Bisected regression on RV610

[Pavel Roskin]

Hello, Alex!

Sorry for delay with the reply!  It turns out I made a huge blunder  
while bisecting.  I assumed that the radeon drivers would load from  
lib/modules, so I didn't bother running "make install" once I saw that  
only the modules are recompiled.  In fact, the video modules were  
loaded from initrd (unlike, say, WiFi drivers I'm used to deal with),  
so skipping "make install" would leave me with the drivers from the  
previous iteration.

The commit that introduces the problem is actually
2d6cc7296d4ee128ab0fa3b715f0afde511f49c2
drm/radeon: use async dma for ttm buffer moves on 6xx-SI

I could fix it simply by reverting the first chunk, that is, the part  
modifying r600_asic.

I tried the current mainline kernel 3.8.0-rc2+ (commit  
5f243b9b46a22e5790dbbc36f574c2417af49a41) and it doesn't have the  
problem!  However, the chunk of the "bad" patch can still be reverted  
with no conflicts.

So I assume the problem was fixed in another place.

If you want, I can make another bisect to find the commit that fixed  
the problem.  But if you have an idea what it was, then that it :)

Just a reminder, the problem is a hang when running unshield on some  
proprietary Windows drivers in xterm on Fedora 17 x86_64 with Radeon  
RV610.

-- 
Regards,
Pavel Roskin