Created attachment 112564 [details] GPU crash dump Error occurs consistently in firefox/chromium and are sometimes recoverable by closing dwm (like the error attached). Usually black blocks are visible beforehand.
It looks very much like the ringbuffer was overwritten with new commands whilst the GPU was still executing them, and the CS parser went off into never-never land.
Could you please grab a few more error states from firefox/chromium to see if they are all of this type?
Created attachment 112568 [details] GPU crash dump 2
(In reply to Arjan Bruin from comment #3) > Created attachment 112568 [details] > GPU crash dump 2 This looks different. Here there isn't the evidence of confusion in the ringbuffer, but that the CS kept reading past the batchbuffer - indicative that the batchbuffer is incoherent. Did this issue start occurring recently? Perhaps a change in 3.17.y? If so, is there any chance you could bisect it?
I haven't had this install for very long. Before gentoo it had arch linux and didn't have the same problem. But so much can be different. By opening 15 firefox youtube windows using a script I can reproduce quite reliably. Is this too artificial too base the bisect on? I'll attach two error dumps I got this way.
Created attachment 112616 [details] Two more gpu crash dumps
Both of those errors have inconsistent values in the batch vs being read by the GPU (what I call incoherence). Looks like you have a reasonable method of reproducing. I think this most likely a kernel issue, and I would start investigating with different kernel versions.
linux-stable 3.18.0 (b2776bf) hasn't shown any problems so far. 3.19-rc5 does hang but that may be another issue. I'll try to capture an error state.
(In reply to Arjan Bruin from comment #8) > linux-stable 3.18.0 (b2776bf) hasn't shown any problems so far. > 3.19-rc5 does hang but that may be another issue. I'll try to capture an > error state. Long time no updates, closing. If the problem still persists with latest kernels, please reopen.
Closing >1 year old resolved+fixed.
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.