Created attachment 115721 [details] error state Running this demo and changing the scene to daylight or so causes a GPU hang https://www.chromeexperiments.com/experiment/a-trip-under-the-moonlight and soon after that the system is dead, but there's time to get the error state. This only happens with Chromium, Firefox is fine. reproduced with drm-intel-nightly (2015y-05m-11d-16h-51m-26s), mesa 10.5.2 + 5255e1a9e6599 & e93566a15c61c3
This is reproducible with a piglit test as well. A bit easier: fbo-depth-array depth-clear Timo, you can get error state if you disable reset FWIW (also you can reboot the machine still) i915.reset=0 Assigning to Mika per Jesse's recommendation
We can only reproduce this on D-step silicon with new BIOS. Older platforms do not have the problem. This was tested on drm-intel-fixes from today (4f47c99a9be7e9b90a7f3c2c4599ea6b7c2ec49d) Mesa master from today: c820407ef0aac87546d1a778e169cfa1a915a219
Hmm. I lied, different error state. Cause of hang is different, result is the same. We need two bugs. One for the GPU hang, one for the system death. Sorry. I'll open a new one.
I can reproduce by SPT-H C0 stepping, with Bios B084.R00. But can not reproduce by SKL PCH-LP B0 stepping, with Bios V85.R00
(In reply to Ben Widawsky from comment #3) > Hmm. I lied, different error state. Cause of hang is different, result is > the same. We need two bugs. One for the GPU hang, one for the system death. Moving this one to Mesa to track the GPU hang. Bug 90854 tracks the system hang.
Can you please try this on mesa master? We're expecting 90754d2df05eafe1a3ee3cd9bb1611a19099fc49 to fix this.
I added that patch to 10.5.2 but it still hangs, also with current -next kernel which has "drm/i915: Reset request handling for gen8+"
applied to 10.6.0 and it just crashes chromium
*** Bug 91625 has been marked as a duplicate of this bug. ***
Turns out that updating mesa to 10.6.3 fixes it, doesn't hang anymore. Pulling just push constant buffer WA to 10.5 wasn't enough, need to sort out what else I need :)
Based on comment #10, this is fixed. We don't yet know what other commits are needed but commit 90754d2df05eafe1a3ee3cd9bb1611a19099fc49 was not enough itself, backported to 10.5.2. The hang remained but with a different signature. Thanks Timo, for all the testing.
bisected it down to commit 74fd226e34d0cf5e9ff43174ae69b4a66f5de1ab Author: Ben Widawsky <benjamin.widawsky@intel.com> Date: Wed Mar 25 16:52:46 2015 -0700 i965/skl: Don't use the PMA depth stall workaround and applying it (cleanly) to 10.5.9 fixed it
Huh. That was not expected; thanks a bunch Timo.
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.