On 6/3/24 10:58 AM, Borislav Petkov wrote:
On Wed, May 29, 2024 at 05:44:32PM -0700, Alexey Makhalov wrote:
While most of the vmware_hypercall callers are executed after alternative
patching applied, there are small amount of hypercalls running before that.
Only for them we have the logic of analyzing vmware_hypercall_mode as a
default alternative code. And there are 2 constraints:
1. vmcall/vmmcall are not supported by old ESXi/Workstation/Fusion. We have
to use in/out instructions. After the end of support of old hypervisors the
alternative can be simplified as follow:
ALTERNATIVE("vmcall", "vmmcall", X86_FEATURE_VMW_VMMCALL);
2. SEV-ES enabled VMs should use _only_ vmcall/vmmcall as in/out
instructions cause faults.
Another approach that we discussed internally was to use
ALTERNATIVE_2("movw %[port], %%dx; "inl (%%dx), %%eax", "vmcall",
X86_FEATURE_VMW_VMCALL, "vmmcall", X86_FEATURE_VMW_VMMCALL) for
vmware_hypercallX family of functions, _and_ to have a separate API
vmware_sev_hypercallX, with the silly dance without an alternative inside,
to be used only by early boot code, before alternative application. But,
it's error prone when things come to boot time related code movements or
rearrangements as it puts additional requirement for SEV-ES
understanding/testing for VMware guests.
Right, so since we're exporting that alternatives_patched thing already,
you might also try to do:
if (unlikely(!alternatives_patched))
return slow_hypercall_X_in_c();
asm_inline volatile(VMWARE_HYPERCALL...
where that slow_hypercall_X_in_c()* set of APIs does the checks in C.
And the VMWARE_HYPERCALL thing is a lot simpler then.
All in all, you'll have a lot less unreadable asm to pay attention to
and those APIs should be all easy and readable.
Thanks for the idea.
I improved the condition to eliminate slow path for modules such as
vmmouse, vmwgfx.
if (unlikely(!alternatives_patched) && !__is_defined(MODULE))
return vmware_hypercall_slow(...);
It also drops the need for exporting vmware_hypercall_mode or
vmware_hypercall_slow symbols.
Will post just Patch0001 for review here before sending v11 out.
