MIPS: fcsr31 may be dirty after execve when kernel preempt is enabled (was: Re: [PATCH v2] LoongArch: Fix and simplify fcsr initialization on execve)

[Xi Ruoyao <xry111@xry111.site>]

On Tue, 2024-01-02 at 18:25 +0800, Xi Ruoyao wrote:
> On Tue, 2024-01-02 at 18:17 +0800, Xi Ruoyao wrote:
> > The only other architecture setting FCSR in SET_PERSONALITY2 is MIPS.
> > They do this for supporting different FP flavors (NaN encodings etc).
> > which do not exist on LoongArch.  I'm not sure how MIPS evades the issue
> > (or maybe it's just buggy too) but I'll investigate it later.
> 
> Phew.  I just managed to recommission my 3A4000 and I can reproduce the
> issue as well with Linux 5.18.1 (the latest kernel release when I
> decommissioned it) and CONFIG_PREEMPT=y.
> 
> % cat measure.c
> #include <fenv.h>
> int main() { return fetestexcept(FE_INEXACT); }
> 
> % echo $((1./3))
> 0.33333333333333331
> 
> % while ./a.out; do ; done
> (stopped in seconds)
> 
> I'm building the mainline kernel on the 3A4000 now, will see if the
> issue still exists...

Still happening with 6.7.0-rc8.  I'm not sure how to fix it for MIPS. 
Maybe lose_fpu in SET_PERSONALITY2? But to me doing so will be really
nasty.  Anyway I'll leave this for MIPS maintainers.

-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University