[PATCH 0/4] media: Track v4l2 buffers through an allocator
Detlev Casanova
detlev.casanova at collabora.com
Thu Sep 24 05:36:15 PDT 2026
Hi Diederik,
Thank you for testing !
On Sunday, September 20, 2026 9:50:24 a.m. Eastern Daylight Time Diederik de
Haas wrote:
> Hi Detlev,
>
> On Wed Sep 16, 2026 at 4:25 PM CEST, Detlev Casanova wrote:
> > Hello,
> >
> > Currently, the only way to track buffers allocated by v4l2 from usespace
> > is to use the subsystem available debug information (e.g.:
> > /sys/kernel/debug/dma_buf/bufinfo).
> > But that information is generic and cannot be matched to a v4l2 driver or
> > to a userspace application: It is merely information about the allocation.
> >
> > Other types of allocations require the developper to find where they are
> > exposed and how to link them to their test.
> >
> > This can become hard to track when mutliple drivers are working at the
> > same time.
> >
> > To improve that, add a small wrapper around buffer allocations to keep
> > track of them at the video device level so that we can add debug
> > information to them like a name, userspace pid/fd that did the
> > allocation,... and expose them to userspace via a debugfs entry.
> >
> > It currently only supports DMA buffer allocations and adds support for
> > VB2 allocations too.
> > Other kind of memory tracking can be added later.
> > The verisilicon and rkvdec drivers have been ported to use the tracked
> > dma alloactions.
> >
> > Note that this depends on the ftrace support patch series[1] that
> > provides fd/pid info in the v4l2_fh struct.
> > That series is a bit old, so this is based on an older linux version, but
> > that only changes things for the last 2 commits.
> >
> > A v4l2top utility[2] has been made, to be used in parallel with the
> > fdinfo patch series[3], to show the list of active streams with their HW
> > and memory usage.
> >
> > With this, debugfs looks like this when decoding a HEVC 1080p stream with
> >
> > rkvdec on rk3588:
> > root # cat /sys/kernel/debug/v4l2/fdc38100.video-codec/mem
> > created-by fd pid size
> > label
> > -----------------------------------------------------------------------
> > -------------- gst-launch-1.0 7 635
> > 39184 vdpu381-hevc-priv-tbl gst-launch-1.0
> > 7 635 4177920
> > cap-00000000ac34392e-7 gst-launch-1.0 7
> > 635 4177920 cap-00000000ac34392e-6 gst-launch-1.0
> > 7 635 4177920
> > cap-00000000ac34392e-5 gst-launch-1.0 7
> > 635 4177920 cap-00000000ac34392e-4 gst-launch-1.0
> > 7 635 4177920
> > cap-00000000ac34392e-3 gst-launch-1.0 7
> > 635 4177920 cap-00000000ac34392e-2 gst-launch-1.0
> > 7 635 4177920
> > cap-00000000ac34392e-1 gst-launch-1.0 7
> > 635 4177920 cap-00000000ac34392e-0 gst-launch-1.0
> > 7 635 3133440
> > out-000000005c0230f6-1 gst-launch-1.0 7
> > 635 3133440 out-000000005c0230f6-0
> > =======================================================================
> > ============== Total size: 39729424
>
> I build a kernel with your 3 patch series and I also build the v4l2top
> utility. Cool tool :-D
> Especially after I found out I should build the 'upstream' branch ;-)
>
> There were a couple of things I noticed:
> 1) With 'plain' kernel 7.3-rc3 but also with all my patches included, but
> excluding the patches from this series, I see this:
>
> ```
> root at nanopc-t6-plus:~# ls -lh /sys/kernel/debug/v4l2/
> total 0
> drwxr-xr-x 3 root root 0 Sep 20 15:10 fdee0000.hdmi_receiver
> ```
>
> Which is presumably the reason that with a kernel which also includes this
> patch series, I get the following kernel error:
> ``[ 10.849405] debugfs: 'v4l2' already exists in '/'``
>
> I don't think this should generate an error/warning/info message at all,
> but just silently dealt with.
