Re: [PATCH 00/12] introduce support for early platform drivers

From: Geert Uytterhoeven
Date: Thu May 31 2018 - 02:43:01 EST


Hi Rob,

On Thu, May 31, 2018 at 12:36 AM, Rob Herring <robh+dt@xxxxxxxxxx> wrote:
> On Wed, May 30, 2018 at 2:40 PM, Michael Turquette
> <mturquette@xxxxxxxxxxxx> wrote:
>> Quoting Rob Herring (2018-05-14 06:20:57)
>>> On Mon, May 14, 2018 at 6:38 AM, Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
>>> > 2018-05-11 22:13 GMT+02:00 Rob Herring <robh+dt@xxxxxxxxxx>:
>>> >> On Fri, May 11, 2018 at 11:20 AM, Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
>>> >>> This series is a follow-up to the RFC[1] posted a couple days ago.
>>> >>>
>>> >>> NOTE: this series applies on top of my recent patches[2] that move the previous
>>> >>> implementation of early platform devices to arch/sh.
>>> >>>
>>> >>> Problem:
>>> >>>
>>> >>> Certain class of devices, such as timers, certain clock drivers and irq chip
>>> >>> drivers need to be probed early in the boot sequence. The currently preferred
>>> >>> approach is using one of the OF_DECLARE() macros. This however does not create
>>> >>> a platform device which has many drawbacks - such as not being able to use
>>> >>> devres routines, dev_ log functions or no way of deferring the init OF function
>>> >>> if some other resources are missing.
>>> >>
>>> >> I skimmed though this and it doesn't look horrible (how's that for
>>> >> positive feedback? ;) ). But before going into the details, I think
>>> >> first there needs to be agreement this is the right direction.
>>> >>
>>> >> The question does remain though as to whether this class of devices
>>> >> should be platform drivers. They can't be modules. They can't be
>>> >> hotplugged. Can they be runtime-pm enabled? So the advantage is ...
>>> >>
>>> >
>>> > The main (but not the only) advantage for drivers that can both be
>>> > platform drivers and OF_DECLARE drivers is that we get a single entry
>>> > point and can reuse code without resorting to checking if (!dev). It
>>> > results in more consistent code base. Another big advantage is
>>> > consolidation of device tree and machine code for SoC drivers used in
>>> > different boards of which some are still using board files and others
>>> > are defined in DT (see: DaVinci).
>>> >
>>> >> I assume that the clock maintainers had some reason to move clocks to
>>> >> be platform drivers. It's just not clear to me what that was.
>>> >>
>>> >>> For drivers that use both platform drivers and OF_DECLARE the situation is even
>>> >>> more complicated as the code needs to take into account that there can possibly
>>> >>> be no struct device present. For a specific use case that we're having problems
>>> >>> with, please refer to the recent DaVinci common-clock conversion patches and
>>> >>> the nasty workaround that this problem implies[3].
>>> >>
>>> >> So devm_kzalloc will work with this solution? Why did we need
>>> >> devm_kzalloc in the first place? The clocks can never be removed and
>>> >> cleaning up on error paths is kind of pointless. The system would be
>>> >> hosed, right?
>>> >>
>>> >
>>> > It depends - not all clocks are necessary for system to boot.
>>>
>>> That doesn't matter. You have a single driver for all/most of the
>>> clocks, so the driver can't be removed.
>>
>> -ECANOFWORMS
>>
>> A couple of quick rebuttals, but I imagine we're going to disagree on
>> the platform_driver thing as a matter of taste no matter what...
>
> It's really more should the clocksource, clockevents and primary
> interrupt controller be drivers. Let's get agreement on that first. If
> yes, then it probably does make sense that their dependencies are
> drivers too. If not, then making only the dependencies drivers doesn't
> seem right to me.

Yes, they should all be drivers.

Assuming clocksources, clockevents, or primary interrupt controllers are
special, and thus forcing everyone to use OF_DECLARE for the whole
subsystem, doesn't fly everywhere.

Clockevents and interrupt controllers can have a module clock.
All three can be part of a PM Domain, which requires a struct device to
be handled properly.

>> 1) There should be multiple clk drivers in a properly modeled system.
>> Some folks still incorrectly put all clocks in a system into a single
>> driver because That's How We Used To Do It, and some systems (simpler
>> ones) really only have a single clock generator IP block.
>>
>> Excepting those two reasons above, we really should have separate
>> drivers for clocks controlled by the PMIC, for the (one or more) clock
>> generator blocks inside of the AP/SoC, and then even more for the
>> drivers that map to IP blocks that have their own clock gens.
>
> I agree those should be separate entities at least. But for a given
> h/w block, if you already have to use OF_DECLARE, why would you try to
> split that into OF_DECLARE and a driver? what advantage does putting
> non-boot essential clocks in a driver or transitioning to a driver get
> you?

You may want to split it because of dependencies. OF_DECLARE
doesn't handle EPROBE_DEFER, while some critical parts may be
needed early.

> And I've seen PMIC clocks could be inputs to the SoC's clock
> controller(s), so the dependencies get more complicated. Then does the
> PMIC driver and its dependencies need to be early drivers too?

Especially if there are clock loops, but that's something different.
Cfr. cs2000 in arch/arm64/boot/dts/renesas/salvator-common.dtsi, which
provides a clock from the main SoC, but also consumes a clock from the main
SoC. The latter is modeled in DT as a fixed-clock as a workaround.

>> Good examples of the latter are display controllers that generate their
>> own PLL and pixel clock. Or MMC controllers that have a
>> runtime-programmable clock divider. Examples of these are merged into
>> mainline.
>
> But those are drivers of types other than a clock controller that
> happen to register some clocks as well. I wasn't saying these cases
> can't or shouldn't be part of the driver model. Look at irqchips. We
> have some that use the driver model (e.g. every GPIO driver) and some
> that don't because there's no need (e.g. GIC).

There is a need for using the driver model for the GIC. On some platforms,
the GIC can be part of a clock and/or power domain.

For secondary GICs, that got fixed in commit 9c8edddfc9924cb4
("irqchip/gic: Add platform driver for non-root GICs that require RPM").
For primary GICs, it's still not fixed, so we have to live with
CLK_IS_CRITICAL, and rely on other critical devices in the same power
domain to avoid the power domain being powered off.

>> 2) Stephen and I wanted clock drivers to actually be represented in the
>> driver model. There were these gigantic clock drivers that exclusively
>> used CLK_OF_DECLARE and they just sort of floated out there in the
>> ether... no representation in sysfs, no struct device to map onto a
>> clock controller struct, etc.
>>
>> I'd be happy to hear why you think that platform_driver is a bad fit for
>> a device driver that generally manages memory-mapped system resources
>> that are part of the system glue and not really tied to a specific bus.
>> Sounds like a good fit to me.
>>
>> If platform_driver doesn't handle the early boot thing well, that tells
>> me that we have a problem to solve in platform_driver, not in the clk
>> subsystem or drivers.
>
> Doing things earlier is not the only way to solve the problems.
> Perhaps we need to figure out how to start things later. Maybe it's
> not feasible here, I don't know.

The fixed probe order imposed by OF_DECLARE() limits this: if your
OF_DECLARE() driver depends on something else, the latter must become an
early device.
If all subsystems would use real devices, EPROBE_DEFER would handle most of
it automatically.

Then, next step would be to avoid triggering EPROBE_DEFER, e.g. by
analyzing dependencies in DT...

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds