I left UDL3 running overnight and it doesn't seem to have a problem. The problem is either gone, or related to remote desktop.

 There is indeed no problem if you just leave it with no work. I suspect that the problem is when there are many crazy redraws, like fullscreen video. The problem may be related to chrome desktop but it makes entire chrome os unusable until a reboot.

Yup in my case it was playing a video and running the webgl fish demo. So I think it's on the remote desktop side probably. Does it repro if you don't have remote desktop?

Unfortunately I will not have access to this setup in the near future. I'll
report as soon as I have it. Thanks

Yup I see that locally on ToT on samus for example. It seems like a sync problem on the Chrome side. Daniele do you have cycles to look at it?

By the way, in the end it has nothing to do with remote desktop, but I can reproduce if I:
- plug a mimo display to my samus
- play a 4K video (https://www.youtube.com/watch?v=6pxRHBw-k8M) and put it in fullscreen

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/00ef929f434505f4d09a130e0bc1d4a4938d490f

commit 00ef929f434505f4d09a130e0bc1d4a4938d490f
Author: dcastagna <dcastagna@chromium.org>
Date: Tue May 02 21:56:23 2017

ozone: Wait on EGLFence before committing buffers. Avoid using GL.

GbmSurfaceless::SwapBuffersAsync used to always call glFlush, to call
glFinish for universal link display device, and to insert an EGL fence
and wait on it when "EGL_ARM_implicit_external_sync" or
"EGL_EXT_image_flush_external" was available.

There is no guarantee that the current GL context when
GbmSurfaceless::SwapBuffersAsync is called is the context that was
used to draw to the buffers we are about to commit.
Additionally a glFlush or glFinish call does not guarantee content
is written and flushed to buffers after those calls return.

This caused some flickering on Mimo (udl device) since we'd submit
for scanout buffers while GL was still drawing to them.

This CL removes all the GL calls in GbmSurfaceless::SwapBuffersAsync,
that might have happened on the wrong context. It changes the code so
that we always insert and wait on an EGL fence, even on devices where
we don't have implicit sync or flush external, making SwapBuffersAsync
behavior more consistent across platforms.

BUG= 692508 

Review-Url: https://codereview.chromium.org/2858693002
Cr-Commit-Position: refs/heads/master@{#468785}

[modify] https://crrev.com/00ef929f434505f4d09a130e0bc1d4a4938d490f/ui/ozone/platform/drm/gpu/gbm_surfaceless.cc
[modify] https://crrev.com/00ef929f434505f4d09a130e0bc1d4a4938d490f/ui/ozone/platform/drm/gpu/gbm_surfaceless.h

Verified in Chrome OS 9534.0.0, 60.0.3092.0.

 Issue 707196  has been merged into this issue.