Indeed, that is something I will fix in the next version.
>
> Checking v4l2 debug dir again results in this:
> ```
> root at nanopc-t6-plus:~# ls -lh /sys/kernel/debug/v4l2/
> total 0
> drwxr-xr-x 2 root root 0 Sep 20 15:13 fdb60000.rga
> drwxr-xr-x 2 root root 0 Sep 20 15:13 fdb80000.rga
> drwxr-xr-x 2 root root 0 Sep 20 15:13 fdc38000.video-codec
> drwxr-xr-x 2 root root 0 Sep 20 15:13 fdee0000.hdmi_receiver
> ```
>
> I noticed my video-codec address was slightly different from yours.
Yes, mine is on rk3588. A different SoC will have a different device register
address.
> Playing a 1080p BBB x264 video with a patched ffmpeg+mpv, results in this:
> ```
> root at nanopc-t6-plus:~# cat /sys/kernel/debug/v4l2/fdc38000.video-codec/mem
> created-by fd pid size
> label
> ---------------------------------------------------------------------------
> ---------- mpv 26 2027
> 4177920 cap-00000000205a3b3a-10 mpv 26
> 2027 4177920 cap-00000000205a3b3a-9 mpv
> 26 2027 4177920
> cap-00000000205a3b3a-8 mpv 26 2027
> 16736 vdpu381-h264-priv-tbl mpv
> 26 2027 3133440 out-0000000079002cb1-3
> mpv 26 2027 3133440
> out-0000000079002cb1-2 mpv 26
> 2027 3133440 out-0000000079002cb1-1 mpv
> 26 2027 3133440
> out-0000000079002cb1-0 mpv 26 2027
> 4177920 cap-00000000205a3b3a-7 mpv
> 26 2027 4177920
> cap-00000000205a3b3a-6 mpv 26 2027
> 4177920 cap-00000000205a3b3a-5 mpv
> 26 2027 4177920
> cap-00000000205a3b3a-4 mpv 26 2027
> 4177920 cap-00000000205a3b3a-3 mpv
> 26 2027 4177920
> cap-00000000205a3b3a-2 mpv 26 2027
> 4177920 cap-00000000205a3b3a-1 mpv
> 26 2027 4177920
> cap-00000000205a3b3a-0
> ===========================================================================
> ========== Total size: 58507616
> ```
>
> And looking at the v4l2top utility, I noticed the following:
> 1. I see differing indications of the load while it's playing :-D
\o/
> 2. The ``Clock rate`` is always 750.0MHz (or 786431991Hz) (expected I
> think?)
Yes, the clock won't change, for now at least. Some decoders could adapt their
clocks depending on the work. The frequency value will also be used to
determine exact core usage for those that expose a used cycle count after each
job (I did not implement that yet in v4l2top)
> 3. The ``Total Memory`` is always 0B and the ``Memory usage
> details`` column is always empty. That's also the case when I press one of
> the arrow keys which seems to select the entry. Am I missing sth or doing
> sth wrong?
My guess would be that you didn't run it as root (this is usually needed to
access the debugfs*). But if you did, we will have to dig a little.
*: Note that later, we'd like to report global memory usage through fdinfo, so
that the metric is accessible as a user, with buffer details still from the
debugfs (so that'd need root access)
Detlev.
