Re: [PATCH] KVM: x86: Handle unexpected MMIO accesses using master abort semantics

From: Sean Christopherson
Date: Thu Sep 12 2019 - 21:54:12 EST


On Thu, Sep 12, 2019 at 04:56:03PM -0700, Sean Christopherson wrote:
> Use master abort semantics, i.e. reads return all ones and writes are
> dropped, to handle unexpected MMIO accesses when reading guest memory
> instead of returning X86EMUL_IO_NEEDED, which in turn gets interpreted
> as a guest page fault.
>
> Emulation of certain instructions, notably VMX instructions, involves
> reading or writing guest memory without going through the emulator.
> These emulation flows are not equipped to handle MMIO accesses as no
> sane and properly functioning guest kernel will target MMIO with such
> instructions, and so simply inject a page fault in response to
> X86EMUL_IO_NEEDED.
>
> While not 100% correct, using master abort semantics is at least
> sometimes correct, e.g. non-existent MMIO accesses do actually master
> abort, whereas injecting a page fault is always wrong, i.e. the issue
> lies in the physical address domain, not in the virtual to physical
> translation.
>
> Apply the logic to kvm_write_guest_virt_system() in addition to
> replacing existing #PF logic in kvm_read_guest_virt(), as VMPTRST uses
> the former, i.e. can also leak a host stack address.
>
> Reported-by: Fuqian Huang <huangfq.daxian@xxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> ---
> arch/x86/kvm/x86.c | 40 +++++++++++++++++++++++++++++++---------
> 1 file changed, 31 insertions(+), 9 deletions(-)
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index b4cfd786d0b6..d1d7e9fac17a 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -5234,16 +5234,24 @@ int kvm_read_guest_virt(struct kvm_vcpu *vcpu,
> struct x86_exception *exception)
> {
> u32 access = (kvm_x86_ops->get_cpl(vcpu) == 3) ? PFERR_USER_MASK : 0;
> + int r;
> +
> + r = kvm_read_guest_virt_helper(addr, val, bytes, vcpu, access,
> + exception);
>
> /*
> - * FIXME: this should call handle_emulation_failure if X86EMUL_IO_NEEDED
> - * is returned, but our callers are not ready for that and they blindly
> - * call kvm_inject_page_fault. Ensure that they at least do not leak
> - * uninitialized kernel stack memory into cr2 and error code.
> + * FIXME: this should technically call out to userspace to handle the
> + * MMIO access, but our callers are not ready for that, so emulate
> + * master abort behavior instead, i.e. writes are dropped.

Dagnabbit, fixed this to make it 'reads return all ones' and forgot to
commit.. v2 on its way.

> */
> - memset(exception, 0, sizeof(*exception));
> - return kvm_read_guest_virt_helper(addr, val, bytes, vcpu, access,
> - exception);
> + if (r == X86EMUL_IO_NEEDED) {
> + memset(val, 0xff, bytes);
> + return 0;
> + }
> + if (r == X86EMUL_PROPAGATE_FAULT)
> + return -EFAULT;
> + WARN_ON_ONCE(r);
> + return 0;
> }
> EXPORT_SYMBOL_GPL(kvm_read_guest_virt);
>
> @@ -5317,11 +5325,25 @@ static int emulator_write_std(struct x86_emulate_ctxt *ctxt, gva_t addr, void *v
> int kvm_write_guest_virt_system(struct kvm_vcpu *vcpu, gva_t addr, void *val,
> unsigned int bytes, struct x86_exception *exception)
> {
> + int r;
> +
> /* kvm_write_guest_virt_system can pull in tons of pages. */
> vcpu->arch.l1tf_flush_l1d = true;
>
> - return kvm_write_guest_virt_helper(addr, val, bytes, vcpu,
> - PFERR_WRITE_MASK, exception);
> + r = kvm_write_guest_virt_helper(addr, val, bytes, vcpu,
> + PFERR_WRITE_MASK, exception);
> +
> + /*
> + * FIXME: this should technically call out to userspace to handle the
> + * MMIO access, but our callers are not ready for that, so emulate
> + * master abort behavior instead, i.e. writes are dropped.
> + */
> + if (r == X86EMUL_IO_NEEDED)
> + return 0;
> + if (r == X86EMUL_PROPAGATE_FAULT)
> + return -EFAULT;
> + WARN_ON_ONCE(r);
> + return 0;
> }
> EXPORT_SYMBOL_GPL(kvm_write_guest_virt_system);
>
> --
> 2.22.0
>