Re: [RFC PATCH v3 08/31] KVM: selftests: Require GCC to realign stacks on function entry

From: Erdem Aktas
Date: Mon Jan 23 2023 - 19:04:40 EST


On Mon, Jan 23, 2023 at 10:53 AM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Mon, Jan 23, 2023, Maciej S. Szmigiero wrote:
> > On 23.01.2023 19:30, Erdem Aktas wrote:
> > > On Fri, Jan 20, 2023 at 4:28 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > > >
> > > > On Sat, Jan 21, 2023, Ackerley Tng wrote:
> > > > > Some SSE instructions assume a 16-byte aligned stack, and GCC compiles
> > > > > assuming the stack is aligned:
> > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=40838. This combination
> > > > > results in a #GP in guests.
> > > > >
> > > > > Adding this compiler flag will generate an alternate prologue and
> > > > > epilogue to realign the runtime stack, which makes selftest code
> > > > > slower and bigger, but this is okay since we do not need selftest code
> > > > > to be extremely performant.
> > > >
> > > > Huh, I had completely forgotten that this is why SSE is problematic. I ran into
> > > > this with the base UPM selftests and just disabled SSE. /facepalm.
> > > >
> > > > We should figure out exactly what is causing a misaligned stack. As you've noted,
> > > > the x86-64 ABI requires a 16-byte aligned RSP. Unless I'm misreading vm_arch_vcpu_add(),
> > > > the starting stack should be page aligned, which means something is causing the
> > > > stack to become unaligned at runtime. I'd rather hunt down that something than
> > > > paper over it by having the compiler force realignment.
> > >
> > > Is not it due to the 32bit execution part of the guest code at boot
> > > time. Any push/pop of 32bit registers might make it a 16-byte
> > > unaligned stack.
> >
> > 32-bit stack needs to be 16-byte aligned, too (at function call boundaries) -
> > see [1] chapter 2.2.2 "The Stack Frame"
>
> And this showing up in the non-TDX selftests rules that out as the sole problem;
> the selftests stuff 64-bit mode, i.e. don't have 32-bit boot code.

Thanks Maciej and Sean for the clarification. I was suspecting the
hand-coded assembly part that we have for TDX tests but it being
happening in the non-TDX selftests disproves it.