[libcamera-devel] [PATCH v2 1/4] pipeline: rpi: Increase buffer import count to 32
Laurent Pinchart
laurent.pinchart at ideasonboard.com
Thu Sep 7 17:29:18 CEST 2023
Hi Naush,
Thank you for the patch.
On Tue, Jul 25, 2023 at 09:55:37AM +0100, Naushir Patuck via libcamera-devel wrote:
> Hardcode the maximum number of buffers imported to the V4L2 video device
> to 32. This only has a minor disadvantage of over-allocating cache slots
> and V4L2 buffer indexes, but does allow more headroom for using dma
> buffers allocated from outside of libcamera.
I fear there's an additional disadvantage :-S
Dmabuf instances are imported implicitly in V4L2, when preparing a
buffer (either with VIDIOC_PREPARE_BUF or VIDIOC_QBUF). At that point,
the reference count to the buffer is increased. The reference is dropped
either when the V4L2 buffer is deleted (currently only possible with
VIDIOC_REQBUFS(0)), or when the dmabuf instance is evicted by importing
a differ dmabuf (again with VIDIOC_PREPARE_BUF or VIDIOC_QBUF) for the
same buffer slot.
The eviction mechanism makes it very expensive to cycle through a pool
of dmabuf buffers if you queue them in different buffer slots all the
time, so the V4L2VideoDevice class tries hard to reuse a buffer slot
that is already associated with the dmabuf instances when queuing a
FrameBuffer. The downside, however, is that a large number of buffer
slots means that dmabuf instances will be held on for longer before they
get evicted.
If an application preallocates a limited set of buffers and cycles
through them, it's likely all fine. But if it destroys existing buffers
and allocates new ones (not necessarily all the time, it could do so for
still images only for instance), then the buffers that are freed by the
application will still have references in the kernel, and their memory
won't actually be freed. You may then end run out of memory.
I think this is an issue with the V4L2 API and its implicit dmabuf
import, there's likely nothing we can do about it in userspace to really
fix the problem, but increasing the number of buffer slots
unconditionally can make things much worse.
> Signed-off-by: Naushir Patuck <naush at raspberrypi.com>
> ---
> .../pipeline/rpi/common/rpi_stream.cpp | 33 +++++--------------
> 1 file changed, 8 insertions(+), 25 deletions(-)
>
> diff --git a/src/libcamera/pipeline/rpi/common/rpi_stream.cpp b/src/libcamera/pipeline/rpi/common/rpi_stream.cpp
> index c158843cb0ed..e9ad1e6f0848 100644
> --- a/src/libcamera/pipeline/rpi/common/rpi_stream.cpp
> +++ b/src/libcamera/pipeline/rpi/common/rpi_stream.cpp
> @@ -95,34 +95,17 @@ int Stream::prepareBuffers(unsigned int count)
> int ret;
>
> if (!(flags_ & StreamFlag::ImportOnly)) {
> - if (count) {
> - /* Export some frame buffers for internal use. */
> - ret = dev_->exportBuffers(count, &internalBuffers_);
> - if (ret < 0)
> - return ret;
> -
> - /* Add these exported buffers to the internal/external buffer list. */
> - setExportedBuffers(&internalBuffers_);
> - resetBuffers();
> - }
> + /* Export some frame buffers for internal use. */
> + ret = dev_->exportBuffers(count, &internalBuffers_);
> + if (ret < 0)
> + return ret;
>
> - /* We must import all internal/external exported buffers. */
> - count = bufferMap_.size();
> + /* Add these exported buffers to the internal/external buffer list. */
> + setExportedBuffers(&internalBuffers_);
> + resetBuffers();
> }
>
> - /*
> - * If this is an external stream, we must allocate slots for buffers that
> - * might be externally allocated. We have no indication of how many buffers
> - * may be used, so this might overallocate slots in the buffer cache.
> - * Similarly, if this stream is only importing buffers, we do the same.
> - *
> - * \todo Find a better heuristic, or, even better, an exact solution to
> - * this issue.
> - */
> - if ((flags_ & StreamFlag::External) || (flags_ & StreamFlag::ImportOnly))
> - count = count * 2;
> -
> - return dev_->importBuffers(count);
> + return dev_->importBuffers(32);
> }
>
> int Stream::queueBuffer(FrameBuffer *buffer)
--
Regards,
Laurent Pinchart
More information about the libcamera-devel
mailing list