i915: occasional hangs on Baytrail with deeper C-states

[Daniel Bilik]

On Thu, 21 Apr 2016 10:39:32 +0200
Daniel Bilik <ddb at neosystem.org> wrote:

> ...
> Stay tuned, I'll report when my "test" kernel hangs. ;-)

Well, I've been running this "test" kernel for almost three weeks without a
hang, so it seems the hack suggested by Chris Wilson finally did the
trick.

My conclusion is that there are two separate problems that can cause hangs
on Baytrail when CPU enters deeper C-state...

First one was introduced in i915_irq.c with Linux commit 31685c2+
(drm/i915/vlv: WA for Turbo and RC6 to work together). Among other
changes, it enabled tracking rps events the specific way for Valleyview.
This one alone causes hangs less frequently, once a week or two for me.
Fix for this is attached as dfly-i915-vlv-avoid-vlv_wa_c0_ei.diff (diff
against recent Dragonfly source).

And second one was introduced in intel_pm.c with Linux commit 8fb5519+
(drm/i915: Agressive downclocking on Baytrail). This causes hangs more
often, depending on usage pattern. For me it's approximately once a day,
but freezes after couple of hours are not uncommon. Fix for this is
attached as dfly-i915-vlv-avoid-aggressive-downclocking.diff.

AFAICT fixing just one _or_ another is not sufficient, they have to be
fixed both to prevent hangs completely. I've just confirmed this myself...
Yesterday I updated my 4.5-DEVELOPMENT system to d920e19+, keeping my
manual "vlv_wa_c0_ei" fix. I believed that "aggressive downclocking" fix
was already in (it was comitted as 5d8e0f4+), but it was (probably
unintentionally) dropped with commit a05eeeb+ (drm/i915: Update to Linux
4.3). Thus I've built the kernel with "vlv_wa_c0_ei" fix but without
"aggressive downclocking" fix. And surprise, after three weeks of
problem-free operation, I've got another nice machine hang today. :)

I believe this is the reason why tens of linux users disscussing the
problem on bugzilla.kernel.org have been unsuccessful in finding reliable
solution so far. The effort to identify the (one) problem causing hangs
leads to approach "apply the patch - test - freeze - revert it and try
another one". Both patches were tested by several people on several linux
kernel versions. And several people reported success for both of them,
just to take it back after couple of days, when they get another machine
hang. Just like me three weeks ago. ;-)

Francois, would you please review attached patches and decide whether they
can be pushed to Dragonfly and kept there (ie. not get lost with next
drm/i915 update :))? Thank you.

--
 						Dan
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dfly-i915-vlv-avoid-vlv_wa_c0_ei.diff
Type: text/x-diff
Size: 1087 bytes
Desc: not available
URL: <http://lists.dragonflybsd.org/pipermail/users/attachments/20160511/f005e54b/attachment-0018.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dfly-i915-vlv-avoid-aggressive-downclocking.diff
Type: text/x-diff
Size: 2532 bytes
Desc: not available
URL: <http://lists.dragonflybsd.org/pipermail/users/attachments/20160511/f005e54b/attachment-0019.bin>