Re: [PATCH] cpu: Add starts_halted() method

[Greg Kurz <groug@kaod.org>]

On Thu, 9 Jul 2020 07:11:09 +0200
Philippe Mathieu-Daudé <philmd@redhat.com> wrote:

> On 7/8/20 11:39 PM, Eduardo Habkost wrote:
> > On Wed, Jul 08, 2020 at 06:45:57PM +0200, Philippe Mathieu-Daudé wrote:
> >> On 7/8/20 5:25 PM, Eduardo Habkost wrote:
> >>> On Wed, Jul 08, 2020 at 02:14:03PM +0100, Peter Maydell wrote:
> >>>> On Wed, 8 Jul 2020 at 12:12, David Gibson <david@gibson.dropbear.id.au> wrote:
> >>>>> On Wed, Jul 08, 2020 at 10:38:29AM +0200, Philippe Mathieu-Daudé wrote:
> >>>>>> Class boolean field certainly sounds better, but I am not sure this
> >>>>>> is a property of the machine. Rather the arch? So move the field
> >>>>>> to CPUClass? Maybe not, let's discuss :)
> >>>>>
> >>>>> It is absolutely a property of the machine.  e.g. I don't think we
> >>>>> want this for powernv.  pseries is a bit of a special case since it is
> >>>>> explicitly a paravirt platform.  But even for emulated hardware, the
> >>>>> board can absolutely strap things so that cpus do or don't start
> >>>>> immediately.
> >>>>
> >>>> It's a property of the individual CPU, I think. One common setup
> >>>> for Arm systems is that the primary CPU starts powered up but
> >>>> the secondaries all start powered down.
> >>>
> >>> Both statements can be true.  It can be a property of the
> >>> individual CPU (although I'm not convinced it has to), but it
> >>> still needs to be controlled by the machine.
> >>
> >> From what said Peter, I understand this is a property of the
> >> chipset. Chipsets are modelled unevenly.
> >>
> >> IIUC QEMU started with single-core CPUs.
> >> CPUState had same meaning for 'core' or 'cpu', 1-1 mapping.
> >>
> >> Then multicore CPUs could be easily modelled using multiple
> >> single-core CPUs, usually created in the machine code.
> >>
> >> Then we moved to SoC models, creating the cores in the SoC.
> >> Some SoCs have array of cores, eventually heterogeneous
> >> (see the ZynqMP). We have containers of CPUState.
> >>
> >> On an ARM-based SoC, you might have the first core started
> >> (as said Peter) or not.
> >>
> >> BCM2836 / BCM2837 and ZynqMP start will all ARM cores off.
> >> On the BCM chipsets, a DSP core will boot the ARM cores.
> >> On the ZynqMP, a MicroBlaze core boots them.
> >> As QEMU doesn't models heterogeneous architectures, we start
> >> modelling after the unmodelled cores booted us, so either one
> >> or all cores on.
> >>
> >> In this case, we narrowed down the 'start-powered-off' field
> >> to the SoC, which happens to be how ARM SoCs are modelled.
> > 
> > I was not aware of the start-powered-off property.  If we make it
> > generic, we can just let spapr use it.
> > 
> >>
> >>
> >> Chipsets providing a JTAG interface can have a SRST signal,
> >> the "system reset". When a JTAG probe is attached, it can
> >> keeps the whole chipset in a reset state. This is equivalent
> >> to QEMU '-S' mode (single step mode).
> >>
> >>
> >> I don't know about pseries hardware, but if this is 'explicit
> >> to paravirt platform', then I expect this to be the same with
> >> other accelerators/architectures.
> >>
> >> If paravirtualized -> cores start off by default. Let the
> >> hypervisor start them. So still a property of the CPUState
> >> depending on the accelerator used?
> > 
> > I don't understand this part.  Why would this depend on the
> > accelerator?
> 
> Because starting a virtualized machine with all cores powered-off
> with TCG accelerator should at least emit a warning? Or change
> the behavior and start them powered-on? This is machine-specific
> although.
> 

FYI, PAPR requires all vCPUs to be "stopped" by default. It is up to the
guest to start them explicitly through an RTAS call. The hypervisor is
only responsible to start a single vCPU (see spapr_cpu_set_entry_state()
called from spapr_machine_reset()) to be able to boot the guest.

So I'm not sure to see how that would depend on the accelerator...

>