[PATCH v7 01/10] ARM: davinci: move private EDMA API to arm/common

[Matt Porter]

On Sat, Feb 02, 2013 at 07:06:06PM +0000, Sergei Shtylyov wrote:
> Hello.
> 
> On 02-02-2013 22:07, Matt Porter wrote:
> 
> >>>>>>> Move mach-davinci/dma.c to common/edma.c so it can be used
> >>>>>>> by OMAP (specifically AM33xx) as well.
> 
> >>>>>> I think this should rather go to drivers/dma/?
> 
> >>>>> No, this is the private EDMA API. It's the analogous thing to
> >>>>> the private OMAP dma API that is in plat-omap/dma.c. The actual
> >>>>> dmaengine driver is in drivers/dma/edma.c as a wrapper around
> >>>>> this...same way OMAP DMA engine conversion is being done.
> 
> >>>>     Keeps me wondering why we couldn't have the same with CPPI 4.1 when I proposed
> >>>> that, instead of waiting indefinitely for TI to convert it to drivers/dma/
> >>>> directly. We could have working MUSB DMA on OMAP-L1x/Sitara all this time... Sigh.
> 
> >>> That is a shame. Yeah, I've pointed out that I was doing this exactly
> >>> the same way as was acceptable for the OMAP DMA conversion since it was
> >>> in RFC. The reasons are sound since in both cases, we have many drivers
> >>> to convert that need to continue using the private DMA APIs.
> 
> >>      In case of CPPI 4.1, we'd only have to convert MUSB DMA driver. Other
> >> in-tree CPPI 4.1 having SoCs don't use it for anything but MUSB -- it even is
> >> sub-block of their MUSB device, AFAIK (I maybe wrong about Sitaras -- I don't
> >> know them well).
> 
> > Well, it's pretty clear to me now that there's good reason for it not
> > landing in arch/arm/ so the obvious path is to do the dmaengine
> > conversion and put it in drivers/dma/ if it's really a generic dma engine.
> > I'm not sure why you express concern over the dma engine api not fitting
> > with CPPI 4.1?
> 
>     It's not a DMA controller only, it's 3 distinct devices, with the DMA 
> controller being one among them and using another one, the queue manager, as 
> some sort of proxy. The third device doesn't exist on OMAP-L1x SoCs -- it's 
> the buffer manager.

Interesting, and you don't see a way to have this split between
dmaengine and the musb driver? This all assumes there's even a match as
RMK did raise concerns elsewhere in this thread.

> > If it doesn't work, work with Vinod to fix the api. It's
> > expected, I'm working on dmaengine API changes right now to deal with a
> > limitation of EDMA that needs to be abstracted.
> 
>     Sorry, now it's TI's task. I no longer have time to work on this (my 
> internal project to push OMAP-L1x support upstream has expired at Sep 2010) 
> and my future in MV is very uncertain at this moment. Most probably I'll leave 
> it (or be forced to leave).

Too bad, it's certainly "somebody's task". The people employed by TI
have names too. ;) I suspect it falls to Felipe or Kishon these days,
but I try to avoid USB as it's generally a source of pain.

> > As pointed out, edma.c is already here in arch/arm/, so moving it doesn't
> > add something new. It does let us regression test all platforms that use it
> > (both Davinci and AM33xx) as I go through the conversion process.
> 
>     Could have been the same with CPPI 4.1 in theory if it was added to 
> mach-davinci/ back in 2009... we'd then only have to move it. EDMA code is 
> much older of course, so it's probably more justified.

Absolutely, timing is everything. I can assure you I've flamed enough
people "internally" about leaving this EDMA dmaengine conversion for so
long. As you might have guessed, AM33xx is a bit of a driver for it, but
all of this would have been quite a bit simpler had somebody taken a
little effort and moved edma to dmaengine years ago when slave dma
support was added to dmaengine. ;)

-Matt