when processing the VMCB (virtual machine
control block) provided by the L1 guest to spawn/handle a nested guest
(L2). Due to improper validation of the "virt_ext" field, this issue could
allow a malicious L1 to disable both VMLOAD/VMSAVE intercepts and VLS
(Virtual VMLOAD/VMSAVE) for the L2 guest. As a result, the L2 guest would
be allowed to read/write physical pages of the host, resulting in a crash
of the entire system, leak of sensitive data or potential guest-to-host
escape.

Maxim Levitsky and Paolo Bonzini discovered that the KVM hypervisor
implementation for AMD processors in the Linux kernel allowed a guest VM to
disable restrictions on VMLOAD/VMSAVE in a nested guest. An attacker in a
guest VM could use this to read or write portions of the host's physical
memory.

Disable nested virtualisation when loading the KVM AMD module:

 modprobe kvm_amd nested=0

cascardo