Re: [PATCH v2 1/1] x86/tdx: Add __tdcall() and __tdvmcall() helper functions

[Sean Christopherson]

On Fri, Apr 23, 2021, Dan Williams wrote:
> On Fri, Apr 23, 2021 at 8:15 AM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> >
> > On Thu, Apr 22, 2021, Andi Kleen wrote:
> > > On Thu, Apr 22, 2021 at 06:21:07PM -0700, Dave Hansen wrote:
> > > > On 4/22/21 6:09 PM, Kuppuswamy, Sathyanarayanan wrote:
> > > > > But let me try to explain it here. What I meant by complication is,
> > > > > for in/out instruction, we use alternative_io() to substitute in/out
> > > > > instructions with tdg_in()/tdg_out() assembly calls. So we have to ensure
> > > > > that we don't corrupt registers or stack from the substituted instructions
> > > > >
> > > > > If you check the implementation of tdg_in()/tdg_out(), you will notice
> > > > > that we have added code to preserve the caller registers. So, if we use
> > > > > C wrapper for this use case, there is a chance that it might mess
> > > > > the caller registers or stack.
> > > > >
> > > > >     alternative_io("in" #bwl " %w2, %" #bw "0",            \
> > > > >             "call tdg_in" #bwl, X86_FEATURE_TDX_GUEST,    \
> >
> > Has Intel "officially" switched to "tdg" as the acronym for TDX guest?  As much
> > as I dislike having to juggle "TDX host" vs "TDX guest" concepts, tdx_ vs tdg_
> > isn't any better IMO.  The latter looks an awful lot like a typo, grepping for
> > "tdx" to find relevant code will get fail (sometimes), and confusion seems
> > inevitable as keeping "TDX" out of guest code/comments/documentation will be
> > nigh impossible.
> >
> > If we do decide to go with "tdg" for the guest stuff, then _all_ of the guest
> > stuff, file names included, should use tdg.  Maybe X86_FEATURE_TDX_GUEST could
> > be left as a breadcrumb for translating TDX->TDG.
> >
> > > > >             "=a"(value), "d"(port))
> > > >
> > > > Are you saying that calling C functions from inline assembly might
> > > > corrupt the stack or registers?  Are you suggesting that you simply
> > >
> > > It's possible, but you would need to mark a lot more registers clobbered
> > > (the x86-64 ABI allows to clobber many registers)
> > >
> > > I don't think the stack would be messed up, but there might be problems
> > > with writing the correct unwind information (which tends to be tricky)
> > >
> > > Usually it's better to avoid it.
> >
> > For me, the more important justification is that, if calling from alternative_io,
> > the input parameters will be in the wrong registers.  The OUT wrapper would be
> > especially gross as RAX (the value to write) isn't an input param, i.e. shifting
> > via "ignored" params wouldn't work.
> >
> > But to Dave's point, that justfication needs to be in the changelog.
> 
> It's not clear to me that in()/out() need to be inline asm with an
> alternative vs out-of-line function calls with a static branch?

Code footprint is the main argument, especially since Intel presumably is hoping
most distros will build their generic kernels with CONFIG_TDX_GUEST=y.  IIRC, a
carefully crafted assembly helper with a custom ABI can keep the overhead to
three bytes (or maybe even one byte?) versus raw IN and OUT.  Using a static
branch would incur significantly more overhead since the register prologues would
be different.

Or are you saying make the inb/outb() helpers themselves functions?  That would
be quite brutal, too.  

On a related topic, this changelog also needs to justify changing "Nd" to "d".
Maybe no one cares too much about imm8 port I/O and code footprint, but
effectively killing off "IN AL, imm8" should at least be called out.