[PATCH 00/23] RFC: exynos multiplatform support

[tomasz.figa@gmail.com (Tomasz Figa)]

On Tuesday 05 of March 2013 19:19:02 Arnd Bergmann wrote:
> On Tuesday 05 March 2013, Tomasz Figa wrote:
> > > With this patch set, we can build mach-exynos as part
> > > of a multiplatform kernel, with the following caveats:
> > > 
> > > * Only DT based boards are supported
> > 
> > As far as I'm aware, there are plans to drop non-DT Exynos support.
> > Kgene should know more on this.
> 
> Yes, that was my understanding as well. It might not actually be too
> hard to get multiplatform working with the ATAGS based board files
> (we do that on some of the other platforms), but there is probably
> no reason to try hard.
> 
> > > * Moving to common-clk breaks things including cpufreq
> > > 
> > >   and others. Thomas is working on a patch for this
> > 
> > We have several patches internally for fixing things up to run
> > correclty with common-clk. I'll see if we can post some patches.
> 
> Ok, excellent.
> 
> > > * We disable the gpio implementation, which also breaks
> > > 
> > >   stuff, but Thomas has a patch already
> > 
> > The legacy GPIO code is needed only for non-DT case. DT-case uses the
> > new pinctrl-samsung driver, which is (AFAIK) multiplatform-aware.
> 
> Please have a closer look at the "ARM: exynos: work around missing gpio
> code on multiplatform" patch, I think there a few files I had to touch
> that actually rely on the legacy gpio code:
> 
> * the pm-gpio.c file
> * the s3c_i2c0_cfg_gpio function
> * the eint irqchip
> 
> If none of these are needed for DT-based systems, we should probably
> build that code conditionally based on the CONFIG_EXYNOS_ATAGS symbol
> I introduced.

Yes, none of them are needed for DT-based systems.

> > > * sparsemem support is not available on multiplatform
> > 
> > There was some discussion some time ago whether we really need
> > sparsemem on Exynos. If I remember correctly, it turned out that we
> > don't. So this is not really an issue.
> 
> Ok, good.
> 
> > > Arnd Bergmann (23):
> > >   ARM: exynos: introduce EXYNOS_ATAGS symbol
> > >   irqchip: exynos: remove dependency on mach/irqs.h
> > >   tty: serial/samsung: prepare for common clock API
> > >   tty: serial/samsung: make register definitions global
> > >   tty: serial/samsung: fix modular build
> > >   ARM: exynos: move debug-macro.S to include/debug/
> > >   i2c: s3c2410: make header file local
> > >   mmc: sdhci-s3c: remove platform dependencies
> > >   usb: exynos: do not include plat/usb-phy.h
> > >   [media] exynos: remove unnecessary header inclusions
> > >   video/exynos: remove unnecessary header inclusions
> > >   thermal/exynos: remove unnecessary header inclusions
> > >   mtd: onenand/samsung: make regs-onenand.h file local
> > >   rtc: s3c: make header file local
> > >   spi: s3c64xx: move to generic dmaengine API
> > >   pwm: samsung: repair the worst MMIO abuses
> > 
> > I'm currently working (in my free time) on a series of cleanup patches
> > sanitizing Samsung PWM code for S3C64xx and S5PV210 DT (and
> > multiplatform) support.
> 
> Ah, nice.
> 
> > On those platforms it is a bit more complex case as there are two
> > blocks of code that access the same hardware - samsung-time (using
> > two PWM channels for clocksource and clock events) and pwm-samsung.
> > 
> > Hopefully, I will have some patches ready soon.
> 
> How are you planning to solve this? Do you want to have a combined
> driver that registers both a clocksource and a pwm?

Let's start with a quick introduction to the s3c-pwm hardware. Each 
channel has its own timer value, compare and reload value registers, so 
they don't need any extra locking. However there are additional shared 
configuration registers, containing things such as start and reload bits 
for all channels, prescaler and main divisor values, etc. Those registers 
needs synchronization.

Now there are several possible approaches:

1) A brute force one - two separate drivers, based on the fact that the
   clocksource driver will be used only on uniprocessor systems, so
   a simple _irqsave when accessing a shared register in both will 
   guarantee correct synchronization.

2) Two separate drivers with some synchronized shared code accessing 
   registers (_start, _stop, _set_reload, _set_prescaler, _set_divisor, 
   etc.; all using a shared spinlock).

3) Single driver registering PWM channels, clocksource and clock event 
   device. This does not seem like a bad idea, since the whole code for 
   configuration of the PWM block would reside in one location and there 
   would be no redundancy. However there is a question where such driver 
   should be placed - drivers/clocksource, drivers/pwm, or maybe somewhere 
   else?

Personally I wanted to go with first option, which would require least 
amount of changes to existing code, at a cost of some code duplication 
(but some PWM code is duplicated already).

Best regards,
Tomasz