[PATCH v6 3/7] tpm: Add the RNG child device

[Simon Glass]

Hi Tom,

On Wed, 13 Jul 2022 at 12:09, Tom Rini <trini at konsulko.com> wrote:
>
> On Wed, Jul 13, 2022 at 09:28:16AM -0600, Simon Glass wrote:
> > Hi Rob,
> >
> > On Tue, 12 Jul 2022 at 08:11, Rob Herring <robh at kernel.org> wrote:
> > >
> > > On Tue, Jul 12, 2022 at 5:04 AM Simon Glass <sjg at chromium.org> wrote:
> > > >
> > > > Hi Ilias,
> > > >
> > > > On Fri, 8 Jul 2022 at 02:24, Ilias Apalodimas
> > > > <ilias.apalodimas at linaro.org> wrote:
> > > > >
> > > > > Hi Simon,
> > > > >
> > > > > [...]
> > > > >
> > > > > > > +
> > > > > > >  UCLASS_DRIVER(tpm) = {
> > > > > > > -       .id             = UCLASS_TPM,
> > > > > > > -       .name           = "tpm",
> > > > > > > -       .flags          = DM_UC_FLAG_SEQ_ALIAS,
> > > > > > > +       .id                     = UCLASS_TPM,
> > > > > > > +       .name                   = "tpm",
> > > > > > > +       .flags                  = DM_UC_FLAG_SEQ_ALIAS,
> > > > > > >  #if CONFIG_IS_ENABLED(OF_REAL)
> > > > > > > -       .post_bind      = dm_scan_fdt_dev,
> > > > > > > +       .post_bind              = dm_scan_fdt_dev,
> > > > > > >  #endif
> > > > > > > +       .post_probe             = tpm_uclass_post_probe,
> > > > > > >         .per_device_auto        = sizeof(struct tpm_chip_priv),
> > > > > > >  };
> > > > > > > --
> > > > > > > 2.25.1
> > > > > > >
> > > > > >
> > > > > > The driver needs a compatible string so it can be in the device tree.
> > > > >
> > > > > Why?  I've tried to hint this on the previous iteration of the patch.
> > > > > The RNG here is not a *device*.  The TPM is the device and you are
> > > > > guaranteed to have an RNG.  The way to get a random number is send a
> > > > > special command to the TPM. So all that we should do here is leverage
> > > > > the fact that the TPM is already in the device tree.
> > > > >
> > > > > And fwiw we should stick to try to stick to what the DT spec defines
> > > > > as much as possible.  I personally don't see this as a special usecase
> > > > > were deviating from the spec is justified.
> > > >
> > > > This is not a deviation from a spec. What spec? Also, I don't want to
> > > > get into another discussion about what a device is. We can disagree on
> > > > that if you like.
> > > >
> > > > One reason is that U-Boot generally requires compatible strings, e.g.
> > > > with of-platdata. But also we can refer to the rand device from
> > > > elsewhere in the tree. I know that Linux does lots of ad-hoc device
> > > > creation and doesn't really have the concept of a uclass consistently
> > > > applied, but this is U-Boot.
> > >
> > > You are letting client/OS details define your binding. Doing so
> > > doesn't result in OS agnostic bindings. Sure, it would be nice if DT
> > > nodes and drivers were always a nice clean 1:1 ratio, but h/w is messy
> > > sometimes and DT is not an abstraction layer. The general guidance on
> > > whether there are child nodes for sub-blocks is do they have their own
> > > resources in DT or are the sub-blocks reusable on multiple devices.
> > > Neither of those are the case here.
> > >
> > > Besides, we already have TPM device bindings defined. Requiring
> > > binding changes when adding a new client/OS feature is not good
> > > practice either.
> >
> > I'm not sure what to do with this, but in general the practice of
> > implied subnodes is not friendly to U-Boot. Yet it seems like a common
> > feature of the bindings at present, for example GPIOs.
> >
> > The device tree is how U-Boot determines which random-number generator
> > to use for a particular function. For example, for VBE, if we need to
> > generate some random numbers we'd like to specify which device creates
> > them. It can't just be 'use the TPM if you find one'. I'm not sure how
> > that works in Linux?
>
> Why can't it be "use the TPM if you find one" since a TPM is a superset
> of an RNG?

You mean look through the available rand devices and choose the one
whose parent is a TPM?

Sure. There are lots of things you can do. But device tree is supposed
to provide the tree of devices and allow such things to be configured
deterministically.

Regards,
Simon