On Tue, 2012-01-31 at 09:59 +0100, Guennadi Liakhovetski wrote:
Hi Vinod
Thanks for your reply.
On Tue, 31 Jan 2012, Vinod Koul wrote:
On Mon, 2012-01-30 at 10:34 +0100, Guennadi Liakhovetski wrote:
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?
If I was sure about that, I'd call the library shmobile-dma or something similar:-)
Btw does you hardware only support single transfers and no sg support,
Well, the controller does have some repeat- and reload-modes, but they are currently not supported by the driver and from the documentation I cannot easily understand, how useful those modes could be to implement SG.
would this remain the same in future?
Who knows?...
(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.
Ok, so, should I just rename the driver to sgsoft? Is this the only change, that you'd like to see?
That would be one :) Also, I would review the other patch patches by today/tomorrow, you can add any changes in next version as well