Re: [PATCH] interconnect: Disallow interconnect core to be built as a module

From: Bjorn Andersson
Date: Mon May 18 2020 - 14:41:38 EST


On Fri 15 May 00:11 PDT 2020, Greg Kroah-Hartman wrote:

> On Fri, May 15, 2020 at 07:48:47AM +0300, Georgi Djakov wrote:
> > On 9/12/19 19:33, Bjorn Andersson wrote:
> > > On Thu, Aug 29, 2019 at 1:07 AM Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> > >>
> > >> Building individual drivers as modules is fine but allowing a core
> > >> framework to be built as a module makes it really complex and should be
> > >> avoided.
> > >>
> > >> Whatever uses the interconnect core APIs must also be built as a module
> > >> if interconnect core is built as module, else we will see compilation
> > >> failures.
> > >>
> > >> If another core framework (like cpufreq, clk, etc), that can't be built
> > >> as module, needs to use interconnect APIs then we will start seeing
> > >> compilation failures with allmodconfig configurations as the symbols
> > >> (like of_icc_get()) used in other frameworks will not be available in
> > >> the built-in image.
> > >>
> > >> Disallow the interconnect core to be built as a module to avoid all
> > >> these issues.
> >
> > Hi Greg,
> >
> > We had a discussion [1] a few months back about frameworks being built as
> > modules. IIUC, you initially expressed some doubts about this patch, so i
> > wanted to check with you again on this.
> >
> > While i think that the possibility for a framework core to be a module is a
> > nice feature, and we should try to be as modular as possible, it seems that
> > handling dependencies between the different core frameworks becomes difficult
> > when one of them is tristate.
> >
> > This of course affects the drivers which use it (every client should express
> > the dependency in Kconfig as a "depends on framework || !framework"), in order
> > to avoid build failures in the case when framework=m and client=y. However, this
> > is not a big issue.
> >
> > But it gets more complex when another framework2 becomes a client of the modular
> > framework and especially when framework2 is "select"-ed in Kconfig by it's
> > users. When selects are used in Kconfig, it forces the option, without ever
> > visiting the dependencies. I am not sure what we should do in this case, maybe
> > we can continue and sprinkle more "depends on framework || !framework" also for
> > every single user which selects framework2.. But i believe that this is very
> > inconvenient.
> >
> > Well, the above is not impossible, but other frameworks (regulator, clk, reset,
> > pinctrl, etc.) are solving this problem by just being bool, instead of tristate.
> > This makes life much easier for everyone. So i am wondering if it wouldn't be
> > more appropriate to use the same approach here too?
>
> Ok, if it makes things easier, perhaps this is the best way to handle
> it.
>

It most certainly does.

With INTERCONNECT as a bool we can handle its absence with stub
functions - like every other framework does. But as a tristate then
every driver with a call to the interconnect api needs an entry in
Kconfig to ensure the client driver must be a module if the interconnect
framework is.

Regards,
Bjorn