Comment 24 for bug 353898

Revision history for this message
In , Carl Worth (cworth) wrote :

(In reply to comment #6)
> Is there additional information/debugging that can be done for this class of
> bug?

Absolutely!

Ben Gamari wrote some excellent patches to our i915 kernel driver that provide through debugfs a dump of the current batchbuffer of commands submitted to the GPU. I did some recent brushing-up of those patches, and they should be going out to the mainline Linux kernel soon.

If you get yourself a driver with those debugfs hooks, then whenever you encounter a GPU hang, you can send us a dump of the batchbuffer causing the hang. From there, Eric's written some batch-buffer parsing code, which I will be packaging up as a new intel_gpu_dump tool that will make the hex-dump from debugfs human-readable. But for now, just sending us the hex dump should be extremely valuable. We've already verified this week that the dump can make it very easy to diagnose an otherwise mystic hang.

This is all going to be easier once all of the debug stuff is nicely upstream and available on the machines of all of our users. Until then, I wrote up some notes on where to get the code and how to use it here:

https://bugs.freedesktop.org/show_bug.cgi?id=17638#c28

I hope that helps, and I'll look forward to any dumps you can provide.

Thanks,

-Carl