Re: [PATCH 1/2] media: docs-rst: Document memory-to-memory video decoder interface

[Nicolas Dufresne]

Le lundi 15 octobre 2018 Ã 19:13 +0900, Tomasz Figa a Ãcrit :
> On Wed, Aug 8, 2018 at 11:55 AM Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
> > 
> > On Tue, Aug 7, 2018 at 4:37 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
> > > 
> > > On 08/07/2018 09:05 AM, Tomasz Figa wrote:
> > > > On Thu, Jul 26, 2018 at 7:57 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
> > > > > > > I wonder if we should make these min buffer controls required. It might be easier
> > > > > > > that way.
> > > > > > 
> > > > > > Agreed. Although userspace is still free to ignore it, because REQBUFS
> > > > > > would do the right thing anyway.
> > > > > 
> > > > > It's never been entirely clear to me what the purpose of those min buffers controls
> > > > > is. REQBUFS ensures that the number of buffers is at least the minimum needed to
> > > > > make the HW work. So why would you need these controls? It only makes sense if they
> > > > > return something different from REQBUFS.
> > > > > 
> > > > 
> > > > The purpose of those controls is to let the client allocate a number
> > > > of buffers bigger than minimum, without the need to allocate the
> > > > minimum number of buffers first (to just learn the number), free them
> > > > and then allocate a bigger number again.
> > > 
> > > I don't feel this is particularly useful. One problem with the minimum number
> > > of buffers as used in the kernel is that it is often the minimum number of
> > > buffers required to make the hardware work, but it may not be optimal. E.g.
> > > quite a few capture drivers set the minimum to 2, which is enough for the
> > > hardware, but it will likely lead to dropped frames. You really need 3
> > > (one is being DMAed, one is queued and linked into the DMA engine and one is
> > > being processed by userspace).
> > > 
> > > I would actually prefer this to be the recommended minimum number of buffers,
> > > which is >= the minimum REQBUFS uses.
> > > 
> > > I.e., if you use this number and you have no special requirements, then you'll
> > > get good performance.
> > 
> > I guess we could make it so. It would make existing user space request
> > more buffers than it used to with the original meaning, but I guess it
> > shouldn't be a big problem.
> 
> I gave it a bit more thought and I feel like kernel is not the right
> place to put any assumptions on what the userspace expects "good
> performance" to be. Actually, having these controls return the minimum
> number of buffers as REQBUFS would allocate makes it very well
> specified - with this number you can only process frame by frame and
> the number of buffers added by userspace defines exactly the queue
> depth. It leaves no space for driver-specific quirks, because the
> driver doesn't decide what's "good performance" anymore.

I agree on that and I would add that the driver making any assumption
would lead to memory waste in context where less buffer will still work
(think of fence based operation as an example).

> 
> Best regards,
> Tomasz