Re: [PATCH 4/8] s390/entry: workaround llvm's IAS limitations

[Heiko Carstens]

On Mon, May 16, 2022 at 11:07:43AM +0200, Alexander Gordeev wrote:
> > diff --git a/arch/s390/kernel/entry.S b/arch/s390/kernel/entry.S
> > index a6b45eaa3450..f2f30bfba1e9 100644
> > --- a/arch/s390/kernel/entry.S
> > +++ b/arch/s390/kernel/entry.S
> > @@ -172,9 +172,19 @@ _LPP_OFFSET	= __LC_LPP
> >  	lgr	%r14,\reg
> >  	larl	%r13,\start
> >  	slgr	%r14,%r13
> > -	lghi	%r13,\end - \start
> > -	clgr	%r14,%r13
> > +#ifdef CONFIG_AS_IS_LLVM
> > +	clgfrl	%r14,.Lrange_size\@
> > +#else
> > +	clgfi	%r14,\end - \start
> > +#endif
> >  	jhe	\outside_label
> > +#ifdef CONFIG_CC_IS_CLANG
> > +	.section .rodata, "a"
> > +	.align 4
> > +.Lrange_size\@:
> > +	.long	\end - \start
> 
> Isn't the machine check handler refers to this memory before checking
> unrecoverable storage errors (with CHKSTG macro) as result of this change?

Yes, indeed. However implementing this without another register will
be quite of a challenge. So what I would prefer in any case: just
assume that this minimal set of memory accesses work. Actually I'd
seriously like to go a bit further, and even move the checks for
storage errors back to C for two reasons:

- this would make the machine check handler entry code easier again
- it would also allow to enter the machine check handler with DAT on

After all we rely anyway on the fact that at least the local lowcore +
the page(s) which contain text are still accessible. Assuming that a
couple of page tables also work won't make this much worse, but the
code much easier.

So I'd suggest: leave this code as is, and at some later point move
"rework" the early machine check handler code.

What do you think?