[RFC] dmaengine/dma_slave: add context parameter to prep_slave_sg callback

Guennadi Liakhovetski g.liakhovetski at gmx.de
Wed Feb 1 22:58:33 EST 2012


On Wed, 1 Feb 2012, Vinod Koul wrote:

> On Wed, 2012-02-01 at 01:09 +0100, Guennadi Liakhovetski wrote:
> > On Mon, 30 Jan 2012, Vinod Koul wrote:
> > 
> > > On Thu, 2012-01-26 at 16:22 -0500, Alexandre Bounine wrote:
> > > > As we agreed during our discussion about adding DMA Engine support for RapidIO
> > > > subsystem, RapidIO and similar clients may benefit from adding an extra context
> > > > parameter to device_prep_slave_sg() callback.
> > > > See https://lkml.org/lkml/2011/10/24/275 for more details.
> > > > 
> > > > Adding the context parameter will allow to pass client/target specific
> > > > information associated with an individual data transfer request.
> > > > 
> > > > In the case of RapidIO support this additional information consists of target
> > > > destination ID and its buffer address (which is not mapped into the local CPU
> > > > memory space). Because a single RapidIO-capable DMA channel may queue data
> > > > transfer requests to different target devices, the per-request configuration
> > > > is required.
> > > > 
> > > > The proposed change eliminates need for new subsystem-specific API.
> > > > Existing DMA_SLAVE clients will ignore the new parameter.
> > > > 
> > > > This RFC only demonstrates the API change and does not include corresponding
> > > > changes to existing DMA_SLAVE clients. Complete set of patches will be provided
> > > > after (if) this API change is accepted.
> > > This looks good to me. But was thinking if we need to add this new
> > > parameter for other slave calls (circular, interleaved, memcpy...)
> > 
> > Yes, we (shdma.c) also need to pass additional slave configuration 
> > information to the dmaengine driver and I also was thinking about 
> > extending the existing API, but my ideas were going more in the direction 
> > of adding a parameter to __dma_request_channel() along the lines of
> So your question is more on the lines of channel mapping/allocation?
> The approach here is to pass controller specific parameters which are
> required to setup the respective transfer. Since this is dependent on
> each transfer, this needs to be passed in respective prepare.
> 
> The two things are completely orthogonal and shouldn't be clubbed.
> For your issue we need a separate debate on how to solve this... I am
> open to ideas...

Well, I'm not sure whether they are necessarily always orthogonal, they 
don't seem so in my case at least. We definitely can use our approach - 
configure the channel during allocation. I _think_ we could also perform 
the configuration on a per-transfer basis, during the prepare stage, as 
this RFC is suggesting, but that definitely would require reworking the 
driver somewhat and changing the concept. The current concept is a fixed 
DMA channel allocation to slaves for as long as the slave is using DMA. 
This is simpler, avoids some overhead during operation and fits well with 
the dmaengine PRIVATE channel concept. So, given the choice, we would 
prefer to perform the configuration during channel allocation.

Maybe there are cases, where the driver absolutely needs this additional 
information during allocation, in which case my proposal would be the only 
way to go for them.

I'll post an RFC soon - stay tuned:-)

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


More information about the Linuxppc-dev mailing list