On Mon, 2012-01-30 at 10:34 +0100, Guennadi Liakhovetski wrote:

quoted

I don't still comprehend the need for a library on top of dmaengine
which gain is just a library between clients and dmacs. Surely we don't
want to write another abstraction on top of one provided?

If the question is to handle scatter-gather even if the hardware doesn't
have the capability, then why don't add that in dmaengine itself rather
than one more layer?

Well, yes, adding new abstraction layers is always a decision, that has to 
be well justified. In this case it does at least make the life easier for 
two sh-mobile drivers: shdma and the new SUDMAC driver.

However, I did name the library in a generic way without reference to sh, 
assuming, that it might with time become useful for other architectures 
too. The reasons why I prefered to keep it as an optional addition to 
dmaengine core, instead of tightly integrating it with it are, that (1) I 
did not want to add useless code to drivers, that do not need it, 

So are we sure that only sh-mobile drivers need this capablity?
Btw does you hardware only support single transfers and no sg support,
would this remain the same in future?

(2) I am 
not sure if and when this library will become useful for other drivers: 
apart from sh I am only familiar with one more dmaengine driver: 
ipu/ipu_idmac.c, and that one supports scatter-gather lists in a limited 
way and has some further peculiarities, that would likely make it a bad 
match for the simple DMA library,

typically the dmacs will support this in some form or other, so your
point is valid :)

 (3) keeping it separate makes its 
further development easier.

OTOH, I'm certainly fine with a tighter library integration with the 
dmaengine core. I think, it still would be better to keep it in a separate 
file and only build it if needed, right? This woult also simplify code 
debugging and further development. I can remove the "simple" notation, 
which does make it look like an additional abstraction layer, and replace 
it with, say, sgsoft (scatter-gather software implementation)?

that would be more apt :)

 A more 
interesting question is what to do with struct dma_simple_dev, struct 
dma_simple_chan, struct dma_simple_desc, that embed struct dma_device, 
struct dma_chan and struct dma_async_tx_descriptor respectively. I don't 
think we want to merge all the additions from those wrapping structs back 
into their dmaengine counterparts?

Sure they should be kept separate. I like the wrapping, this keeps it
simple.

How would you like to do this? Don't you think, it would be good to allow 
both: either implement a dmaengine driver directly, exactly as all drivers 
are doing now, or use the additional helper library for suitable (simple) 
hardware types? I see it similar to I2C, where you either implement an I2C 
driver directly, or you use the bitbanging abstraction for simpler 
hardware.

I think it would be good to have both, this can be used by folks who
don't have sg support available.

-- 
~Vinod