I get the above message from time to time - once every 1-2 weeks. After that the desktop is extremely laggy, no vsync (moving windows moves several blocks), scrolling is slow as hell. I only get normal behaviour by restarting X.
Created attachment 76656 [details]
Created attachment 76657 [details]
Created attachment 76658 [details]
*** This bug has been marked as a duplicate of bug 54226 ***
Hmm, it really shouldn't be that noticeable after a gpu hang... unless you are using a compositor? Certain operations will be unavailable (accelerated GL, vsync, etc), but for everything else it should fallback to a shadow buffer and for typical rendering although it may be an order of magnitude slower it shouldn't actually impact upon latency. Moving windows and scrolling should still be crisp. So if you can, please grab a 'sudo perf record -f -g -a' after such an event.
Created attachment 77229 [details]
perf.data (xz compressed)
happened again, so here the requested perf record.
And yes, I am using a compositor (kwin).
You have to parse the perf.data locally so that it can resolve the symbols etc. Can you please do 'perf report -i /path/to/perf.data | head -1500'? Sorry for skipping that detail before.
Created attachment 77242 [details]
85.26% X libc-2.15.so [.] __memcpy_ssse3_back
I also recorded just now where everything is fine (Should I post that, too?). That is the top line:
5.09% swapper [kernel.kallsyms] [k] mwait_idle_with_hints
I was hoping for just a little more information from the stacktraces. The most obvious cause for the GTT reads would be the DRI2 copies, but having that confirmed would have been useful. However, those cannot be eliminated due to the API constraints. So other than working around the broken hw, we also need to prevent the false positive EIO - which should be fixed in v3.9.