Re: [Ksummit-discuss] [CORE TOPIC] (group) maintainership models

[Peter Ujfalusi <peter.ujfalusi@ti.com>]

On 08/02/16 10:27, Arnd Bergmann wrote:
> On Tuesday, August 2, 2016 9:48:53 AM CEST Peter Ujfalusi wrote:
>> On 08/02/16 07:46, Vinod Koul wrote:
>>> On Mon, Aug 01, 2016 at 04:10:29PM +0200, Arnd Bergmann wrote:
>>>> On Monday, August 1, 2016 10:36:01 AM CEST Laurent Pinchart wrote:
>>>>> On Monday 01 Aug 2016 08:56:30 Arnd Bergmann wrote:
>>>>>> On Sunday, July 31, 2016 11:27:53 PM CEST Vinod Koul wrote:
>>>>>>> On Thu, Jul 28, 2016 at 10:44:23AM +0200, Arnd Bergmann wrote:
>>>>>>>> config FOO
>>>>>>>>
>>>>>>>>       bool "foo driver" if COMPILE_TEST && !ARCH_FOO
>>>>>>>>       default ARCH_FOO
>>>>>>>>       depends on GPIOLIB && I2C && OF && WHATEVER
>>>>>>>>
>>>>>>>> This becomes a silent always-on symbol if the platform is used,
>>>>>>>> and user-selectable on every other platform with COMPILE_TEST.
>>>>>>>
>>>>>>> Yeah this seems a bit more better. I am perhaps thinking to add top
>>>>>>> level arch dependency if required. I will try this out on DMAengine to
>>>>>>> start with and ease my build tests
>>>>>>
>>>>>> Sounds good. Let me know if you end up having to add a 'depends on ARM'
>>>>>> dependency anywhere, as that might indicate that we are doing something
>>>>>> different on ARM that should be done in a more generic way.
>>>>>
>>>>> OMAP DMA comes to mind, we still depend on the implementation provided by 
>>>>> arch/arm/plat-omap/dma.c.
>>>>
>>>> That's a dependency on MACH_OMAP instead of ARM though, which is a bit
>>>> different.
>>>>
>>>> (Getting offtopic and) speaking of this one, is anyone working
>>>> on converting the last four or five drivers to the dmaengine framework?
>>>
>>> Rather lets' move this conversation to dmaengine list?
>>>
>>>> drivers/media/platform/omap/omap_vout_vrfb.c
>>
>> Needs the interleaved transfer type support in omap-dma. Now it is in -next.
> 
> Ok
> 
>>>> drivers/mtd/onenand/omap2.c
>>
>> Using memcpy, I have sent the patches:
>> https://patchwork.kernel.org/patch/7842891/
>> https://patchwork.kernel.org/patch/7851881/
>>
>> but I need to revisit. I think the 'fix' will be to drop the DMA support from it.
> 
> Sounds like it might even possible to remove the entire driver based
> on that thread.
> 
>>>> drivers/usb/gadget/udc/omap_udc.c
>>>> drivers/usb/musb/tusb6010_omap.c
>>
>> They are mostly slave implementations, but. As I recall they do some really
>> interesting non symmetric setup for the TX/RX DMA which I still need to figure
>> out. And I'm looking for HW to check for regression...
> 
> Ok
> 
>>>> (and a bit of drivers/video/fbdev/omap/omapfb_main.c)
>>
>> Hrm, that is OMAP1. Looks like highly coupled with mach-omap1/lcd_dma.c. This
>> is going to be a bit complicated IMHO.
> 
> It's just a single function call: 
> 
> omap_set_dma_priority(0, OMAP_DMA_PORT_EMIFF, 15);
> 
> We can probably open-code that in mach-omap1/lcd_dma.c and
> have a callback pointer in the platform_data.

Looking at the omap_set_dma_priority(), it will grant high priority for _all_
DMA access via EMIFF (or EMIFS, OCPT1, OCPT2). It is global priority over
targets. The same register is used to give 'priority' to OCPI, DMA, DSP or MPU
masters.
The whole priority configuration should get it's own API, something like:
omap1_set_ocpi_priority(master, target, priority);

I will see what is the best when I get to this.

>>>> I guess we could improve build coverage by moving 
>>>> arch/arm/plat-omap/dma.c into drivers/dma/omap-dma.c, but that
>>>> has the disadvantage of exposing a nonstandard interface from
>>>> somewhere inside of the dmaengine subsystem.
>>>
>>> yeah that won't be nice. I think we should get these removed... We have omap
>>> as well as edma driver in dmaengine.
>>
>> I'm not going to move the legacy API under dmaengine, it makes no sense.
>>
>> My plan is to do what I did with the eDMA driver stack:
>> - convert drivers using the legacy API to dmaengine
>> - 'merge' the code from plat-omap to the dmaengine driver
>>   - while doing this the legacy API will vanish along with the plat-omap/dma.c
>>
>> As with the eDMA, the sDMA stack will need some cleanup, but the good thing is
>> that the dmaengine driver have minimal dependency on the legacy
>> plat-omap/dma.c API.
> 
> Ok, thanks for the background information. I agree on the approach, this
> is the right way to do it. I just wasn't sure if there was someone actually
> working on it, and it's clear that you are, even if it will take a while.

The eDMA got higher priority over the sDMA. Now that we have interleaved
support via omap-dma, we can start working on this. Only the most problematic
drivers were left to convert (or the ones no one cares about?).

-- 
Péter