Re: [Qemu-devel] [PATCH 2/7] target/openrisc: add shutdown logic

[Jason A. Donenfeld]

Hi Stafford,

On Thu, Apr 28, 2022 at 06:48:27AM +0900, Stafford Horne wrote:
> On Wed, Apr 27, 2022 at 07:47:33PM +0100, Peter Maydell wrote:
> > On Wed, 27 Apr 2022 at 18:46, Jason A. Donenfeld <Jason@zx2c4.com> wrote:
> > >
> > > Hey Stafford,
> > >
> > > On Mon, Apr 17, 2017 at 08:23:51AM +0900, Stafford Horne wrote:
> > > > In openrisc simulators we use hooks like 'l.nop 1' to cause the
> > > > simulator to exit.  Implement that for qemu too.
> > > >
> > > > Reported-by: Waldemar Brodkorb <wbx@openadk.org>
> > > > Signed-off-by: Stafford Horne <shorne@gmail.com>
> > >
> > > I'm curious as to why this never got merged. I noticed I'm entirely able
> > > to shutdown or to reboot (which is mostly what I care about) Linux from
> > > OpenRISC. It just hangs.
> > 
> > This kind of thing needs to be either:
> >  (1) we're modelling real hardware and that real hardware has a
> > device or other mechanism guest code can prod to cause a power-off
> > or reboot. Then we model that device, and guest code triggers a
> > shutdown or reboot exactly as it would on the real hardware.
> >  (2) there is an architecturally defined ABI for simulators, debug
> > stubs, etc, that includes various operations typically including
> > an "exit the simulator" function. (Arm semihosting is an example
> > of this.) In that case we can implement that functionality,
> > guarded by and controlled by the appropriate command line options.
> > (This is generally not as nice as option 1, because the guest code
> > has to be compiled to have support for semihosting and also because
> > turning it on is usually also giving implicit permission for the
> > guest code to read and write arbitrary host files, etc.)
> > 
> > Either way, undocumented random hacks aren't a good idea, which
> > is why this wasn't merged.
> 
> Yes, this is what was brought up before.  At that time semihosting was 
> mentioned
> and I tried to understand what it was but didn't really understand it as a 
> general
> concept.  Is this something arm specific?
> 
> Since the qemu or1k-sim defines our "simulator", I suspect I could add a
> definition of our simulator ABI to the OpenRISC architecture specification.  
> The
> simulation uses of l.nop N as ABI hooks is a de-facto standard for OpenRISC.
> From the way you describe this now I take it if we document this as a
> architecture simulation ABI the patch would be accepted.

If that's what it takes, then that'd make sense.

By the way, would this also help the reboot case? That's
`reboot(RB_AUTOBOOT);`, which does:

machine_restart() ->
  do_kernel_restart() ->
    atomic_notifier_chain_register(&restart_handler_list, nb) ->
      ???

As far as I can tell, nothing is wired into the reboot case for
OpenRISC. Is this something that could be fixed in the kernel without
having to patch QEMU? If so, then I could effectively get shutdown for
my CI with the -no-reboot option, which is what I'm already doing for a
few platforms.

Jason