Re: [PATCH] drm/msm/dpu: ensure device suspend happens during PM sleep

On 2020-03-25 21:20, Doug Anderson wrote:

Hi,

On Wed, Mar 25, 2020 at 8:40 AM Rob Clark [off-list ref] wrote:

quoted

On Tue, Mar 24, 2020 at 7:35 AM Doug Anderson [off-list ref] 
wrote:

quoted

Hi,

On Sun, Mar 22, 2020 at 11:14 PM Kalyan Thota [off-list ref] wrote:

quoted

"The PM core always increments the runtime usage counter
before calling the ->suspend() callback and decrements it
after calling the ->resume() callback"

DPU and DSI are managed as runtime devices. When
suspend is triggered, PM core adds a refcount on all the
devices and calls device suspend, since usage count is
already incremented, runtime suspend was not getting called
and it kept the clocks on which resulted in target not
entering into XO shutdown.

Add changes to manage runtime devices during pm sleep.

Changes in v1:
 - Remove unnecessary checks in the function
     _dpu_kms_disable_dpu (Rob Clark).

I'm wondering what happened with my feedback on v1, AKA:

https://lore.kernel.org/r/CAD=FV=VxzEV40g+ieuEN+7o=34+wM8MHO8o7T5zA1Yosx7SVWg@mail.gmail.com (local)

Maybe you didn't see it?  ...or if you or Rob think I'm way off base
(always possible) then please tell me so.

-- I didn't notice your comments earlier. Apologies !!

quoted

At least w/ the current patch, disable_dpu should not be called for
screen-off (although I'd hope if all the screens are off the device
would suspend).

OK, that's good.

-- Rob has answered it, with current change disable_dpu will only be 
called during pm_suspend.

quoted

But I won't claim to be a pm expert.. so not really
sure if this is the best approach or not.  I don't think our
arrangement of sub-devices under a parent is completely abnormal, so
it does feel like there should be a simpler solution..

I think the other arguments about asymmetry are still valid and I've
fixed bugs around this type of thing in the past.  For instance, see
commit f7ccbed656f7 ("drm/rockchip: Suspend DP late").

* What happens if suspend is aborted partway through (by getting a
wakeup even as you're suspending, for instance)?  In such a case some
of the normal suspend calls will be called but "suspend_late" won't be
called.  Does that mess up your counting?

-- I understand this concern, i'll explore a bit more on how to handle 
"failed to suspend","early awake"
cases (to restore the usage_count) since suspend_late wont be called.

*From your description, it sure seems like this part of the
runtime_pm.rst doc is relevant to you:

Did I misunderstand and this isn't what you want?  Looking a bit
further, maybe the right thing is to use the "SMART_SUSPEND" flag?

-- if you notice in the device_prepare 
(https://elixir.bootlin.com/linux/latest/source/drivers/base/power/main.c#L1913)
there is a pm_runtime_get_noresume at L1931, which will increment the 
usagecount before triggering client prepare call, hence implementing 
prepare wont fetch us much.

This appears to be more for the cases when device is runtime suspended 
and suspend followed later
"one example usecase that i can think of, is screen timeout after that 
suspend is triggered"

currently the problem i am looking at is that
	PM Core does +1 in device prepare
		DPU driver does -1 in suspend
		DPU driver does +1 in suspend late  ( look for right place )
	PM core does -1 in device complete

i'll get back after exploring a bit.

-Doug