Commen=
t # 26
on bug 93341<=
/a>
from Jean-Fran=C3=A7ois Fortin Tam
OK, I've got good news... Julien, thanks to the crazy furry do=
nut "torture
test" you suggested, I was able to finally pinpoint the real trigger f=
or this
bug.
My understanding is that on Radeons (well, at least the Radeon HD 7770), th=
ere
is an emergency mechanism in the hardware (or firmware/microcode maybe) that
activates self-throttling of performances when the GPU reaches a critical
temperature. Normally, the video driver is supposed to handle this state ch=
ange
gracefully, however the radeonsi/radeon/amdgpu driver on Linux does not, so=
the
kernel panics because the driver went belly up.
During additional testing today, where I forced my GPU to overheat, I was a=
ble
to determine that the critical point is the same as on Windows: 113 degrees
Celsius. As soon as you go over 112... boom, dead radeonsi driver + kernel =
oops
(with the same error messages as my previous logs above). Additionally,
lm_sensors thinks the temperature has instantly jumped to 511 degrees Celsi=
us
(!), and the readings stay stuck at 511 Celsius.
"Duh! Just get better cooling!" might sound like a workaround (ju=
st like
keeping the case open), but nope, technically, it's still a software/driver
issue: the Linux driver should handle such scenarios gracefully just as wel=
l as
the Windows driver. In Windows, breaching the 110-113 degrees Celsius limit
results in the video driver simply dropping frames massively, continuing to
function at reduced performance (ie: going from 40-60 fps to 10-15 fps on o=
ne
of my benchmarks). The system never crashes.
So the bug here, as I understand it, is that the radeonsi driver on Linux d=
oes
not handle the event where the hardware force-throttles itself.
---------
Contextual notes:
The reason why I only started experiencing this issue in December 2015 (as =
I've
had the GPU since 2012) was that I changed my PC case then, which means a
different airflow and cooling behavior... And the reason why it was so hard=
to
get consistent crashes here was that when I was trying to troubleshoot it, I
was sometimes doing it with the case closed, sometimes with the case open (=
when
trying with a different power supply unit using a "siamese transplant&=
quot; across
another computer, for example). If I keep my case open, the card will never
reach the critical temperature and so the issue will not happen. I might ge=
t a
system "freeze" (possibly saying "*ERROR*
si_restrict_performance_levels_before_switch failed") after many hours=
of
torture testing, but the symptoms are different (the screen does not turn o=
ff,
image stays on with everything frozen, and nothing else in the logs) and so=
I
presume that to be a different issue.