Re: [PATCH 1/4] dt-bindings: mailbox: qcom,apcs: Add separate node for clock-controller

[Stephan Gerhold <stephan.gerhold@linaro.org>]

On Wed, May 14, 2025 at 11:08:41AM -0500, Rob Herring wrote:
> On Tue, May 13, 2025 at 02:16:59PM +0100, Stephan Gerhold wrote:
> > On Sun, May 11, 2025 at 05:48:11PM -0500, Bjorn Andersson wrote:
> > > On Tue, May 06, 2025 at 03:10:08PM +0200, Stephan Gerhold wrote:
> > > > APCS "global" is sort of a "miscellaneous" hardware block that combines
> > > > multiple registers inside the application processor subsystem. Two distinct
> > > > use cases are currently stuffed together in a single device tree node:
> > > > 
> > > >  - Mailbox: to communicate with other remoteprocs in the system.
> > > >  - Clock: for controlling the CPU frequency.
> > > > 
> > > > These two use cases have unavoidable circular dependencies: the mailbox is
> > > > needed as early as possible during boot to start controlling shared
> > > > resources like clocks and power domains, while the clock controller needs
> > > > one of these shared clocks as its parent. Currently, there is no way to
> > > > distinguish these two use cases for generic mechanisms like fw_devlink.
> > > > 
> > > > This is currently blocking conversion of the deprecated custom "qcom,ipc"
> > > > properties to the standard "mboxes", see e.g. commit d92e9ea2f0f9
> > > > ("arm64: dts: qcom: msm8939: revert use of APCS mbox for RPM"):
> > > >   1. remoteproc &rpm needs mboxes = <&apcs1_mbox 8>;
> > > >   2. The clock controller inside &apcs1_mbox needs
> > > >      clocks = <&rpmcc RPM_SMD_XO_CLK_SRC>.
> > > >   3. &rpmcc is a child of remoteproc &rpm
> > > > 
> > > > The mailbox itself does not need any clocks and should probe early to
> > > > unblock the rest of the boot process. The "clocks" are only needed for the
> > > > separate clock controller. In Linux, these are already two separate drivers
> > > > that can probe independently.
> > > > 
> > > 
> > > Why does this circular dependency need to be broken in the DeviceTree
> > > representation?
> > > 
> > > As you describe, the mailbox probes and register the mailbox controller
> > > and it registers the clock controller. The mailbox device isn't affected
> > > by the clock controller failing to find rpmcc...
> > > 
> > 
> > That's right, but the problem is that the probe() function of the
> > mailbox driver won't be called at all. The device tree *looks* like the
> > mailbox depends on the clock, so fw_devlink tries to defer probing until
> > the clock is probed (which won't ever happen, because the mailbox is
> > needed to make the clock available).
> > 
> > I'm not sure why fw_devlink doesn't detect this cycle and tries to probe
> > them anyway, but fact is that we need to split this up in order to avoid
> > warnings and have the supplies/consumers set up properly. Those device
> > links are created based on the device tree and not the drivers.
> 
> Does "post-init-providers" providers solve your problem?
> 

I would expect that it does, but it feels like the wrong solution to the
problem to me. The clock is not really a post-init provider: It's not
consumed at all by the mailbox and needed immediately to initialize the
clock controller. The real problem in my opinion is that we're
describing two essentially distinct devices/drivers in a single device
node, and there is no way to distinguish that.

By splitting up the two distinct components into separate device tree
nodes, the relation between the providers/consumers is clearly
described.

Thanks,
Stephan