[PATCH] dmaengine: add CSR SiRFprimaII DMAC driver

[Barry Song]

2011/9/12 Linus Walleij <linus.walleij at linaro.org>:
> On Thu, Sep 8, 2011 at 11:38 PM, Vinod Koul <vkoul at infradead.org> wrote:
>> On Thu, 2011-09-08 at 22:11 +0200, Arnd Bergmann wrote:
>>> On Thursday 08 September 2011 20:48:26 Linus Walleij wrote:
>>> > 2011/9/8 Arnd Bergmann <arnd at arndb.de>:
>>> > > On Thursday 08 September 2011, Barry Song wrote:
>>> > >>
>>> > >> this filter is used by all drivers with DMA since every dma channel is
>>> > >> fixed to be assigned to one device.
>>> > >
>>> > > Ok, I see now. I think it would be best to introduce a generic
>>> > > 'filter by device tree property' function or alternatively an
>>> > > dma_of_request_channel function like this:
>>> >
>>> > You'd have to discuss that with Vinod, the thing is that x86 Atom
>>> > systems are using dmaengine for device slave transfers too, and
>>> > IIRC these things don't use devicetrees. I may be wrong...
>>>
>>> Some of them use device tree, some don't.
>>>
>>> I'm not saying that we have to convert all drivers to use this, but
>>> for platforms that always have device tree available, it seems by far
>>> the cleanest solution.
>> We don't have a very clean solution for filter function in case of slave
>> dmaengine. How should the client specify which channel it wants is not
>> really clear.
>>
>> We can look at device tree but that's something which wont work in case
>> of non device tree platforms (atom x86).
>> What we need is this information of channel mapping, which IMO is
>> platform specific and needs to come from platform data, we can abstract
>> it actually from the device tree data/PCI/firmware etc but essentially a
>> mechanism to publish channels and slaves uniquely and match them in this
>> kind of data
>>
>> Linus W, any progress on that patches you posted??
>
> I haven't written any, and I felt the whole issue was pretty inflamed
> too so I felt bad about it and avoided to think about it even since
> I have no real problem with this in my current setups.
>
> Currently there is a strong coupling between platforms and filter
> functions and I can live with it in the systems I use since they have
> just one DMAC and need only one filter function per device,
> that is specified in platform data for the device. This would likely
> also work for the SiRFprimaII if it has only a single DMAC.
> The people facing an immediate issue with this are IIRC the
> Samsung S5Ps.
>
> If you think this is in need of solving soon and want me to propose
> patches for channel mapping to devices using the approach used in
> clkdev and regulator APIs to create an attributed mapping table
> using struct device * or its string representations, I can
> try it out of course, but if it gets flamy I will just back off again.

the attributed mapping table is something that is rubbish Linus
Torvalds wants to say as people are waiting for the new clkdev
framework to eliminate the long and trivial table, those are totally
useless codes to present hardware details.

if we want the mapping table, we want it in dt. maybe something like
current gpio makes more sense. devices just list the gpio range in
dts. i guess DMA is just like iomem and gpio, which is just device
resource. then we might make it just like and "reg" and "gpio".

for platforms without dt, write the dma resouce in mach and add a
common filter function which filters the channel id might be simple.

>
> Thanks,
> Linus Walleij
>
-barry