On 22/09/21 17:52, Sean Christopherson wrote:
On Wed, Sep 22, 2021, Paolo Bonzini wrote:
On 22/09/21 16:46, Sean Christopherson wrote:
On Wed, Sep 22, 2021, Paolo Bonzini wrote:
On 13/09/21 16:09, Maxim Levitsky wrote:
KVM: x86: nVMX: re-evaluate emulation_required on nested VM exit
...
Queued, thanks. However, I'm keeping patch 1 for 5.16 only.
I'm pretty sure the above patch is wrong, emulation_required can simply be
cleared on emulated VM-Exit.
Are you sure?
Pretty sure, but not 100% sure :-)
I think you can at least set the host segment fields to a data segment that
requires emulation. For example the DPL of the host DS is hardcoded to zero,
but the RPL comes from the selector field and the DS selector is not
validated.
HOST_DS_SEL is validated:
In the selector field for each of CS, SS, DS, ES, FS, GS and TR, the RPL
(bits 1:0) and the TI flag (bit 2) must be 0.
Ah, I think that's a bug in the manual. In "27.5.2 Loading Host Segment
and Descriptor-Table Registers" the reference to 26.3.1.2 should be
26.2.3 ("Checks on Host Segment and Descriptor-Table Registers"). That
one does cover all segment registers. Hmm, who do we ask now about
fixing Intel manuals?
So yeah, a WARN_ON_ONCE might be in order. But I don't feel super safe
making it false when it is possible to make KVM do something that is at
least sensible.
Paolo
Therefore a subsequent vmentry could fail the access rights tests of 26.3.1.2
Checks on Guest Segment Registers:
Yes, but this path is loading host state on VM-Exit.
DS, ES, FS, GS. The DPL cannot be less than the RPL in the selector field if
(1) the “unrestricted guest” VM-execution control is 0; (2) the register is
usable; and (3) the Type in the access-rights field is in the range 0 – 11
(data segment or non-conforming code segment).
Paolo
