[Intel-gfx] Haswell Regression in 5.9-rcX and lower 5.7.x, 5.8.x #2024 - Revert 47f8253d2b89 ("drm/i915/gen7: Clear all EU/L3 residual contexts") ?

[Dirk Neukirchen]

Hi David, Chris and lists

I am inquiring about the current status of #2024 [1]

Problem: Kernels 5.7.x , 5.8.x, current 5.9-rcs and drm-tip have a large regression on (some?) Haswell (HSW)

This is verified by _multiple people_ using different methods.
All his is documented in [1]  , Kernel 5.6.19 was fine
Result: no output, no usable desktop due to gpu crashes

Hardware : 2x Acer C720 Chromebooks and potentially others

Methods:
- running mpv with vaapi enabled
- glxgears
- automatic desktop environment start (probably with acceleration)

There is atm no fix in drm-tip integrated (5.9.0_rc5bisect_g30b3e38bd6d5 for example still errors)
 and no activity from Intel in the bug report beside the "working" suggested patches / hacks there

There should be enough bisection info an there.
Nobody asked there to get more info or different bisects/trees.

So - the ticket is still open.

Because Fedora uses newish Kernels I would like to use it "normally" without
going for 5.4 LTS or waiting an unknown amount of time - as 5.9.x currently seems to have the
same regression.

Affected Hardware at my end:
- C720 with a Intel Celeron 2955U
- UEFI Bios instead of original Bios (mrchromebox.tech) if thats relevant for hardware init related bugs;
But the Kernel should take care of the correct initialization anyway

Feel free to comment which trees, branches, tags or patches we should try to help
and what _more_ to report for "Report-By" , "Tested-By", "Verified-By"
tagging so that [1] can be closed.

Greetings, Dirk

[1]https://gitlab.freedesktop.org/drm/intel/-/issues/2024