Re: [PATCH kvmtool 3/6] arm64: Initialise the PMU last

[Alexandru Elisei <alexandru.elisei@arm.com>]

Hi Will,

On Sun, May 17, 2026 at 09:33:42AM +0100, Will Deacon wrote:
> On Mon, Mar 23, 2026 at 03:02:18PM +0000, Alexandru Elisei wrote:
> > PMU initialisation is done in:
> > 
> > setup_fdt()
> >   vcpu->generate_fdt_nodes()
> >     pmu__generate_fdt_nodes()
> > 
> > The PMU ioctl KVM_ARM_VCPU_PMU_V3_INIT requires that the GIC has been
> > initialised, which is done in gic__init_gic().
> > 
> > gic__init_gic() and setup_fdt() are part of the same initialisation list.
> > The relative order on the list depends on the order of compilation: gic.c
> > is compiled after fdt.c and this places gic__init_gic() first on the
> > initialisation list.
> > 
> > If the compilation order changes and gic.c is compiled *before* fdt.c,
> > gic__init_gic() will be executed *after* setup_fdt() and PMU initialisation
> > will fail with an error message:
> > 
> > PMU KVM_SET_DEVICE_ATTR: No such device
> > 
> > This is fragile and hard to debug. Improve the situation by creating a new
> > init list, last_init(), execute it after late_init(), and move PMU
> > initialisation on that list.
> 
> Hmm, not a big fan of that. There's always _somebody_ who wants to be
> later/earlier than everybody else (see the CCA probing converstion!).

Sure.

> 
> If the PMU genuinely depends on the GIC, then perhaps we should just
> call the PMU init function from the GIC init function, along with a
> comment to explain the dependency? This is all arm64 code so I think
> that's fine.

Sounds good to me.

Thanks,
Alex