[Intel-gfx] [DMC_BUGFIX_SKL_V2 3/5] drm/i915/skl: Making DC6 entry is the last call in suspend flow.

[Daniel Vetter]

On Tue, Sep 29, 2015 at 2:35 PM, Patrik Jakobsson
<patrik.jakobsson at linux.intel.com> wrote:
> On Tue, Sep 29, 2015 at 11:01:35AM +0200, Daniel Vetter wrote:
>> On Tue, Sep 29, 2015 at 11:08:25AM +0530, Animesh Manna wrote:
>> >
>> >
>> > On 09/28/2015 12:51 PM, Daniel Vetter wrote:
>> > >On Wed, Aug 26, 2015 at 01:36:07AM +0530, Animesh Manna wrote:
>> > >>Mmio register access after dc6/dc5 entry is not allowed when
>> > >>DC6 power states are enabled according to bspec (bspec-id 0527),
>> > >>so enabling dc6 as the last call in suspend flow.
>> > >We unconditionaly grab a power well reference for everything in our
>> > >suspend/resume functions, which means dc6/5 should be prevented for long
>> > >enough.
>> > >
>> > >Also since dc6/5 are part of the power well framework you can't just
>> > >enable/disable them directly, instead you need to get/put the
>> > >corresponding power wells.
>> > >
>> > >Finally this patch doesn't just do that, but also frobs around a lot in
>> > >set power well code itself, and I have no idea what it does there and why.
>> > >It does smell a bit like you're just breaking dc6 for runtime pm though,
>> > >which wouldn't be good.
>> > >
>> > >Anyway I decided to just merge this since this patch series has been
>> > >floating around since forever, but then the patch didn't apply cleanly and
>> > >so dropped it.
>> >
>> > I mentioned in my commit message that we have a h/w workaround.
>> > "Mmio register access after dc6/dc5 entry is not allowed when
>> > DC6 power states are enabled"
>> >
>> > This patch is mandatory to solve the pc10 entry issue for skylake.
>> > Planning to send again after rebase on top of tree.
>
> Hi Animesh and Daniel
>
> I'm trying to shed some light on this. It seems rather confusing atm. Animesh,
> please correct me if I'm wrong below.
>
>> The problem isn't that the patch didn't apply cleanly, but that it seems
>> to have fundamental issues:
>> - We already grab power well references around suspend/resume code, see
>>   the call to intel_power_domains_init_hw and the calls to
>>   intel_display_set_init_power. That code is supposed to ensure that while
>>   suspend/resume is going on _all_ power wells are on (and hence dc6/5 are
>>   forbidden). It's a bit a big hammer, but we already have the code to do
>>   exactly what you're trying to do here. It could be though that for skl
>>   it's in the wrong order (but then the commit message must state very
>>   clearly what's the exact depency, the gpu is assembled from a pile of
>>   various different blocks which are all mostly independent).
>
> DC6 can only be enabled when _all_ power wells are disabled. Hence this patch
> moves DC6 enabled/disabled to the runtime suspend/resume callbacks. Previously
> we only disabled DC6 when PW2 was enabled.
>
> The DC state table says:
> - Everything off: DC6 enabled
> - PW0 enabled: DC5 enabled but DC6 is disabled
> - PW1 enabled: DC5 and DC6 disabled
> - PW2 enabled: DC5 and DC6 disabled

Ok, so what the patch does (and that is at least self-consistent) is
to move the DC6 enable/disable from PW2 to the top-level domain. That
doesn't quite match what you're describing here since we still have
PW1 managed by the driver, but that's short-circuited now.

>> - Your patch directly calls skl_enable/disable_dc6 instead of going
>>   through the power well abstraction. Breaking these abstraction needs a
>>   really good reason, adequate assurance that the refcounts are ok and
>>   nothing breaks (using WARN_ON) and an explanation why the refcount
>>   interface for display power management doesn't work.
>
> This is indeed a hack but since all we need is the guarantee that a power well
> is in fact enabled we can see the operation as: enable pw + disable dc state.
> This is currently hacked into skl_set_power_well(). It's not nice but it takes
> away the complexity of modeling DC states as power wells.
>
> The alternative would be to turn DC5 and DC6 into power wells and add proper
> dependencies. What I've discovered is that it's quite a tight fit.

Nah, if the goal is to move just DC6 the patch seems fine. I was just
really confused that the commit message talked about both dc6 _and_
dc5 being impossible to do, and then only moving half of the story
around. And I was also a bit confused with runtime pm vs. system
suspend (again).

So if we can confirm that this is really only about DC6 and not also
about dc5 like the commit message claims, then I can clean up the
commit message and apply the patch.

>> - You also do changes to the power well code itself, which looks like it
>>   might break dc6 for runtime pm (by only going into dc5 at most). That
>>   needs to be a separate patch or have a much better explanation.
>
> If I understood correctly, that is the entire point of this patch. Don't allow
> DC6 unless all power wells are disabled. My question though is why we don't
> disable DC5 for PW1 as well?

Afaiui dmc manages pw1 for us and we shouldn't ever touch it from the
driver side. I'm not entirely sure about that though since the current
fix is a hack. If we really shouldn't touch pw1 on skl then we should
remove that power domain from our skl code entirely.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch