[Intel-gfx] [5.10.y regression] i915 clear-residuals mitigation is causing gfx issues

[Greg Kroah-Hartman]

On Sun, Feb 14, 2021 at 05:00:44PM +0100, Hans de Goede wrote:
> Hi,
> 
> On 2/11/21 1:26 PM, Hans de Goede wrote:
> > Hi,
> > 
> > On 2/11/21 11:49 AM, Chris Wilson wrote:
> >> Quoting Hans de Goede (2021-02-11 10:36:13)
> >>> Hi,
> >>>
> >>> On 2/10/21 1:48 PM, Chris Wilson wrote:
> >>>> Quoting Hans de Goede (2021-02-10 10:37:19)
> >>>>> Hi,
> >>>>>
> >>>>> On 2/10/21 12:07 AM, Chris Wilson wrote:
> >>>>>> Quoting Hans de Goede (2021-02-09 11:46:46)
> >>>>>>> Hi,
> >>>>>>>
> >>>>>>> On 2/9/21 12:27 AM, Chris Wilson wrote:
> >>>>>>>> Quoting Hans de Goede (2021-02-08 20:38:58)
> >>>>>>>>> Hi All,
> >>>>>>>>>
> >>>>>>>>> We (Fedora) have been receiving reports from multiple users about gfx issues / glitches
> >>>>>>>>> stating with 5.10.9. All reporters are users of Ivy Bridge / Haswell iGPUs and all
> >>>>>>>>> reporters report that adding i915.mitigations=off to the cmdline fixes things, see:
> >>>>>>>>
> >>>>>>>> I tried to reproduce this on the w/e on hsw-gt1, to no avail; and piglit
> >>>>>>>> did not report any differences with and without mitigations. I have yet
> >>>>>>>> to test other platforms. So I don't yet have an alternative.
> >>>>>>>
> >>>>>>> Note the original / first reporter of:
> >>>>>>>
> >>>>>>> https://bugzilla.redhat.com/show_bug.cgi?id=1925346
> >>>>>>>
> >>>>>>> Is using hsw-gt2, so it seems that the problem is not just the enabling of
> >>>>>>> the mitigations on ivy-bridge / bay-trail but that there actually is
> >>>>>>> a regression on devices where the WA worked fine before...
> >>>>>>
> >>>>>> There have been 3 crashes uploaded related to v5.10.9, and in all 3
> >>>>>> cases the ACTHD has been in the first page. This strongly suggests that
> >>>>>> the w/a is scribbling over address 0. And there's then a very good
> >>>>>> chance that
> >>>>>>
> >>>>>> commit 29d35b73ead4e41aa0d1a954c9bfbdce659ec5d6
> >>>>>> Author: Chris Wilson <chris at chris-wilson.co.uk>
> >>>>>> Date:   Mon Jan 25 12:50:33 2021 +0000
> >>>>>>
> >>>>>>     drm/i915/gt: Always try to reserve GGTT address 0x0
> >>>>>>     
> >>>>>>     commit 489140b5ba2e7cc4b853c29e0591895ddb462a82 upstream.
> >>>>>>
> >>>>>> in v5.10.14 is sufficient to hide the issue.
> >>>>>
> >>>>> That one actually is already in v5.10.13 and the various reportes of these
> >>>>> issues have already tested 5.10.13. They did mention that it took longer
> >>>>> to reproduce with 5.10.13 then with 5.10.10, but that could also be due to:
> >>>>>
> >>>>> "drm/i915/gt: Clear CACHE_MODE prior to clearing residuals"
> >>>>> https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?h=linux-5.10.y&id=520d05a77b2866eb4cb9e548e1d8c8abcfe60ec5
> >>>>
> >>>> Started looking for scratch page overwrites, and found this little gem:
> >>>> https://patchwork.freedesktop.org/patch/420436/?series=86947&rev=1
> >>>>
> >>>> Looks promising wrt the cause of overwriting random addresses -- and
> >>>> I hope that is the explanation for the glitches/hangs. I have a hsw gt2
> >>>> with gnome shell, piglit is happy, but I suspect it is all due to
> >>>> placement and so will only occur at random.
> >>>
> >>> If you can give me a list of commits to cherry-pick then I can prepare
> >>> a Fedora 5.10.y kernel which those added for the group of Fedora users
> >>> who are hitting this to test.
> >>
> >> e627d5923cae ("drm/i915/gt: One more flush for Baytrail clear residuals")
> >> d30bbd62b1bf ("drm/i915/gt: Flush before changing register state")
> >> 1914911f4aa0 ("drm/i915/gt: Correct surface base address for renderclear")
> > 
> > Thanks, the test-kernel is building now. I will let you know when I have
> > heard back from the Fedora users (this will likely take 1-2 days).
> 
> I've heard back from 2 of the reporters who were seeing issues with 5.10.9+
> 
> And I'm happy to report 5.10.15 + the 3 commits mentioned above cherry-picked
> on top fixes the graphics glitches for them.
> 
> So if we can get these 3 commits into 5.10.y and 5.11.y then this should be
> resolved.

Great!

Hopefully these will show up in Linus's tree soon...