Re: [RFC PATCH v1 11/38] KVM: arm64: CCA: register host tsm platform device

[Aneesh Kumar K . V]

<dan.j.williams@xxxxxxxxx> writes:

> Jason Gunthorpe wrote:
>> On Wed, Jul 30, 2025 at 11:38:27AM +0100, Jonathan Cameron wrote:
>> > On Wed, 30 Jul 2025 14:12:26 +0530
>> > "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxx> wrote:
>> > 
>> > > Jason Gunthorpe <jgg@xxxxxxxx> writes:
>> > > 
>> > > > On Tue, Jul 29, 2025 at 06:10:45PM +0100, Jonathan Cameron wrote:
>> > > >  
>> > > >> > +static struct platform_device cca_host_dev = {  
>> > > >> Hmm. Greg is getting increasingly (and correctly in my view) grumpy with
>> > > >> platform devices being registered with no underlying resources etc as glue
>> > > >> layers.  Maybe some of that will come later.  
>> > > >
>> > > > Is faux_device a better choice? I admit to not knowing entirely what
>> > > > it is for..
>> > 
>> > I'll go with a cautious yes to faux_device. This case of a glue device
>> > with no resources and no reason to be on a particular bus was definitely
>> > the intent but I'm not 100% sure without trying it that we don't run
>> > into any problems.
>> > 
>> > Not that many examples yet, but cpuidle-psci.c looks like a vaguely similar
>> > case to this one.  
>> > 
>> > All it really does is move the location of the device and
>> > smash together the device registration with probe/remove.
>> > That means the device disappears if probe() fails, which is cleaner
>> > in many ways than leaving a pointless stub behind.
>> > 
>> > Maybe it isn't appropriate it if is actually useful to rmmod/modprobe the
>> > driver. 
>> 
>> Yeah, exactly. Can a TSM driver even be modular? If it has to be built
>> in then there is no reason to do this:
>
> For example, CRYPTO_DEV_CCP_DD, the AMD PCI device driver that will call
> tsm_register(), is already modular.
>
>> > > The goal is to have tsm class device to be parented by the platform
>> > > device.
>> 
>> IMHO the only real point of that is to trigger module autoloading.
>
> Right. For TDX, and I expect CCA as well, the arch code that knows that
> PCI/TSM functionality is available and can register a device, may be
> running too early to attach a driver to that device.
>
> I.e. I would like to just use faux_device, but without the ability to do
> EPROBE_DEFER, for example to await the plaform IOMMU driver. It needs to
> move to its own bus so the attach event can be handled at a better time.
>

One of the issues I’ve run into after switching to the faux_device model
is determining how to automatically load the guest and host TSM drivers
based on the availability of the device assignment feature.

The platform device previously provided a clean abstraction for this
behavior, which made autoloading straightforward

>
>> Otherwise the tsm core should accept NULL as the parent pointer during
>> registration, it probably already does..
>
> Yes, NULL @parent "just works" with tsm_register().
>
> However, I expect all tsm_register() callers to be from modular drivers.

-aneesh