Re: [PATCH v2 4/5] KVM: selftests: Add exception handling support for aarch64

[Ricardo Koller <ricarkol@google.com>]

On Fri, May 07, 2021 at 03:31:56PM +0100, Marc Zyngier wrote:
> On Mon, 03 May 2021 20:12:21 +0100,
> Ricardo Koller <ricarkol@google.com> wrote:
> > 
> > On Mon, May 03, 2021 at 11:32:39AM +0100, Marc Zyngier wrote:
> > > On Sat, 01 May 2021 00:24:06 +0100,
> > > Ricardo Koller <ricarkol@google.com> wrote:
> 
> [...]
> 
> > > > +	.if \vector >= 8
> > > > +	mrs	x1, sp_el0
> > > 
> > > I'm still a bit perplexed by this. SP_EL0 is never changed, since you
> > > always run in handler mode. Therefore, saving/restoring it is only
> > > overhead. If an exception handler wants to introspect it, it is
> > > already available in the relevant system register.
> > > 
> > > Or did you have something else in mind for it?
> > > 
> > 
> > Not really. The reason for saving sp_el0 in there was just for
> > consistency, so that handlers for both el0 and el1 exceptions could
> > get the sp at regs->sp.
> 
> We already have sp_el0 consistency by virtue of having it stored in in
> a sysreg.
> 
> > Restoring sp_el0 might be too much. So, what do you think of this
> > v3: we keep the saving of sp_el0 into regs->sp (to keep things the
> > same between el0 and el1) and delete the restoring of sp_el0?
> 
> To me, the whole purpose of saving some some context is to allow the
> exception handling code to run C code and introspect the interrupted
> state. But saving things that are not affected by the context change
> seems a bit pointless.
> 
> One thing I'd like to see though is to save sp_el1 as it was at the
> point of the exception (because that is meaningful to get the
> exception context -- think of an unaligned EL1 stack for example),
> which means correcting the value that gets saved.

Got it. Replacing:
	mov     x1, sp
with:
	add     x1, sp, #16 * 17

> 
> So I would suggest to *only* save sp_el1, to always save it
> (irrespective of the exception coming from EL0 or EL1), and to save a
> retro-corrected value so that the handler can directly know where the
> previous stack pointer was.

Sounds good, will send a V3 accordingly.

Thanks!
Ricardo

> 
> Thanks,
> 
> 	M.
> 
> -- 
> Without deviation from the norm, progress is not possible.