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

[Guennadi Liakhovetski <g.liakhovetski@gmx.de>]

On Fri, 16 Mar 2012, Linus Walleij wrote:

> On Fri, Mar 16, 2012 at 10:36 AM, Guennadi Liakhovetski
> <g.liakhovetski@gmx.de> wrote:
> 
> > Ok, let me try to summarise, what this would mean for sh-mobile:
> >
> > 1. this proposal introduces a new special case: with or without a mapping,
> > that will have to be handled in affected client and DMA controller
> > drivers. E.g., on sh-mobile some devices might on some systems use
> > channels from "general purpose" DMA controllers (no mapping), on other
> > systems it will be a dedicated controller (fixed mapping).
> >
> > 2. this will break, if we get more than 1 "general purpose" type with
> > different supported client sets. So, we develop a new API with a
> > pre-programmed limitation.
> 
> I fail to see why this would not be solved by a one-to-many mapping?
> 
> Flag for each device which channels it may use in a mapping
> table in platform data or device tree, I don't see the problem.
> 
> You don't even have to specify that on a per-channel basis if
> you can come up with something more clever in the mapping
> table, such as "this device can use any channel on this DMAC,
> and channels 1-7 on that DMAC" - no problem?

Sure, everything is possible. So, would something like this make you 
happy:

struct dma_channel_range {
	const char *dma_device;
	int channel_start;
	int channel_end;
};

struct dma_map {
	const char *name;
	const char *client;
	const struct dma_channel_range *chan_range;
	int chan_range_num;
};

You really want to do this?...

And the least important question: who and when will implement the core 
support for this?

> > 3. this will mean a substantial driver and platform code modification.
> > Nothing super-complex, but still some.
> 
> Big deal. Refactoring is fun... ;-)
> 
> > 4. we'll need a 3-stage channel allocation / configuration: request,
> > filter, config.
> 
> In my world: channel request with *NO* filter function.

How??? Again:

1. the client issues a dma_request_channel() with _just_ a capability mask 
and a filter and its argument as parameters - _nothing_ about channel 
restrictions.

2. you propose to eliminate a filter - the core has no way to know, which 
channel to pick up...

3. the wrapper, proposed by Russell, now calls dmaengine_slave_config(), 
which fails, because that's a wrong channel (hope I get this right this 
time - configuration has nothing to do with selection:-))

4. that's it, if you start again - the dmaengine core will enumerate the 
same channels again and propose the same unsuitable channel to you - 
there's no way to continue to the next channel / device.

What am I missing? How is the mapping going to be used, if you eliminate 
the filter function?

> Filter functions are part of the problem. So we refactor these
> away as part of this change. That's the whole point...
> 
> The core gathers information from the platform and the
> DMAC driver(s) to build up the constraints necessary to
> hand out workling channels to each device that request
> one.
> 
> And Russell IIRC already suggested a request-and-config
> channel inline for the simple cases, and if you still need to
> explicitly runtime-reconfigure then that's for a good
> reason.
> 
> > Whereas with my configuration-parameter proposal it's just
> > one stage: allocate-and-configure.
> 
> For one specific hardware, yes. For DMAengine at large
> and the majority of the drivers, no.

Sorry, why? I don't think I saw an answer to it apart from - maintenance 
burden... You can use that parameter to actually pass information to be 
used by the core to scan your mapping tables, I really don't see how you 
want to use those tables with the existing dmaengine channel-allocation 
API.

Thanks
Guennadi
---
Guennadi Liakhovetski, Ph.D.
Freelance Open-Source Software Developer
http://www.open-technology.de/