[alsa-devel] [PATCH 2/3] ASoC: Davinci: pcm: add support for sram-support-less platforms

[Matt Porter]

On Tue, Oct 02, 2012 at 06:50:14PM +0200, Daniel Mack wrote:
> On 02.10.2012 18:41, Matt Porter wrote:
> > On Tue, Oct 02, 2012 at 03:42:47PM +0200, Daniel Mack wrote:
> >> On 02.10.2012 13:06, Sekhar Nori wrote:
> >>> On 10/2/2012 4:03 PM, Daniel Mack wrote:
> >>>> On 02.10.2012 11:37, Mark Brown wrote:
> >>>>> On Tue, Oct 02, 2012 at 10:48:53AM +0300, Peter Ujfalusi wrote:
> >>>>>
> >>>>>> I also agree that ifdef is not a good solution.
> >>>>>> It is better to have this information passed as device_data and via DT it can
> >>>>>> be decided based on the compatible property for the device.
> >>>>>
> >>>>> That's not really the problem here - the problem is that the APIs used
> >>>>> to get the SRAM are DaVinci only so it's not possible to build on OMAP
> >>>>> or other platforms.  The SRAM code needs to move to a standard API.
> >>>>
> >>>> What about following Matt Porter's idea and ignore the SRAM code
> >>>> entirely and port the entire PCM code to generic dmaengine code first?
> >>>> The EDMA driver needs to learn support for cyclic DMA for that, and I
> >>>> might give that a try in near future.
> >>>>
> >>>> Later on, the SRAM ping-pong code can get added back using genalloc
> >>>> functions, as Sekhar proposed. That needs to be done by someone who has
> >>>> access to a Davinci board though, I only have a AM33xx/OMAP here.
> >>>
> >>> We cannot "get rid" of SRAM code and add it back "later". It is required
> >>> for most DaVinci parts. The SRAM code can be converted to use genalloc
> >>> (conversion should be straightforward and I can help test it) and the
> >>> code that uses SRAM can probably keep using the private EDMA API till
> >>> the dmaengine EDMA driver has cyclic DMA support. Matt has already
> >>> posted patches to move private EDMA APIs to a common location. That
> >>> should keep AM335x build from breaking. Is this something that is feasible?
> >>
> >> Yes - by "later" I just meant in a subsequent patch. But you're probably
> >> right, we can also do that first.
> >>
> >> I'm looking at that right now and the problem seems that we don't have a
> >> sane way to dynamically look up gen_pools independently of the actual
> >> run-time platform. All users of gen_pools seem to know which platform
> >> they run on and access a platform-specific pool. So I don't currently
> >> see how we could implement multi-platform code, gen_pools are fine but
> >> don't solve the problem here.
> >>
> >> Would it be an idea to add a char* member to gen_pools and a function
> >> that can be used to dynamically look it up again? If a buffer with a
> >> certain name exists, we can use it and install that ping-pong buffer,
> >> otherwise we just don't. While that would be easy to do, it's a
> >> tree-wide change.
> >>
> >> Is there a better way that I miss?
> > 
> > At the high level there's two platform models we have to handle, the
> > boot from board file !DT case, and then the boot from DT case. Since
> > Davinci is just starting DT conversion, we mostly care about the !DT
> > base in which the struct gen_pool * is passed in pdata to the ASoC
> > driver. It is then selectable on a per-platform basis where the decision
> > should be made.
> > 
> > Given a separate discussion with Sekhar, we're only going to have one
> > SRAM pool on any DaVinci part right now...this was only a question on
> > L138 anyway. But regardless, the created gen_pool will be passed via
> > pdata.
> 
> I thought about this too, as mmp does it that way.
> 
> > Since DT conversion is starting and we need to consider that now,
> > the idea there is to use the DT-based generic sram driver [1] such that
> > when we do boot from DT on Davinci, the genpool is provided via phandle
> > and the pointer extracted with the OF helpers that are part of the
> > series.
> 
> A phandle is the cleanest way I think, yes.

See the of_get_named_gen_pool() helper example in the series
http://comments.gmane.org/gmane.linux.kernel/1352210

> > That's pretty much it. I'm reworking the backend support as discussed
> > with Sekhar wrt to my uio_pruss series. I can post a standalone series
> > that just replaces sram_* with genalloc for davinci ASoC.
> 
> As you can also test it, it would be easiest if you came up with a patch
> for that, yes. I can have a look at the dma bits laters, once my OMAP
> board finally works with the code as it currently stands. I'm still
> fighting with the mcasp driver right now ...

Ok. Also, as Sekhar pointed out, dmaengine itself isn't a blocker since
we can have AM33xx use the private EDMA API for ASoC until I finish
cyclic DMA support. Handling the ping-pong and normal case transparently
within the dmaengine driver will require a further look but at least
genalloc is our first step there.

-Matt