[libcamera-devel] [PATCH] libcamera: v4l2_videodevice: Fix V4L2 buffer cache entry availabity check

Niklas Söderlund niklas.soderlund at ragnatech.se
Sun Feb 16 17:22:02 CET 2020


Hi Naushir,

On 2020-02-12 15:43:42 +0000, Naushir Patuck wrote:
> Hi Niklas,
> 
> On Wed, 12 Feb 2020 at 15:07, Niklas Söderlund
> <niklas.soderlund at ragnatech.se> wrote:
> >
> > Hi Naushir,
> >
> > On 2020-02-12 11:47:42 +0000, Naushir Patuck wrote:
> > > Hi,
> > >
> > > Would anybody - hopefully Niklas :) - be able to do a quick review of
> > > the below patch please?  It does seem we need it (or similar
> > > alternative) for good performance on our platform.
> >
> > I had a quick look and I agree this fix solves a real issue. However I'm
> > concerned it introduce another one. If more imported buffers are used
> > then there are entries in the cache there is no eviction of the oldest
> > occupied entry not in use.
> >
> 
> Yes, this was my thoughts on the fix as well.  However, I noticed that
> the cache size is always initialised to the number of buffers
> allocated, so we would always have the same number of entries as
> buffers - maybe my understanding is wrong?

That is mostly correct, in the Android HAL code we operate using 
externally provided buffers and we do not know how many they will use 
and if it will reflect the number provided at stream configuration time.

I just sent out a patch that I hope will solve your issues as well as 
address the Android HAL needs [1]. It would be super if you could test 
it and verify it solves your problem.

1. [PATCH 0/2] libcamera: V4L2BufferCache: Improve cache eviction 
   strategy

> 
> 
> 
> > This is my fault as I have it on my todo list to write a proper test for
> > the V4L2BufferCache and address any faults like the one you address in
> > this change. If you don't beat me to it I hope to do so beginning next
> > week.
> >
> 
> Not a problem.  I am running locally with the patch below, so this is
> not impacting me in any way.
> 
> > >
> > > Regards,
> > > Naush
> > >
> > > On Thu, 30 Jan 2020 at 15:43, <naush at raspberrypi.com> wrote:
> > > >
> > > > From: Naushir Patuck <naush at raspberrypi.com>
> > > >
> > > > The logic for keeping track of the cache holding dmabuf file descriptors
> > > > did not distinguish between slots that are occuiped but free and slots
> > > > that are not occupied.  In devices using imported dmabufs this would
> > > > result in ping-ponging between the top two cache entries - and that in
> > > > turn would result in unnecessary cache missies.
> > > >
> > > > Fix this by tracking if an entry is occupied separately from if an entry
> > > > is free.
> > > >
> > > > Fixes: 9e71540ebbde ("libcamera: v4l2_videodevice: Add V4L2BufferCache to deal with index mapping")
> > > > Signed-off-by: Naushir Patuck <naush at raspberrypi.com>
> > > > ---
> > > >  src/libcamera/include/v4l2_videodevice.h | 1 +
> > > >  src/libcamera/v4l2_videodevice.cpp       | 6 +++---
> > > >  2 files changed, 4 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/src/libcamera/include/v4l2_videodevice.h b/src/libcamera/include/v4l2_videodevice.h
> > > > index e4d35ab..a77b62e 100644
> > > > --- a/src/libcamera/include/v4l2_videodevice.h
> > > > +++ b/src/libcamera/include/v4l2_videodevice.h
> > > > @@ -125,6 +125,7 @@ private:
> > > >                 bool operator==(const FrameBuffer &buffer);
> > > >
> > > >                 bool free;
> > > > +               bool occupied;
> > > >
> > > >         private:
> > > >                 struct Plane {
> > > > diff --git a/src/libcamera/v4l2_videodevice.cpp b/src/libcamera/v4l2_videodevice.cpp
> > > > index 37b8d33..b887a38 100644
> > > > --- a/src/libcamera/v4l2_videodevice.cpp
> > > > +++ b/src/libcamera/v4l2_videodevice.cpp
> > > > @@ -212,7 +212,7 @@ int V4L2BufferCache::get(const FrameBuffer &buffer)
> > > >                 if (!entry.free)
> > > >                         continue;
> > > >
> > > > -               if (use < 0)
> > > > +               if (use < 0 && !entry.occupied)
> > > >                         use = index;
> > > >
> > > >                 /* Try to find a cache hit by comparing the planes. */
> > > > @@ -245,12 +245,12 @@ void V4L2BufferCache::put(unsigned int index)
> > > >  }
> > > >
> > > >  V4L2BufferCache::Entry::Entry()
> > > > -       : free(true)
> > > > +       : free(true), occupied(false)
> > > >  {
> > > >  }
> > > >
> > > >  V4L2BufferCache::Entry::Entry(bool free, const FrameBuffer &buffer)
> > > > -       : free(free)
> > > > +       : free(free), occupied(true)
> > > >  {
> > > >         for (const FrameBuffer::Plane &plane : buffer.planes())
> > > >                 planes_.emplace_back(plane);
> > > > --
> > > > 2.17.1
> > > >
> > > _______________________________________________
> > > libcamera-devel mailing list
> > > libcamera-devel at lists.libcamera.org
> > > https://lists.libcamera.org/listinfo/libcamera-devel
> >
> > --
> > Regards,
> > Niklas Söderlund
> 
> Best regards,
> Naush

-- 
Regards,
Niklas Söderlund


More information about the libcamera-devel mailing list