Re: [RFC][PATCH 12/15] videobuf2: add begin/end cpu_access callbacks to dma-sg

[Tomasz Figa]

On Tue, Feb 4, 2020 at 11:50 AM Sergey Senozhatsky
<senozhatsky@xxxxxxxxxxxx> wrote:
>
> On (20/02/03 19:04), Tomasz Figa wrote:
> [..]
> > > I very much agree with that. But this should be very clearly documented.
> > > Should V4L2_CAP_MEMORY_NON_CONSISTENT always be set in this case?
> > >
> >
> > Yes, IMHO that would make sense. My understanding is that currently
> > the consistency of allocated memory is unspecified, so it can be
> > either. With V4L2_FLAG_MEMORY_NON_CONSISTENT, the userspace can
> > explicitly ask for inconsistent memory.
> >
> > Moreover, I'd vote for setting V4L2_CAP_MEMORY_NON_CONSISTENT when
> > V4L2_FLAG_MEMORY_NON_CONSISTENT is guaranteed to return inconsistent
> > memory to avoid "optional" features or "hints" without guaranteed
> > behavior.
>
> Documentation/DMA-attributes.txt says the following
>
>   DMA_ATTR_NON_CONSISTENT
>   -----------------------
>
>   DMA_ATTR_NON_CONSISTENT lets the platform to choose to return either
>   consistent or non-consistent memory as it sees fit.  By using this API,
>   you are guaranteeing to the platform that you have all the correct and
>   necessary sync points for this memory in the driver.

Good point. And I also realized that some platforms just have no way
to make the memory inconsistent, because they may have hardware
coherency.

Then we need to keep it a hint only.

Best regards,
Tomasz