Re: [RFC][PATCH 13/15] videobuf2: do not sync buffers for DMABUF queues

[Tomasz Figa]

On Thu, Jan 23, 2020 at 8:35 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>
> On 1/22/20 6:05 AM, Sergey Senozhatsky wrote:
> > On (20/01/10 11:30), Hans Verkuil wrote:
> > [..]
> >>> diff --git a/drivers/media/common/videobuf2/videobuf2-v4l2.c b/drivers/media/common/videobuf2/videobuf2-v4l2.c
> >>> index 1762849288ae..2b9d3318e6fb 100644
> >>> --- a/drivers/media/common/videobuf2/videobuf2-v4l2.c
> >>> +++ b/drivers/media/common/videobuf2/videobuf2-v4l2.c
> >>> @@ -341,8 +341,22 @@ static void set_buffer_cache_hints(struct vb2_queue *q,
> >>>                                struct vb2_buffer *vb,
> >>>                                struct v4l2_buffer *b)
> >>>  {
> >>> -   vb->need_cache_sync_on_prepare = 1;
> >>> +   /*
> >>> +    * DMA exporter should take care of cache syncs, so we can avoid
> >>> +    * explicit ->prepare()/->finish() syncs.
> >>> +    */
> >>> +   if (q->memory == VB2_MEMORY_DMABUF) {
> >>> +           vb->need_cache_sync_on_finish = 0;
> >>> +           vb->need_cache_sync_on_prepare = 0;
> >>> +           return;
> >>> +   }
> >>>
> >>> +   /*
> >>> +    * For other ->memory types we always need ->prepare() cache
> >>> +    * sync. ->finish() cache sync, however, can be avoided when queue
> >>> +    * direction is TO_DEVICE.
> >>> +    */
> >>> +   vb->need_cache_sync_on_prepare = 1;
> >>
> >> I'm trying to remember: what needs to be done in prepare()
> >> for a capture buffer? I thought that for capture you only
> >> needed to invalidate the cache in finish(), but nothing needs
> >> to be done in the prepare().
> >
> > Hmm. Not sure. A precaution in case if user-space wrote to that buffer?
>
> But whatever was written in the buffer is going to be overwritten anyway.
>
> Unless I am mistaken the current situation is that the cache syncs are done
> in both prepare and finish, regardless of the DMA direction.
>
> I would keep that behavior to avoid introducing any unexpected regressions.
>

It wouldn't be surprising if the buffer was first filled with default
values (e.g. all zeroes) on the CPU. That would make the cache lines
dirty and they could overwrite what the device writes. So we need to
flush (aka clean) the write-back caches on prepare for CAPTURE
buffers.

> Then, if q->allow_cache_hint is set, then default to a cache sync (cache clean)
> in the prepare for OUTPUT buffers and a cache sync (cache invalidate) in the
> finish for CAPTURE buffers.

I'd still default to the existing behavior even if allow_cache_hint is
set, because of what I wrote above. Then if the userspace doesn't ever
write to the buffers, it can request no flush/clean by setting the
V4L2_BUF_FLAG_NO_CACHE_CLEAN flag when queuing the buffer.

>
> This also means that any drivers that want to access a buffer in between the
> prepare...finish calls will need to do a begin/end_cpu_access. But that's a
> separate matter.

AFAIR with current design of the series, the drivers can opt-in for
userspace cache sync hints, so by default even if the userspace
requests sync to be skipped, it wouldn't have any effect unless the
driver allows so. Then I'd imagine migrating all the drivers to
request clean/invalidate explicitly. Note that the DMA-buf
begin/end_cpu_access isn't enough here. We'd need something like
vb2_begin/end_cpu_access() which also takes care of syncing
inconsistent MMAP and USERPTR buffers.

>
> Regards,
>
>         Hans
>
> >
> > +     if (q->dma_dir == DMA_FROM_DEVICE)
> > +             q->need_cache_sync_on_prepare = 0;
> >
> > ?
> >
> >       -ss
> >
>