Just to update, I've captured three different traces but none have been able to
reproduce the problem on replay. I've also tried the following:

* Looping a trace replay over 24 hours continuously -- no repro
* Running with a -O0 Mesa build -- hang remains
* Going directly to fallback in all cases during si_dma_copy (wild guess based
on code comments) -- hang remains

I don't think traces will be a fruitful method of debugging, unless someone can
suggest something I'm doing wrong. I'm continuing to look at this. If anyone
has a hypothesis and would like to send a patch then I could build and test
with it.