Hi Vinod, On Fri, Jan 03, 2020 at 02:59:18AM +0200, Laurent Pinchart wrote: > On Fri, Dec 20, 2019 at 06:02:32PM +0200, Laurent Pinchart wrote: > > On Fri, Dec 20, 2019 at 09:10:59PM +0530, Vinod Koul wrote: > >> On 20-12-19, 14:35, Laurent Pinchart wrote: > >>> On Fri, Dec 20, 2019 at 01:31:27PM +0530, Vinod Koul wrote: > >>>> On 20-12-19, 07:13, Laurent Pinchart wrote: > >>>> > >>>>>> OK, in the light of this information I'll keep the two separate and will > >>>>>> switch to vchan as requested by Vinod. > >>>>> > >>>>> I've moved forward with this task, but eventually ran into one hack in > >>>>> the driver that is more difficult to get rid of than the other ones. > >>>>> > >>>>> For display operation, the DPSUB driver needs to submit cyclic > >>>>> interleaved transfer requests. There's no such thing (as far as I can > >>>>> tell) in the DMA engine API, so the DPDMA drive simply keeps processing > >>>> > >>>> we do support interleave, you need to implement > >>>> .device_prep_interleaved_dma and use dmaengine_prep_interleaved_dma() > >>>> from the client > >>> > >>> I mean both interleaved and cyclic at the same time. > >>> > >>>>> the same descriptor over and over again until a new one is issued. The > >>>>> hardware supports this with the help of hardware-based chaining of > >>>>> descriptors, and the DPDMA driver simply sets the next pointer of the > >>>>> descriptor to itself. > >>>>> > >>>>> How can I solve this in a way that wouldn't abuse the DMA engine API ? > >>>> > >>>> Is this not a cyclic case of descriptor? > >>> > >>> Exactly my point :-) It's cyclic, but has to be interleaved too as it's > >>> a 2D transfer. > >> > >> IIRC the interleaved descriptor can be set in such a way that last chunk > >> points to the first one.. > > > > I don't see a way to do this in the existing API, am I missing something > > ? And how would the completion handler be called in that case, once per > > frame still ? I don't think vchan supports this at the moment > > > >> I think Jassi had good ideas for generic interleave API which can do > >> all this :) > > > > How do I get this driver moving forward in the meantime ? :-) > > Happy new year, and gentle ping :-) Would it be fine if, in the meantime, I used the vchan helpers by hardcoded interleaved descriptors to always be cyclic ? The DPDMA IP is tied to the DPSUB in existing platforms, so there's no way to reuse it in a generic fashion at the moment anyway. We could then extend this to merge interleaved and cyclic modes in a single API, and I would update both the DPDMA and the DPSUB driver accordingly. Please let me know if that would be acceptable as an interim upstreamable solution. -- Regards, Laurent Pinchart
