Re: BOOKE KVM calling load_up_fpu from C?

[Scott Wood <scottwood@freescale.com>]

On 02/12/2013 03:01:07 AM, Bhushan Bharat-R65777 wrote:
>=20
>=20
> > -----Original Message-----
> > From: Michael Neuling [mailto:mikey@neuling.org]
> > Sent: Tuesday, February 12, 2013 9:46 AM
> > To: Bhushan Bharat-R65777
> > Cc: Wood Scott-B07421; linuxppc-dev@lists.ozlabs.org
> > Subject: Re: BOOKE KVM calling load_up_fpu from C?
> >
> > Bhushan Bharat-R65777 <R65777@freescale.com> wrote:
> >
> > >
> > >
> > > > -----Original Message-----
> > > > From: Michael Neuling [mailto:mikey@neuling.org]
> > > > Sent: Tuesday, February 12, 2013 9:16 AM
> > > > To: Bhushan Bharat-R65777
> > > > Cc: Wood Scott-B07421; linuxppc-dev@lists.ozlabs.org
> > > > Subject: Re: BOOKE KVM calling load_up_fpu from C?
> > > >
> > > > Look further down...
> > > >
> > > > #ifdef CONFIG_PPC32
> > > > 	mfspr	r5,SPRN_SPRG_THREAD		/* current =20
> task's THREAD (phys) */
> > > > 	lwz	r4,THREAD_FPEXC_MODE(r5)
> > > > 	ori	r9,r9,MSR_FP		/* enable FP for =20
> current */
> > > > 	or	r9,r9,r4
> > > > #else
> > > > 	ld	r4,PACACURRENT(r13)
> > > > 	addi	r5,r4,THREAD		/* Get THREAD */
> > > > 	lwz	r4,THREAD_FPEXC_MODE(r5)
> > > > 	ori	r12,r12,MSR_FP
> > > > 	or	r12,r12,r4
> > > > 	std	r12,_MSR(r1)
> > > > #endif
> > > >
> > > > R12 is loaded with SRR1 in the exception prolog before =20
> load_up_fpu is
> > called.
> > >
> > > Yes it is SRR1 not MSR.
> >
> > Yes, SRR1 =3D=3D the MSR of the user process, not the current MSR.
> >
> > > Also on 32bit it looks like that R9 is assumed to have SRR1.
> >
> > Yep that too.
> >
> > So any idea how it's suppose to work or is it broken?
>=20
> To me this looks wrong. And this seems to works because the =20
> thread->reg->msr is not actually used to write SRR1 (and eventually =20
> the thread MSR) when doing rfi to enter guest. Infact =20
> Guest(shadow_msr) MSR is used as SRR1 and which will have proper MSR =20
> (including FP set).
>=20
> But Yes, Scott is right person to comment, So let us wait for him =20
> comment.

I don't think it's actually a problem on 32-bit, since r9 is modified =20
but never actually used for anything.  On 64-bit, though, there's a =20
store to the caller's stack frame (yuck) which the kvm/booke.h caller =20
is not prepared for.  Indeed, book3s's kvmppc_load_up_fpu creates an =20
interrupt-like stack frame, but does not load r9 or r12.

It would be really nice if assumptions like these were put in a code =20
comment above load_up_fpu...  and if we didn't have so many random =20
differences between 32-bit and 64-bit. :-P

-Scott=