Re: [PATCH qemu] spapr: Force 32bit when resetting a core

[Cédric Le Goater]

On 1/10/22 03:52, Alexey Kardashevskiy wrote:
> On 08/01/2022 00:39, Greg Kurz wrote:
> > On Fri, 7 Jan 2022 23:19:03 +1100
> > David Gibson <david@gibson.dropbear.id.au> wrote:
> >
> > > On Fri, Jan 07, 2022 at 12:57:47PM +0100, Greg Kurz wrote:
> > > > On Fri, 7 Jan 2022 18:24:23 +1100
> > > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> > > >
> > > > > "PowerPC Processor binding to IEEE 1275" says in
> > > > > "8.2.1. Initial Register Values" that the initial state is defined as
> > > > > 32bit so do it for both SLOF and VOF.
> > > > >
> > > > > This should not cause behavioral change as SLOF switches to 64bit very
> > > > > early anyway.
> > > >
> > > > Only one CPU goes through SLOF. What about the other ones, including
> > > > hot plugged CPUs ?
> > >
> > > Those will be started by the start-cpu RTAS call which has its own
> > > semantics.
> >
> > Ah indeed, there's code in linux/arch/powerpc/kernel/head_64.S to switch
> > secondaries to 64bit... but then, as noted by Cedric, ppc_cpu_reset(),
> > which is called earlier sets MSR_SF but the changelog of commit 8b9f2118ca40
> > doesn't provide much details on the motivation. Any idea ?
>
> https://patchwork.kernel.org/project/qemu-devel/patch/1458121432-2855-1-git-send-email-lvivier@redhat.com/
>
> this is probably it:
>
> ===
> Reset is properly defined as an exception (0x100). For exceptions, the
> 970MP user manual for example says:
>
> 4.5 Exception Definitions
> When an exception/interrupt is taken, all bits in the MSR are set to
> ‘0’, with the following exceptions:
> • Exceptions always set MSR[SF] to ‘1’.
> ===
>
> but it looks like the above is about emulation bare metal 970 rather than 
> pseries VCPU so that quote does not apply to spapr.

Yes, more info here :

  
https://patchwork.kernel.org/project/qemu-devel/patch/1458121432-2855-1-git-send-email-lvivier@redhat.com/

mac99+970 only boots with a 64bit kernel. 32bit are not supported because
of the use of the rfi instruction which was removed in v2.01. 32bit user
space is supported though.

However I was not able to build a disk with a compatible boot partition
for OpenBIOS. The above support only applies for kernel loaded in memory.
May be Mark knows how to do this ?

Anyhow, I didn't see any regression on PAPR with this patch, TCG or KVM.

Thanks,

C.