On 6/2/25 17:42, Jan Beulich wrote:
This is possible when:
1. The malicious domain has nested HVM capabilities.
2. The CPU is running on top of VMX and supports shadow VMCS.
To trigger the bug, the domain must first enable VMX operation for
itself, execute VMXON and then finally execute VMPTRLD on a guest
physical address that is backed by a non-writable p2m mapping.
In `nvmx_handle_vmptrld`, after attempting to map the nested VMCS, Xen
will check whether or not this mapping is suitable for writing and if
not immediately unmap the nested VMCS again and abort the setup of
`nvcpu->nv_vvmcx`. However, Xen at this point erroneously continues
emulation of the VMPTRLD. In particular, if VMCS shadowing is available,
Xen will nonetheless attempt to link up the nested VMCS to its own VMCS
in `nvmx_set_vmcs_pointer`. Importantly, Xen here attempts to
dereference the presumably mapped nested VMCS (which now is merely a
NULL pointer) in order to mark it as a shadow VMCS by applying the
`VMCS_RID_TYPE_MASK` to its revision identifier. Following, the page
fault handler will panic Xen.
I've attached an XTF reproducer that triggers the bug. To setup such a
non-writable p2m mapping for the malicious VMCS, I first setup an
appropriate grant table entry. I've tested it on Xen version 4.20.0.
I expect this to not work anymore on current staging or 4.20.1-pre.
See a8325f981ce4 ("x86/P2M: synchronize fast and slow paths of
p2m_get_page_from_gfn()").
On first glance I don't see how that would impact the type of the
established p2m mapping.
Thing is that with said change grant mappings will cause
hvm_map_guest_frame_rw() to simply fail, rather than returning a r/o
mapping for r/o grant entries.
I see, that makes sense. Thanks for the clarification!
Best,
Manuel
