[Intel-gfx] ✗ Fi.CI.BAT: failure for series starting with [01/29] drm/i915: Mark up ips for RCU protection

[Chris Wilson]

Quoting Patchwork (2019-04-08 11:28:00)
> == Series Details ==
> 
> Series: series starting with [01/29] drm/i915: Mark up ips for RCU protection
> URL   : https://patchwork.freedesktop.org/series/59155/
> State : failure
> 
> == Summary ==
> 
> CI Bug Log - changes from CI_DRM_5887 -> Patchwork_12723
> ====================================================
> 
> Summary
> -------
> 
>   **FAILURE**
> 
>   Serious unknown changes coming with Patchwork_12723 absolutely need to be
>   verified manually.
>   
>   If you think the reported changes have nothing to do with the changes
>   introduced in Patchwork_12723, please notify your bug team to allow them
>   to document this new failure mode, which will reduce false positives in CI.
> 
>   External URL: https://patchwork.freedesktop.org/api/1.0/series/59155/revisions/1/mbox/
> 
> Possible new issues
> -------------------
> 
>   Here are the unknown changes that may have been introduced in Patchwork_12723:
> 
> ### IGT changes ###
> 
> #### Possible regressions ####
> 
>   * igt at gem_exec_suspend@basic-s3:
>     - fi-apl-guc:         PASS -> DMESG-WARN

guc resume is snafu; we don't try to reenable guc during resume
apparently.

>   * igt at i915_selftest@live_execlists:
>     - fi-bdw-5557u:       PASS -> INCOMPLETE

Uh. Not sure, first time seeing that error.

>   * igt at i915_selftest@live_hangcheck:
>     - fi-cfl-8109u:       PASS -> INCOMPLETE [fdo#106070]
>     - fi-icl-u2:          PASS -> INCOMPLETE [fdo#108569]
>     - fi-kbl-x1275:       PASS -> INCOMPLETE [fdo#108744]
>     - fi-bxt-dsi:         PASS -> INCOMPLETE [fdo#103927]

I have a working theory for these after staring at the dumps over the
w/e. When we reset the engine, we clobber the context, but since the
request had already completed, we don't restore the RING registers. So
on the next request, it begins from the start of the ring -- causing the
seqno to be reset, and that just coincides with us retiring. Aiui, not a
new bug, but easier to hit. Explains a lot though. We need to track the
active context independently of the request during retire, and I feel
more confident in doing that after preempt-to-busy.
-Chris