On Fri, Aug 19, 2022 at 12:37 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Fri, Aug 19, 2022, Vishal Annapurve wrote:
> > On Wed, Aug 10, 2022 at 8:20 AM Peter Gonda <pgonda@xxxxxxxxxx> wrote:
> > > void ucall(uint64_t cmd, int nargs, ...)
> > > {
> > > - struct ucall uc = {};
> > > + struct ucall *uc;
> > > + struct ucall tmp = {};
> >
> > This steps seems to result in generating instructions that need SSE
> > support on x86:
> > struct ucall tmp = {};
> > movaps %xmm0,0x20(%rsp)
> > movaps %xmm0,0x30(%rsp)
> > movaps %xmm0,0x40(%rsp)
> > movaps %xmm0,0x50(%rsp)
> >
> > This initialization will need proper compilation flags to generate
> > instructions according to VM configuration.
>
> Can you be more specific as to why generating SSE instructiions is problematic?
> The compiler emitting fancy instructions for struct initialization is not out of
> the ordinary.
When executing these instructions, I was hitting #GP in the guest VM.
CR4/CPUID settings seem to allow a VM to execute instructions that
need SSE support enabled (which I was doubting earlier). After some
more digging, it looks like the compiler is not able to ensure that
operands for such instructions are 16 byte aligned as required by the
hardware. There is a bug discussing this same issue:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=40838.
Adding -mstackrealign seems to help avoid running into #GP and things
seem to work normally. Though I am not sure if SSE instruction
execution is reliable from guest VM within selftests and so was
suggesting usage of "-mno-sse" or equivalent flags.
