Re: AMD DC graphics display code enables -mhard-float, -msse, -msse2 without any visible FPU state protection

[Jann Horn <jannh@google.com>]

On Thu, Apr 2, 2020 at 9:34 AM Christian König <christian.koenig@amd.com> wrote:
> Am 02.04.20 um 04:34 schrieb Jann Horn:
> > [x86 folks in CC so that they can chime in on the precise rules for this stuff]
> > I noticed that several makefiles under drivers/gpu/drm/amd/display/dc/
> > turn on floating-point instructions in the compiler flags
> > (-mhard-float, -msse and -msse2) in order to make the "float" and
> > "double" types usable from C code without requiring helper functions.
> >
> > However, as far as I know, code running in normal kernel context isn't
> > allowed to use floating-point registers without special protection
> > using helpers like kernel_fpu_begin() and kernel_fpu_end() (which also
> > require that the protected code never blocks). If you violate that
> > rule, that can lead to various issues - among other things, I think
> > the kernel will clobber userspace FPU register state, and I think the
> > kernel code can blow up if a context switch happens at the wrong time,
> > since in-kernel task switches don't preserve FPU state.
> >
> > Is there some hidden trick I'm missing that makes it okay to use FPU
> > registers here?
> >
> > I would try testing this, but unfortunately none of the AMD devices I
> > have here have the appropriate graphics hardware...
>
> yes, using the floating point calculations in the display code has been
> a source of numerous problems and confusion in the past.
>
> The calls to kernel_fpu_begin() and kernel_fpu_end() are hidden behind
> the DC_FP_START() and DC_FP_END() macros which are supposed to hide the
> architecture depend handling for x86 and PPC64.

Hmm... but as far as I can tell, you're using those macros from inside
functions that are already compiled with the FPU on:

 - drivers/gpu/drm/amd/display/dc/calcs/dcn_calcs.c uses the macros,
but is already compiled with calcs_ccflags
 - drivers/gpu/drm/amd/display/dc/dcn20/dcn20_resource.c uses the
macros, but is already compiled with "-mhard-float -msse -msse2"
 - drivers/gpu/drm/amd/display/dc/dcn21/dcn21_resource.c uses the
macros, but is already compiled with "-mhard-float -msse -msse2"

AFAIK as soon as you enter any function in any file compiled with FPU
instructions, you may encounter SSE instructions, e.g. via things like
compiler-generated memory-zeroing code - not just when you're actually
using doubles or floats.