Well, "fixed" :p

I've used the patched module for 5-6 hours without issues now. So this version
checks out.

Next is with sclk_dpm_key_disabled = 0 and mclk_dpm_key_disabled = 0;

My guess is that this'll make the freeze happen again. I've dug through the
sources in an attempt to understand what's going on. My hypothesis is that the
lockup happens under rare circumstances when the driver switches from one
performance_level to another. Forcing the highest or lowest perf level is fine,
but 'auto' switches perf levels often enough that the bug will happen
relatively soon.

At first I thought it was an issue with the writeback feature that caches
certain register values because it also caches an rptr value that is used in
the driver's gpu_lockup_check and is, to my knowledge, never actually written
to.

Buuut using radeon.no_wb=1 doesn't help. So if I've found a bug it is not the
culprit of the lockups.