Re: [PATCHv4 00/11] Implement V4L2_BUF_FLAG_NO_CACHE_* flags

[Hans Verkuil]

On 02/03/2020 05:12, Sergey Senozhatsky wrote:
> Hello,
> 
> 	v4 of the series. Typos and grammar fixes.
> 
> Previous series:
> v3 link: https://lore.kernel.org/lkml/20200226111529.180197-1-senozhatsky@xxxxxxxxxxxx
> v2 link: https://lore.kernel.org/lkml/20200204025641.218376-1-senozhatsky@xxxxxxxxxxxx/
> v1 link: https://lore.kernel.org/lkml/20191217032034.54897-1-senozhatsky@xxxxxxxxxxxx/
> 
> Series Intro
> ========================================================================
> 
>         This is a reworked version of the vb2 cache hints
> (V4L2_BUF_FLAG_NO_CACHE_INVALIDATE / V4L2_BUF_FLAG_NO_CACHE_CLEAN)
> support patch series which previsouly was developed by Sakari and
> Laurent [0].
> 
> The patch set attempts to preserve the existing behvaiour - cache
> sync is performed in ->prepare() and ->finish() (unless the buffer
> is DMA exported). User space can request âdefault behaviorâ override
> with cache management hints, which are handled on a per-buffer basis
> and should be supplied with v4l2_buffer ->flags during buffer
> preparation. There are two possible hints:
> 
> - V4L2_BUF_FLAG_NO_CACHE_INVALIDATE
>         No cache sync on ->finish()
> 
> - V4L2_BUF_FLAG_NO_CACHE_CLEAN
>         No cache sync on ->prepare()
> 
> In order to keep things on the safe side, we also require driver
> to explicitly state which of its queues (if any) support user space
> cache management hints (such queues should have ->allow_cache_hints
> bit set).
> 
> The patch set also (to some extent) simplifies allocators' ->prepare()
> and ->finish() callbacks. Namely, we move cache management decision
> making to the upper - core - layer. For example, if, previously, we
> would have something like this
> 
>         vb2_buffer_done()
>           vb2_dc_finish()
>             if (buf->db_attach)
>                return;
> 
> where each allocators' ->finish() callback would either bail
> out (DMA exported buffer, for instance) or sync, now that "bail
> out or sync" decision is made before we call into the allocator.
> 
> Along with cache management hints, user space is also able to
> adjust queue's memory consistency attributes. Memory consistency
> attribute (dma_attrs) is per-queue, yet it plays its role on the
> allocator level, when we allocate buffersâ private memory (planes).
> For the time being, only one consistency attribute is supported:
> DMA_ATTR_NON_CONSISTENT.

As mentioned in my v4 review I found a serious bug when testing with
v4l2-compliance. That meant that this series was not tested properly,
which is a requirement for something that touches the core framework.

I've posted an RFC patch with my v4l-utils changes (assumes you've run
'make sync-with-kernel' first), but that's just very basic testing. You
can use it as your starting point.

It needs to be expanded to test the various combinations of flags and
capabilities. I don't think there is a reliable way of actually testing
the cache hint functionality, so that can be skipped, but the compliance
test should at least test the basic behavior depending on whether or not
the cache hints capability is set.

I also would like to see a patch adding cache hint support to an existing
driver (more than one if possible) and the compliance output when tested
against that driver.

You should also test with the test-media script in contrib/test: run as
'sudo test-media mc' to test with all the virtual drivers. If it all passes,
then that's a good indication that there are at least no regressions.

Sorry, but this bug scared me a little, it suggests that not much testing
has been done.

Regards,

	Hans

> 
> [0] https://www.mail-archive.com/linux-media@xxxxxxxxxxxxxxx/msg112459.html
> 
> Sergey Senozhatsky (11):
>   videobuf2: add cache management members
>   videobuf2: handle V4L2 buffer cache flags
>   videobuf2: add V4L2_FLAG_MEMORY_NON_CONSISTENT flag
>   videobuf2: add queue memory consistency parameter
>   videobuf2: handle V4L2_FLAG_MEMORY_NON_CONSISTENT flag
>   videobuf2: factor out planes prepare/finish functions
>   videobuf2: do not sync caches when we are allowed not to
>   videobuf2: check ->synced flag in prepare() and finish()
>   videobuf2: add begin/end cpu_access callbacks to dma-contig
>   videobuf2: add begin/end cpu_access callbacks to dma-sg
>   videobuf2: don't test db_attach in dma-contig prepare and finish
> 
>  Documentation/media/uapi/v4l/buffer.rst       |  29 +++++
>  .../media/uapi/v4l/vidioc-create-bufs.rst     |   7 +-
>  .../media/uapi/v4l/vidioc-reqbufs.rst         |  18 ++-
>  .../media/common/videobuf2/videobuf2-core.c   | 110 +++++++++++++-----
>  .../common/videobuf2/videobuf2-dma-contig.c   |  39 ++++++-
>  .../media/common/videobuf2/videobuf2-dma-sg.c |  36 ++++--
>  .../media/common/videobuf2/videobuf2-v4l2.c   |  82 ++++++++++++-
>  drivers/media/dvb-core/dvb_vb2.c              |   2 +-
>  drivers/media/v4l2-core/v4l2-ioctl.c          |   5 +-
>  include/media/videobuf2-core.h                |  28 ++++-
>  include/uapi/linux/videodev2.h                |  11 +-
>  11 files changed, 310 insertions(+), 57 deletions(-)
>