Re: [PATCH] [media] vb2: map dmabuf for planes on driver queue instead of vidioc_qbuf

From: Nicolas Dufresne <nicolas@ndufresne.ca>
To: Sakari Ailus <sakari.ailus@iki.fi>,
	Javier Martinez Canillas <javier@osg.samsung.com>
Cc: linux-kernel@vger.kernel.org,
	Marek Szyprowski <m.szyprowski@samsung.com>,
	Mauro Carvalho Chehab <mchehab@s-opensource.com>,
	Kyungmin Park <kyungmin.park@samsung.com>,
	Pawel Osciak <pawel@osciak.com>,
	linux-media@vger.kernel.org, Hans Verkuil <hverkuil@xs4all.nl>,
	Shuah Khan <shuahkh@osg.samsung.com>,
	Luis de Bethencourt <luisbg@osg.samsung.com>
Subject: Re: [PATCH] [media] vb2: map dmabuf for planes on driver queue instead of vidioc_qbuf
Date: Tue, 06 Sep 2016 14:34:52 -0400	[thread overview]
Message-ID: <1473186892.2668.14.camel@ndufresne.ca> (raw)
In-Reply-To: <20160720132005.GC7976@valkosipuli.retiisi.org.uk>

[-- Attachment #1: Type: text/plain, Size: 2178 bytes --]

Le mercredi 20 juillet 2016 à 16:20 +0300, Sakari Ailus a écrit :
> Hi Javier,
> 
> On Fri, Jul 15, 2016 at 12:26:06PM -0400, Javier Martinez Canillas
> wrote:
> > The buffer planes' dma-buf are currently mapped when buffers are queued
> > from userspace but it's more appropriate to do the mapping when buffers
> > are queued in the driver since that's when the actual DMA operation are
> > going to happen.
> > 
> > Suggested-by: Nicolas Dufresne <nicolas.dufresne@collabora.com>
> > Signed-off-by: Javier Martinez Canillas <javier@osg.samsung.com>
> > 
> > ---
> > 
> > Hello,
> > 
> > A side effect of this change is that if the dmabuf map fails for some
> > reasons (i.e: a driver using the DMA contig memory allocator but CMA
> > not being enabled), the fail will no longer happen on VIDIOC_QBUF but
> > later (i.e: in VIDIOC_STREAMON).
> > 
> > I don't know if that's an issue though but I think is worth mentioning.
> 
> I have the same question has Hans --- why?
> 
> I rather think we should keep the buffers mapped all the time. That'd
> require a bit of extra from the DMA-BUF framework I suppose, to support
> streaming mappings.
> 
> The reason for that is performance. If you're passing the buffer between a
> couple of hardware devices, there's no need to map and unmap it every time
> the buffer is accessed by the said devices. That'd avoid an unnecessary
> cache flush as well, something that tends to be quite expensive. On a PC
> with resolutions typically used on webcams that might not really matter. But
> if you have an embedded system with a relatively modest 10 MP camera sensor,
> it's one of the first things you'll notice if you check where the CPU time
> is being spent.

That is very interesting since the initial discussion started from the
idea of adding an implicit fence wait to the map operation. This way we
could have a dma-buf fence attached without having to modify the
drivers to support it. Buffer handles could be dispatched before there
is any data in it. Though, if we keep it mapped, I believe this idea is
simply incompatible and fences should remain explicit for extra
flexibility.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 181 bytes --]