[Mesa-dev] very strange intermittent frame-dropping
Luke Kenneth Casson Leighton
lkcl at lkcl.net
Tue Apr 4 06:23:51 UTC 2017
https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=848895
there's a really strange and comprehensively cross-application
intermittent bug that's only occurring on opengl-based applications,
that's been introduced some time in the past year. to be absolutely
honest nobody's even sure it's actually opengl-related or whether it's
x11 related, but it's so weird that it's really hard to pin down.
the above bugreport on debian with chromium-browser is *just one* of
the occurrences, but it occurs with glxgears, openscad and a whole
stack of other applications.
the simplest 100% guaranteed way discovered so far to reproduce the issue is:
* install debian / testing
* install fvwm2
* set up fvwm2 to use "wireframe" for "resize window"
* run glxgears
* go ahead and resize the glxgears window with a mouse
that is *guaranteed* to result in glxgears "freezing". the only known
guaranteed way to "recover" it is to switch to text console
(ctrl-alt-f1) followed by switching back to X11 (ctrl-alt-f7).
there are various teams trying to "fix" this bug, believing it to be
the fault of their application. this includes the chromium-browser
team. what they've managed to do so far is to simply... avoid *most*
circumstances under which the lock-up occurs... but all that happens
is that the processor running the application goes into multi-tasking
meltdown... *and then still* gets a lock-up.
it doesn't matter what underlying hardware is used: intel graphics,
nvidia, ATI. it doesn't even matter if you use optirun: that *still*
results in a lockup although there you get a rather useful error
message indicating that the GPU is trying to do its job, but reports
constantly that the frame had to be dropped.
anyway, apologies: nobody whose applications are affected by this
really has a clue where the *actual* bug is so i am escalating it down
the chain of library dependencies, making people aware. it could be
X11, it could be mesa, we just don't know.
l.
More information about the mesa-dev
mailing list