Re: [PATCH v9 15/19] media: ti: j721e-csi2rx: Change the drain architecture for multistream

[Rishikesh Donadkar]

On 15/01/26 21:32, Jai Luthra wrote:
> Hi Tomi,
>
> Quoting Tomi Valkeinen (2026-01-15 18:07:07)
> > Hi,
> >
> > On 30/12/2025 10:32, Rishikesh Donadkar wrote:
> > > On buffer starvation the DMA is marked IDLE, and the stale data in the
> > > internal FIFOs gets drained only on the next VIDIOC_QBUF call from the
> > > userspace. This approach works fine for a single stream case.
> > >
> > > But in multistream scenarios, buffer starvation for one stream i.e. one
> > > virtual channel, can block the shared HW FIFO of the CSI2RX IP. This can
> > > stall the pipeline for all other virtual channels, even if buffers are
> > > available for them.
> > One stream is filtered based on VC & DT, but the above only mentions VC.
> > And then later uses VC when referring to the stream. I think you can
> > drop the VC parts, and just talk about streams.
> >
> > > This patch introduces a new architecture, that continuously drains data
> > > from the shared HW FIFO into a small (32KiB) buffer if no buffers are made
> > > available to the driver from the userspace. This ensures independence
> > > between different streams, where a slower downstream element for one
> > > camera does not block streaming for other cameras.
> > >
> > > Additionally, after a drain is done for a VC, the next frame will be a
> > > partial frame, as a portion of its data will have already been drained
> > > before a valid buffer is queued by user space to the driver.
> > This makes it sounds we drain a single 32KB piece. But won't we continue
> > draining that stream until the stream is stopped or the user provides a
> > buffer?
> >
> > Also, does the DMA not offer us ways to drain a full frame? There's no
> > way to e.g. set the DMA TX increment to 0, so that it would just write
> > to a single location in memory? Or just set the target to void.
> + Vignesh
>
> AFAIU the dmaengine API is the first limitation here, and the actual
> Transfer Record (TR) structure for BCDMA might support writing to a single
> address. It also might be possible to use dmaengine_prep_dma_cyclic to
> setup a auto-repeating TR like it's used for audio.. maybe that can be
> explored separate from this series.
Makes sense, if we have such API we can drain a full frame as Tomi 
suggested and also not care about the next frame being a partial one.
> But yes, ideally we need to set the target to void, which I don't think is
> supported by the HW (TI folks please correct me if I'm wrong :)
>
> Thanks,
>      Jai
>
> >   Tomi
> >
> > > Use wait for completion barrier to make sure the shared hardware FIFO
> > > is cleared of the data at the end of stream after the source has stopped
> > > sending data.
> > >
> > > Reviewed-by: Jai Luthra <jai.luthra@xxxxxxxxxxxxxxxx>
> > > Reviewed-by: Yemike Abhilash Chandra <y-abhilashchandra@xxxxxx>
> > > Signed-off-by: Rishikesh Donadkar <r-donadkar@xxxxxx>
> [...]