>
> > [1]:
> > https://lore.kernel.org/all/20260610-v4l2-add-ftrace-v2-0-9756edf72ac1@co
> > llabora.com/ [2]: https://github.com/cazou/v4l2top/tree/upstream
> > [3]:
> > https://lore.kernel.org/all/20260706-v4l2-add-fdinfo-v3-0-d556568cf38e@co
> > llabora.com/
> >
> > Signed-off-by: Detlev Casanova <detlev.casanova at collabora.com>
> > ---
> >
> > Detlev Casanova (4):
> > media: Add a v4l2 memory allocations tracker
> > media: Store v4l2_fh in the vb_queue
> > media: verisilicon: Switch to tracked dma allocations
> > media: rkvdec: Switch to tracked dma allocations
> >
> > drivers/media/common/videobuf2/Makefile | 1 +
> > drivers/media/common/videobuf2/v4l2-allocator.c | 186
> > +++++++++++++++++++++ .../media/common/videobuf2/videobuf2-dma-contig.c
> > | 27 ++-
> > .../media/platform/rockchip/rkvdec/rkvdec-h264.c | 14 +-
> > .../media/platform/rockchip/rkvdec/rkvdec-hevc.c | 14 +-
> > .../media/platform/rockchip/rkvdec/rkvdec-rcb.c | 21 ++-
> > .../platform/rockchip/rkvdec/rkvdec-vdpu381-h264.c | 14 +-
> > .../platform/rockchip/rkvdec/rkvdec-vdpu381-hevc.c | 14 +-
> > .../platform/rockchip/rkvdec/rkvdec-vdpu383-h264.c | 14 +-
> > .../platform/rockchip/rkvdec/rkvdec-vdpu383-hevc.c | 14 +-
> > .../media/platform/rockchip/rkvdec/rkvdec-vp9.c | 33 ++--
> > drivers/media/platform/rockchip/rkvdec/rkvdec.c | 4 +
> > drivers/media/platform/verisilicon/hantro.h | 1 +
> > drivers/media/platform/verisilicon/hantro_drv.c | 4 +
> > drivers/media/platform/verisilicon/hantro_h264.c | 7 +-
> > drivers/media/platform/verisilicon/hantro_hevc.c | 100 ++++++-----
> > drivers/media/platform/verisilicon/hantro_mpeg2.c | 16 +-
> > .../media/platform/verisilicon/hantro_postproc.c | 14 +-
> > drivers/media/platform/verisilicon/hantro_vp8.c | 25 +--
> > drivers/media/platform/verisilicon/hantro_vp9.c | 34 +++-
> > .../verisilicon/rockchip_vpu981_hw_av1_dec.c | 160
> > +++++++++++-------
> > drivers/media/v4l2-core/v4l2-device.c | 4 +-
> > include/media/v4l2-allocator.h | 26 +++
> > include/media/v4l2-device.h | 2 +
> > include/media/videobuf2-core.h | 3 +
> > 25 files changed, 555 insertions(+), 197 deletions(-)
> >
> > ---
> > base-commit: 66affa37cfac0aec061cc4bcf4a065b0c52f7e19
> > change-id: 20260612-v4l2-add-mem-tracker-da0088c74a64
> > prerequisite-change-id: 20260608-v4l2-add-ftrace-aec6e7f60a6c:v2
> > prerequisite-patch-id: bd45b4df66799a7f9f22b49973a78d1bd3590a4a
> > prerequisite-patch-id: 6a5ed5615f08257cf854d441785066b5ff4d5d47
> > prerequisite-patch-id: 02594c869498b9e91e416d70242afbaa6b3a8d6a
> > prerequisite-patch-id: 9f25ec78f3b99ac1c5d42eb732f5bd9eae73fadd
> > prerequisite-patch-id: 4ed2c2e8c3abfa7f6328d0ea4f488ca5a3eec8fb
> > prerequisite-patch-id: 8787d50eb80e1a63a57a123268cade23c9e64618
> > prerequisite-patch-id: bfe85822378fa771f9bbc3e1e8c8ff6bce0eb7fd
> > prerequisite-patch-id: 784d0801da320f7b41194f4632ee422c32f1b5b8
> > prerequisite-patch-id: 362c8366e2efbfb5e6f1a3f71cfddbd46e8d1506
> >
> > Best regards,
> > --
> > Detlev Casanova <detlev.casanova at collabora.com>
> >
> >
> > _______________________________________________
> > Linux-rockchip mailing list
> > Linux-rockchip at lists.infradead.org
> > http://lists.infradead.org/mailman/listinfo/linux-rockchip
More information about the Linux-rockchip
mailing list