Re: [PATCH/RFC] dmaengine: add a slave parameter to __dma_request_channel()

[Vinod Koul <vinod.koul@intel.com>]

On Thu, 2012-03-08 at 17:04 +0530, Vinod Koul wrote:
> On Thu, 2012-03-08 at 12:22 +0100, Guennadi Liakhovetski wrote:
> > On Thu, 8 Mar 2012, Vinod Koul wrote:
> > 
> > > On Thu, 2012-03-08 at 11:16 +0100, Guennadi Liakhovetski wrote:
> > > > I still have the impression, that my specific use-case (sh-mobile), where 
> > > > channels can be freely configured for use by _ANY_ client on one of 
> > > > _SEVERAL_ DMAC instances, is not fully understood or taken into account. 
> > > > For this driver any kind of fixed mapping means, that we'd have to use 
> > > > both virtual channels and controllers, adding _a lot_ of complexity to the 
> > > > DMAC driver and making the dmaengine core just an "obfuscation layer." 
> > > > Yes, I remember Russell proposing core helpers for this. They would help, 
> > > > but (1) when would they be available, (2) how well would they be suitable 
> > > > for us, (3) they'd take the coding / maintainance burden away, but 
> > > > wouldn't reduce complexity and run-time overhead. 
> > > Lets try to address you case as well.
> > > On a typical platform
> > 
> > Let's take the mackerel board with the sh7372 SoC. it's not the state of 
> > the art, but that's what I'm currently working with and it should give us 
> > a good enough idea
> > 
> > > 1) how many dma controllers you have?
> > 
> > currently supported 5 of 3 types (3 of type A, 1 of each of the types B 
> > and C), all handled by the same driver
> > 
> > > 2) how many clients you have
> > 
> > huh... many. Maybe like 20 or more, and more, that are not yet supported, 
> > using type A, and 1 for each of types B and C
> > 
> > > 3) which client can use what controller channel? How is mapping decided,
> > > do you have a mux, is it hard wired by soc designers,....?
> > 
> > In general - with all the current sh-mobile hardware, that I'm aware of - 
> > there can be several controller instances on an SoC of each controller 
> > type. Inside each type all instances and all channels are freely 
> > configurable. So, of 20 Type A clients they can use any channels on any 
> > one of the 3 type A controllers. Types B and C are "degenerate" cases, 
> > there clients are practically hard-wired to a specific DMA controller.
> > 
> > So, we don't have to decide on mappings for type A. We just pick up any 
> > free channels on any controller and configure them accordingly. Whether 
> > there's a mux somewhere - you can say so, but it's all inside the SoC, and 
> > it's configured automatically ones you configure a physical channel to 
> > serve a specific client.
> > 
> > > Can you pls give a description so that we ensure all models fit in the
> > > final solution?
> > 
> > That's what I've been trying to do since several days now... I've been 
> > saying "multiple controllers with multiple channels all freely 
> > configurable for any device from a list" again and again... Seems I'm 
> > speaking some strange language, that noone understands.
> Okay. One more question before I can tell you how it can work for you
> without you sweating it out :-)
> 
> So you have:
> case A: Here you have N dmacs and M controllers, any controller can use
> any channel, No constraints on channel assignments, right?
> case B: Some hardwired controllers P which can only be used by a set
> clients Q?
> 
> Anything else I missed in your description?
Assuming I didn't miss...

The case B can be handled without sweat by platforms channel mapping
information.

Case A where we don't find that devices exist in map, thus being treated
as generic DMA channels and can be handled easily in sequence. So when
someone in Q request a channel it would get first channel in Ps

This way we handle both of them in a transparent manner to both clients
and controllers. 

Perhaps we can also add capability to know that if channel is to be
searched in map or not - would be anyway required for non slave cases.

-- 
~Vinod