[RFC PATCH 4/4] firmware: arm_ffa: check pkvm initailised when initailise ffa driver

[Yeoreum Yun]

Hi Will,

> [+Seb for the pKVM FFA bits]
>
> On Mon, Apr 20, 2026 at 10:25:29AM +0100, Yeoreum Yun wrote:
> > > On Sun, Apr 19, 2026 at 12:12:44PM +0100, Yeoreum Yun wrote:
> > > > > On Sat, 18 Apr 2026 11:34:30 +0100,
> > > > > Yeoreum Yun <yeoreum.yun at arm.com> wrote:
> > > > > >
> > > > > > > > @@ -2035,6 +2037,16 @@ static int __init ffa_init(void)
> > > > > > > >  	u32 buf_sz;
> > > > > > > >  	size_t rxtx_bufsz = SZ_4K;
> > > > > > > >
> > > > > > > > +	/*
> > > > > > > > +	 * When pKVM is enabled, the FF-A driver must be initialized
> > > > > > > > +	 * after pKVM initialization. Otherwise, pKVM cannot negotiate
> > > > > > > > +	 * the FF-A version or obtain RX/TX buffer information,
> > > > > > > > +	 * which leads to failures in FF-A calls.
> > > > > > > > +	 */
> > > > > > > > +	if (IS_ENABLED(CONFIG_KVM) && is_protected_kvm_enabled() &&
> > > > > > > > +	    !is_kvm_arm_initialised())
> > > > > > > > +		return -EPROBE_DEFER;
> > > > > > > > +
> > > > > > >
> > > > > > > That's still fundamentally wrong: pkvm is not ready until
> > > > > > > finalize_pkvm() has finished, and that's not indicated by
> > > > > > > is_kvm_arm_initialised().
> > > > > >
> > > > > > Thanks. I miss the TSC bit set in here.
> > > > >
> > > > > That's the least of the problems. None of the infrastructure is in
> > > > > place at this stage...
> > > > >
> > > > > > IMHO, I'd like to make an new state check function --
> > > > > > is_pkvm_arm_initialised() so that ff-a driver to know whether
> > > > > > pkvm is initialised.
> > > > >
> > > > > Doesn't sound great, TBH.
> > > > >
> > > > > > or any other suggestion?
> > > > >
> > > > > Instead of adding more esoteric predicates, I'd rather you build on an
> > > > > existing infrastructure. You have a dependency on KVM, use something
> > > > > that is designed to enforce dependencies. Device links spring to mind
> > > > > as something designed for that.
> > > > >
> > > > > Can you look into enabling this for KVM? If that's possible, then it
> > > > > should be easy enough to delay the actual KVM registration after pKVM
> > > > > is finalised.
> > > >
> > > > or what about some event notifier? Just like:
> > >
> > > This seems a bit over-engineered to me. Why don't you just split the
> > > FF-A initialisation into two steps: an early part which does the version
> > > negotiation and then a later part which can fit in with whatever
> > > dependencies you have on the TPM?
> >
> > Sorry, I may have misunderstood your suggestion and
> > I might be in missing your point.
> >
> > But, The issue here is that FFA_VERSION, FFA_RXTX_MAP, and
> > FFA_PARTITION_INFO_GET, which are invoked from ffa_init()
> > as part of early initialisation, must be trapped by pKVM.
> >
> > In other words, even the early part of the initialization,
> > including version negotiation, needs to happen after pKVM
> > is initialized.
> >
> > Because of this dependency, simply splitting the FF-A
> > initialization into two phases within the driver does not
> > seem sufficient, as it still requires knowing when pKVM
> > has been initialized.
> >
> > Am I missing something?
>
> Ah sorry, I mixed up the ordering of 'module_init' vs 'rootfs_initcall'
> and thought you wanted to probe the version earlier. But then I'm still
> confused because, prior to 0e0546eabcd6 ("firmware: arm_ffa: Change
> initcall level of ffa_init() to rootfs_initcall"), ffa_init() was a
> 'device_initcall' which is still called earlier than finalize_pkvm().

Right, and this is what I missed when writing patch
0e0546eabcd6 ("firmware: arm_ffa: Change initcall level of ffa_init() to rootfs_initcall").
and it still exists even if it's device call.

However, rather than changing ffa_init to rootfs_initcall, moving ima_init
to late_initcall_sync is a better approach, as it also addresses similar
issues for TPM devices that do not use FF-A. For this reason,
the FF-A-related changes were reverted.

As a result, patch 4/4 addresses an issue that existed independently of
0e0546eabcd6, as you pointed out.

--
Sincerely,
Yeoreum Yun