[Intel-gfx] [PATCH v2 3/4] drm/i915: make assert_device_not_suspended more precise

[Imre Deak]

On ke, 2015-11-18 at 16:01 +0100, Daniel Vetter wrote:
> On Wed, Nov 18, 2015 at 04:58:46PM +0200, Imre Deak wrote:
> > On ke, 2015-11-18 at 16:44 +0200, Imre Deak wrote:
> > > On ke, 2015-11-18 at 15:37 +0100, Daniel Vetter wrote:
> > > > On Mon, Nov 09, 2015 at 09:13:45PM +0200, Imre Deak wrote:
> > > > > Atm, we assert that the device is not suspended after the point
> > > > > when the
> > > > > HW is truly put to a suspended state. This is fine, but we can
> > > > > catch
> > > > > more problems if we check the RPM refcount. After that one drops
> > > > > to
> > > > > zero
> > > > > we shouldn't access the HW any more, although the actual suspend
> > > > > may be
> > > > > delayed. The only complication is that we want to avoid asserts
> > > > > while
> > > > > the suspend handler itself is running, so add a flag to handle
> > > > > this
> > > > > case.
> > > > 
> > > > Why do we want to avoid asserts firing while we go through the
> > > > suspend
> > > > handler? Calling assert_device_not_suspended from within rpm
> > > > suspend/resume code sounds like a bug. Where/why does this happen?
> > > 
> > > Yea, disable_rpm_asserts() is misnamed. Should be
> > > disable_rpm_wakelock_asserts(). Will change that in the next
> > > iteration.
> > 
> > Ok, misunderstood your question. assert_device_not_suspended() is
> > called during runtime suspend since we're accessing the HW until the
> > point we set dev_priv->pm.suspended = true. Atm this wouldn't trigger a
> > WARN, since assert_device_not_suspended() only checks pm.suspended and
> > that will check out fine, but once we start to check HW accesses
> > against the actual RPM refcount we want to disable the asserts on those
> > in the handlers, since there the refcount is zero. Hence disabling it
> > explicitly around the handlers, but we would still keep checking
> > pm.suspended.
> 
> That seems like we're mixing up 2 asserts:
> - assert_device_not_suspended: To be used in runtime_suspend code.
> - assert_holding_rpm_wakelock (or whatever, I'm bad at names): check the
>   count.

We call this assert (atm assert_device_not_suspended()) from low level
register access helpers, so we can't distinguish between calling one or
the other assert depending on whether we are on the rpm suspend path or
not. What this patch does is to switch all the places where call
assert_device_not_suspended() to assert_rpm_wakelock_held(), since that
one provides a bigger coverage. Since this change will also affect the
low level reg access functions which are called during rpm suspend, we
need to disable part of the assert that checks for the refcount which
is known to be zero there.

Otherwise assert_rpm_wakelock_held() also includes
assert_device_not_suspended(), since that should be true in all other
cases.

> 
> What am I missing?
> -Daniel