Re: [PATCH v4 5/5] drm/vc4: dsi: Recover from a timed-out transfer

From: Henry Zhang

Date: Wed Nov 18 2026 - 18:55:40 EST


On Wed, Nov 18, 2026 at 1:11 PM Dave Stevenson
<dave.stevenson@xxxxxxxxxxxxxxx> wrote:
> Whilst it's a belt and braces approach to work around the issue, I'll
> have a dig next week to see if I can reproduce it and see what is
> going on.

Thanks, that would help. In case it saves you time, this is how we hit
it, on a Pi 4 with the panel on DSI1 in video mode, display on:

- It starts with a DCS read whose bus turnaround the peripheral does
not complete, so the read gets the LP receive timeout.
- From then on, every transfer times out with no interrupt and STAT
reading back 0, writes included, until the next modeset.

I don't have a reliable way to trigger the first failed read. I have
not tried a peripheral that never answers a read, so I don't know if
that alone is enough to get the block stuck.

> There are very few use cases that have a need for sending DCS commands
> whilst transmitting video, so it's a bit niche. The reset stopping
> video isn't nice although I guess it should be fairly brief.

I understand it is niche in general, but for us it is a requirement: we
need reads (with the bus turnaround) in both video mode and command
mode. In video mode our test tools read panel registers only when no
video content is being played, but the display is still on, so the
host is still sending video frames during the reads. The same reads
have to work with the panel in command mode. So a stuck block hurts us
in either mode.

I checked both today, on our downstream kernel with this series and
some local changes: 1000 reads in video mode (display on, static
picture) and 1000 in command mode all completed. We did not reproduce a
timeout in this run, so it did not exercise the recovery. That does not
change the failure described above, which we did capture earlier; we
still have no reliable way to trigger it.

I haven't measured how long video stops during the reset. It is only
done after a timeout, when the link is already unusable for commands.

If you find the cause and a fix that avoids the stuck state, I am happy
to drop this patch, or rework it, and test whatever you have on our
setup. Until then I'll hold 5/5 in v5 as it is.

Thanks,
Henry